动态特征优化机制下的跨尺度红外行人检测

郝帅; 何田; 马旭; 杨磊; 孙思雅

doi:10.37188/OPE.20223019.2390

English Version Light学术出版中心

您当前的位置：

首页 >

文章列表页 >

动态特征优化机制下的跨尺度红外行人检测

信息科学 | 更新时间：2022-10-24

- 动态特征优化机制下的跨尺度红外行人检测
- Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism
- 光学精密工程
- 作者机构：
  
  1.西安科技大学电气与控制工程学院，陕西西安 710054
- 作者简介：
  
  [ "郝帅（1986-），男，博士，硕士生导师，主要从事人工智能、智能电网方面的研究工作。E-mail：haoxust@163.com" ]
  
  郝帅（1986-），男，博士，硕士生导师，主要从事人工智能、智能电网方面的研究工作。E-mail：haoxust@163.com
- 基金信息：
  
  国家自然科学基金资助项目(51804250);中国博士后科学基金资助项目(2019M653874XB;2020M683522);陕西省科技计划项目(2021JQ-572;2020JQ-757);陕西省教育厅科研计划项目(18JK0512;21JK0769);西安市碑林区科技计划项目(GX2116)
- DOI：10.37188/OPE.20223019.2390 中图分类号： TP391.41;TN219
- 纸质出版日期：2022-10-10，
  
  收稿日期：2022-06-17，
  
  修回日期：2022-07-05，
扫描看全文
引用本文
郝帅,何田,马旭等.动态特征优化机制下的跨尺度红外行人检测[J].光学精密工程,2022,30(19):2390-2403.

HAO Shuai,HE Tian,MA Xu,et al.Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism[J].Optics and Precision Engineering,2022,30(19):2390-2403.
郝帅,何田,马旭等.动态特征优化机制下的跨尺度红外行人检测[J].光学精密工程,2022,30(19):2390-2403. DOI： 10.37188/OPE.20223019.2390.

HAO Shuai,HE Tian,MA Xu,et al.Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism[J].Optics and Precision Engineering,2022,30(19):2390-2403. DOI： 10.37188/OPE.20223019.2390.

动态特征优化机制下的跨尺度红外行人检测 Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism first-author 郝帅（1986-），男，博士，硕士生导师，主要从事人工智能、智能电网方面的研究工作。E-mail：haoxust@163.com http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037178&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037183&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037181&type= 22.01332855 32.00399780 郝帅（1986-），男，博士，硕士生导师，主要从事人工智能、智能电网方面的研究工作。E-mail： haoxust@163.com haoxust@163.com corresp E-mail：414548542@qq.com E-mail：414548542@qq.com 414548542@qq.com 西安科技大学电气与控制工程学院，陕西西安 710054 西安科技大学电气与控制工程学院，陕西西安 710054 College of Electrical and Control Engineering， Xi'an University of Science and Technology， Xi'an 710054， China College of Electrical and Control Engineering， Xi'an University of Science and Technology， Xi'an 710054， China 针对红外行人图像中待检测目标存在多尺度及部分遮挡导致传统算法难以准确检测的问题，提出一种动态特征优化机制下的跨尺度红外行人检测算法。为解决复杂环境中行人目标特征难以有效表达进而造成目标检测精度低的问题，提出一种动态特征优化机制，通过设计亮度感知模块及EG-Chimp优化模型在增强输入图像局部对比度的同时抑制背景信息；搭建了CSPDarkNet特征提取网络，并在其基础上构建CSFF-BiFPN特征金字塔结构以及跨尺度特征融合模块，以提高检测网络对多尺度及部分遮挡行人目标的检测精度；为进一步精确定位行人目标，引入CIOU损失函数加速网络收敛，从而提升检测性能。选取9种经典检测算法在KAIST数据集上进行对比测试，实验结果表明，本文算法能够对复杂环境中的多尺度及部分遮挡红外行人目标进行准确检测，检测精度可达90.7%，验证了所提出检测网络的优势。 Multi-scale and partial occlusions in infrared pedestrian images for target detection make it difficult for traditional algorithms to achieve accurate detection. This study developed a cross-scale infrared pedestrian detection algorithm based on a dynamic feature optimization mechanism. First， to alleviate the limitation that pedestrian target features are difficult to express effectively in complex environments， which results in low target detection accuracy， a dynamic feature optimization mechanism is presented. The luminance perception module and EG-Chimp optimization model are designed to enhance the local contrast of the input image and suppress background information. Second， the CSPdarknet53 structure is utilized as the backbone feature extraction network. Accordingly， a CSFF-BiFPN feature pyramid structure and cross-scale feature fusion module are constructed to improve the detection accuracy of multi-scale and partially occluded pedestrian targets. Finally， the CIOU loss function is introduced to accelerate network convergence rate and improve detection performance to locate pedestrian targets more accurately. To verify the advantages of the proposed detection network， nine classical detection algorithms are selected as baseline methods and tested on KAIST datasets. Experimental results demonstrate that the proposed algorithm can accurately detect multi-scale and partially occluded infrared pedestrian targets in complex environments， with detection accuracies of up to 90.7 %. 红外行人检测跨尺度动态特征优化亮度感知特征融合 infrared pedestrian detection cross scale dynamic feature optimization luminance perception feature fusion 1 引言 1 行人检测旨在利用计算机视觉方法判断图像或者视频序列中是否存在行人目标并给予精确定位［ 1 1 ］。目前，该技术已广泛应用于目标跟踪［ 2 2 ］、车辆辅助驾驶［ 3 3 ］、人体行为分析［ 4 4 ］等领域。由于热红外传感器受光照变化及天气条件的影响小，具有较强的抗干扰能力，能够满足全天候检测任务需求［ 5 5 ］。因此，基于热红外图像的行人检测（简称红外行人检测）具有重要的研究意义，也成为行人检测领域的研究热点。传统红外行人检测算法［ 6 6 ］主要是利用滑窗技术实现人工特征提取并结合分类判别器完成行人检测任务。该算法虽然在某些特定场景下能够实现目标的准确检测，但难以捕捉红外行人目标的高级语义信息，在复杂环境下泛化能力欠佳。相较于传统的目标检测算法，基于深度学习的红外行人检测算法利用卷积神经网络出色的特征提取和非线性拟合能力自动学习目标特征，在检测精度、速度以及泛化能力上都表现更好。该类算法通常可分为两类：两阶段（two-stage）检测和一阶段（one-stage）检测。快速区域卷积神经网络（Faster Region-based Convolutional Neural Network，Faster R-CNN）［ 7 7 ］作为two-stage的代表算法之一，首先寻找候选区域，然后在候选区域上对检测结果进行分类回归，能够达到较高的检测精度，但因主干特征提取网络仅提取单层特征图，对红外行人目标的特征表达能力不足，导致密集多目标的检测性能受限。为此，在Faster R-CNN检测网络基础上，刘琼［ 8 8 ］等考虑到红外图像低清晰度对检测精度的不利影响，设计了膨胀最大值滤波器以增强红外行人目标显著性，进而提高检测精度。Chen ［ 9 9 ］等针对红外行人目标多尺度特性，构建区域分解分支，结合多区域特征实现了跨尺度红外行人检测任务。Xu ［ 10 10 ］等为降低检测网络误检率，基于视点几何约束原理，提出地面上下文聚合网络，从而剔除远离地面区域的红外行人误检目标。然而，two-stage红外行人检测算法易丢失部分遮挡目标空间信息，且候选框之间存在大量冗余，算法计算成本较高，往往难以完成复杂环境下的实时检测任务。 One-stage算法基于回归思想，直接利用检测网络预测行人目标类别概率并定位坐标，大大提高了检测速率。Pei ［ 11 11 ］等设计了一种基于Retinanet的目标检测器以融合多层语义信息，从而提高红外图像中小尺度行人目标的检测能力。Wang ［ 12 12 ］等结合短路连接和多重空洞卷积，使检测算法在抑制网络退化现象的同时扩大检测感受野，进而增强检测算法对红外行人检测目标的特征表达能力。李经宇等［ 13 13 ］针对复杂环境下多尺度目标显著度低的问题，引入注意力模型使检测网络在多层特征融合分支上执行权重分配，从而提高了行人目标显著度。刘怡帆等［ 14 14 ］结合CLAHE（Contrast Limited Adaptive Histogram Equalization）算法及YOLO（You Only Look Once）检测网络，通过提高红外图像中待检测目标的对比度，提升检测精度。Liu等［ 15 15 ］在YOLO检测网络基础上构建anchor-based和anchor-free协同预测检测头，通过设置权重参数在线选择损失函数来提高检测网络对密集小尺度红外行人目标的检测能力。然而，上述算法并未测试多目标相互遮挡或重叠区域时的检测性能。综上所述，one-stage算法相比于two-stage算法能够更好地兼顾检测精度和速度。然而，受红外传感器固有特性限制及复杂环境干扰，one-stage算法检测行人目标时依然存在以下问题：红外传感器受热灵敏性限制，行人目标纹理特征较弱且存在边缘模糊现象，易淹没于冗余背景中，从而导致红外待检测目标特征难以有效表达，影响检测网络性能；在密集多目标场景下，红外行人目标存在多尺度及部分遮挡，进而导致红外行人特征形变甚至丢失。针对上述问题，本文提出一种动态特征优化机制下的红外行人检测算法（Cross-scale Detection Network based on Dynamic Feature Optimization Mechanism，DFOM-CSNet）。首先，设计了亮度感知模块（Luminance Perception Module，LPM）对输入图像进行亮度特征提取；接着设计了EG-Chimp（Chimp optimization with Energy loss and Gradient variation）优化模型对构建的目标函数进行迭代寻优，从而在增强输入图像局部对比度和细节表达能力的同时抑制背景信息干扰。为增强检测网络对多尺度及部分遮挡行人目标的检测性能，提出了一种CSFF-BiFPN（Bi-Directional Feature Pyramid Network Based on Cross Scale Feature Fusion）结构。该结构采用尺度跳跃连接的方式，增强颈部网络特征聚合能力。同时，构建跨尺度特征融合模块（Cross-scale Feature Fusion，CSFF）来学习不同尺度特征之间的权重参数，从而改善多尺度以及部分遮挡情况下的行人检测效果。为精确定位检测目标，在原网络中引入完全交并比（Complete Intersection Over Union，CIOU）［ 16 16 ］回归损失函数，通过计算真实框与预测框的重叠面积、中心点距离及长宽比，加速算法收敛，进而提高检测精度。 2 DFOM-CSNet红外行人检测算法 1 针对复杂环境下行人尺度多样性以及图像对比度低等特性造成红外行人目标难以准确定位的问题，本文在YOLOv5特征提取网络CSPDarkNet ［ 17 17 ］的基础上，提出DFOM-CSNet算法，其框架如图1 图1 所示，由DFOM、主干网络、颈部网络及预测层构成。其中，DFOM机制旨在增强输入图像的对比度及细节表达能力，由LPM模块和EG-Chimp优化模型两部分组成。LPM模块对红外图像进行亮度特征提取，同时EG-Chimp优化模型对构建的目标函数进行迭代寻优，从而增强红外行人目标的对比度和细节表达能力并抑制背景信息干扰。主干网络由Focus和CSPNet（Cross Stage Partial Network）两部分构成。其中，Focus切片模块不仅扩大了网络感受野，还能够有效抑制图像特征信息损失，从而加快训练速度。CSPNet结构旨在解决网络优化过程中梯度信息重复进而造成计算成本过高的问题。颈部网络借鉴CSPNet构造了CSP2_X结构以加强网络特征融合能力，并设计CSFF-BiFPN特征金字塔结构通过尺度跳跃连接方式及跨尺度特征融合模块，改善多尺度以及部分遮挡情况下的行人检测效果。预测层通过引入CIOU回归损失函数提高训练过程中预测框的回归速度和精度。 3 动态特征优化机制 1 受红外传感器热辐射成像机理的影响，红外图像相较于可见光图像分辨率往往较低，细节分辨能力较弱。针对该问题，本文提出一种DFOM机制，它利用亮度感知模块将输入图像分解为亮度特征图像及红外背景图像，并设计EG-Chimp优化模型对构建的目标函数进行迭代寻优，从而在增强输入图像局部对比度和细节表达能力的同时抑制背景信息干扰。算法1 ：动态特征优化机制 Input：红外图像 <math id="M1"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037198&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037195&type= 2.37066650 3.21733332 ，最大优化迭代次数 <math id="M2"><mi mathvariant="normal">M</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">x</mi><mo>_</mo><mi mathvariant="normal">i</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037205&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037201&type= 18.03400040 2.79399991 ，寻优参数 <math id="M3"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037247&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037244&type= 1.69333339 2.37066650 ，搜索种群 <math id="M4"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">A</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">k</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037266&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037264&type= 8.72066593 3.21733332 ， <math id="M5"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037269&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037268&type= 7.45066643 3.21733332 ， <math id="M6"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037273&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037271&type= 7.45066643 3.21733332 和 <math id="M7"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">D</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037278&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037276&type= 7.11199999 3.21733332 . 1. LPM模块构造四叉树-贝塞尔插值算子重构初始红外背景图像；引入引导滤波平滑噪音并得到红外亮度特征图像 <math id="M8"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">L</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037233&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037231&type= 3.47133350 3.21733332 和红外背景图像 <math id="M9"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037237&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037235&type= 3.47133350 3.21733332 ； 2. EG-Chimp优化模型构建动态特征优化图像： <math id="M10"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>=</mo><mi>α</mi><mo>×</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">L</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>+</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037241&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037239&type= 24.55333328 3.30200005 ；设计目标函数对参数 <math id="M11"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037247&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037244&type= 1.69333339 2.37066650 寻优： <math id="M12"><mi>F</mi><mo>=</mo><mi mathvariant="normal">m</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">n</mi><mo>{</mo><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">S</mi><mi mathvariant="normal">F</mi></mrow></msub><mo>+</mo><mi>λ</mi><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi></mrow></msub><mo>}</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037252&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037249&type= 31.58066559 3.89466691 ； While <math id="M13"><mi>t</mi><mo><</mo><mi mathvariant="normal">M</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">x</mi><mo>_</mo><mi mathvariant="normal">i</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037257&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037255&type= 23.11400032 2.79399991 For each chimp 计算各人猿种群的位置向量；更新 <math id="M14"><mi>f</mi><mo>,</mo><mi>m</mi><mo>,</mo><mi>c</mi><mo>,</mo><mi>a</mi><mo>,</mo><mi>D</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037262&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037260&type= 15.40933418 3.13266683 ； End For For each search chimp 更新目前搜索种群的位置向量； End For 更新 <math id="M15"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">A</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">k</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037266&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037264&type= 8.72066593 3.21733332 ， <math id="M16"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037269&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037268&type= 7.45066643 3.21733332 ， <math id="M17"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037273&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037271&type= 7.45066643 3.21733332 和 <math id="M18"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">D</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037278&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037276&type= 7.11199999 3.21733332 ； <math id="M19"><mi>t</mi><mo>=</mo><mi>t</mi><mo>+</mo><mn>1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037282&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037279&type= 11.76866722 2.53999996 ； End While Output：动态特征优化图像 <math id="M20"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037286&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037284&type= 3.64066648 3.21733332 3.1　LPM模块 2 为解决红外行人目标特征难以准确表达的问题，设计LPM模块。首先，构建四叉树-贝塞尔插值算子，通过调节部分采样控制点来恢复大规模像素矩阵，进而得到初始背景图像。然后，为避免由于边缘控制点采样不均而造成图像块效应，利用引导滤波的局部线性拟合特性平滑图像边缘信息，最终得到重构背景图像和亮度特征图像，为后续动态特征优化奠定基础。 3 . 1 . 1　四叉树-贝塞尔插值 3 为准确重构红外背景图像，本文引入四叉树分解方法［ 18 18 ］，通过尽可能多地采样除红外亮度区域外的控制点来估计背景轮廓信息。具体分解步骤如下：（1）调整原始红外图像尺寸为 <math id="M21"><mn mathvariant="normal">512</mn><mo>×</mo><mn mathvariant="normal">512</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037295&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037293&type= 16.00200081 2.79399991 ，且将其分解为4个等尺寸的一级图像块；（2）逐块比较各图像块内的最大灰度 <math id="M22"><msub><mrow><mi>p</mi></mrow><mrow><mi mathvariant="normal">m</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">x</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037299&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037297&type= 5.84200001 4.57200003 和最小灰度 <math id="M23"><msub><mrow><mi>p</mi></mrow><mrow><mi mathvariant="normal">m</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">n</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037303&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037301&type= 5.50333309 4.57200003 ，若满足分解条件： <math display="block" id="M24"><msub><mrow><mi>p</mi></mrow><mrow><mi mathvariant="normal">m</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">x</mi></mrow></msub><mo>-</mo><msub><mrow><mi>p</mi></mrow><mrow><mi mathvariant="normal">m</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">n</mi></mrow></msub><mo>></mo><mi mathvariant="normal">t</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">l</mi><mi mathvariant="normal">d</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037308&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037305&type= 34.54399872 4.57200003 ，式中： <math id="M25"><mi mathvariant="normal">t</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">l</mi><mi mathvariant="normal">d</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037312&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037310&type= 13.97000027 2.79399991 表示分解阈值，这里取10。则进一步将该图像块分解为4个二级图像块。（3）重复步骤（2），直至所有图像块均不满足公式（1）公式（1），则四叉树分解结束。通过四叉树分解算法可将输入图像分割为多个图像块，从各图像块中均匀采样16个控制点后，利用贝塞尔插值运算子重建红外背景图像，即： <math display="block" id="M26"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">b</mi></mrow></msub><mo stretchy="false">(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo stretchy="false">)</mo><mo>=</mo><mi mathvariant="bold-italic">X</mi><mi mathvariant="bold-italic">M</mi><mi mathvariant="bold-italic">P</mi><msup><mrow><mi mathvariant="bold-italic">M</mi></mrow><mrow><mi mathvariant="normal">T</mi></mrow></msup><msup><mrow><mi mathvariant="bold-italic">Y</mi></mrow><mrow><mi mathvariant="normal">T</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037316&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037314&type= 35.47533035 5.16466665 ，式中： <math id="M27"><mo stretchy="false">(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037321&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037318&type= 9.05933285 3.97933316 表示插值采样点； <math id="M28"><mi mathvariant="bold-italic">P</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037325&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037323&type= 2.70933342 2.79399991 表示由16个采样控制点组成的 <math id="M29"><mn mathvariant="normal">4</mn><mo>×</mo><mn mathvariant="normal">4</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037340&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037339&type= 8.63599968 2.79399991 矩阵； <math id="M30"><mi mathvariant="bold-italic">X</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037333&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037330&type= 3.13266683 2.79399991 和 <math id="M31"><mi mathvariant="bold-italic">Y</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037337&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037335&type= 3.13266683 2.79399991 分别为由 <math id="M32"><mn mathvariant="normal">4</mn><mo>×</mo><mn mathvariant="normal">4</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037340&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037339&type= 8.63599968 2.79399991 均匀采样控制点的行列坐标值构成的矢量； <math id="M33"><mi mathvariant="bold-italic">M</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037342&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037341&type= 3.97933316 2.96333337 表示常系数插值矩阵。各参数的具体取值如下： <math display="block" id="M34"><mfenced open="{" close="" separators="|"><mrow><mtable><mtr><mtd><mi mathvariant="bold-italic">X</mi><mo>=</mo><mo stretchy="false">[</mo><msup><mrow><mi>x</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup><mo>,</mo><msup><mrow><mi>x</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup><mo>,</mo><msup><mrow><mi>x</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msup><mo>,</mo><msup><mrow><mi>x</mi></mrow><mrow><mn mathvariant="normal">0</mn></mrow></msup><mo stretchy="false">]</mo></mtd></mtr><mtr><mtd><mi mathvariant="bold-italic">Y</mi><mo>=</mo><mo stretchy="false">[</mo><msup><mrow><mi>y</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup><mo>,</mo><msup><mrow><mi>y</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup><mo>,</mo><msup><mrow><mi>y</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msup><mo>,</mo><msup><mrow><mi>y</mi></mrow><mrow><mn mathvariant="normal">0</mn></mrow></msup><mo stretchy="false">]</mo></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037344&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037343&type= 31.32666588 10.41399956 ， <math display="block" id="M35"><mi mathvariant="bold-italic">M</mi><mo>=</mo><mfenced open="[" close="]" separators="|"><mrow><mtable><mtr><mtd><mn mathvariant="normal">1</mn></mtd><mtd><mn mathvariant="normal">0</mn></mtd><mtd><mn mathvariant="normal">0</mn></mtd><mtd><mn mathvariant="normal">0</mn></mtd></mtr><mtr><mtd><mo>-</mo><mn mathvariant="normal">3</mn></mtd><mtd><mn mathvariant="normal">3</mn></mtd><mtd><mn mathvariant="normal">0</mn></mtd><mtd><mn mathvariant="normal">0</mn></mtd></mtr><mtr><mtd><mn mathvariant="normal">3</mn></mtd><mtd><mo>-</mo><mn mathvariant="normal">6</mn></mtd><mtd><mn mathvariant="normal">3</mn></mtd><mtd><mn mathvariant="normal">1</mn></mtd></mtr><mtr><mtd><mo>-</mo><mn mathvariant="normal">1</mn></mtd><mtd><mn mathvariant="normal">3</mn></mtd><mtd><mo>-</mo><mn mathvariant="normal">3</mn></mtd><mtd><mn mathvariant="normal">1</mn></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037346&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037345&type= 40.38600159 18.45733452 . 3 . 1 . 2　引导滤波 3 为解决采样过程中各图像块边缘区域控制点因选取不一导致重构背景图像出现块效应的问题，本文采用引导滤波［ 19 19 ］对初始背景图像进行平滑处理，即： <math display="block" id="M36"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>=</mo><msub><mrow><mi>a</mi></mrow><mrow><mi>k</mi></mrow></msub><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">b</mi></mrow></msub><mo>+</mo><msub><mrow><mi>b</mi></mrow><mrow><mi>k</mi></mrow></msub><mo>,</mo><mtext> </mtext><mo>∀</mo><mi>i</mi><mo>∈</mo><msub><mrow><mi>w</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037348&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037347&type= 36.91466904 4.23333359 ，式中： <math id="M37"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037350&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037349&type= 3.97933316 3.72533321 表示红外背景图像； <math id="M38"><msub><mrow><mi>w</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037375&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037374&type= 3.97933316 3.72533321 表示以像素 <math id="M39"><mi>k</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037354&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037353&type= 1.69333339 2.79399991 为中心的图像块； <math id="M40"><msub><mrow><mi>a</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037388&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037387&type= 2.96333337 3.72533321 和 <math id="M41"><msub><mrow><mi>b</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037359&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037358&type= 2.96333337 3.72533321 为两线性参数，具体取值如下： <math display="block" id="M42"><msub><mrow><mi>a</mi></mrow><mrow><mi>k</mi></mrow></msub><mo>=</mo><mfrac><mrow><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mfenced open="|" close="|" separators="|"><mrow><mi>w</mi></mrow></mfenced></mrow></mfrac><munder><mo movablelimits="false">∑</mo><mrow><mi>i</mi><mo>∈</mo><msub><mrow><mi>w</mi></mrow><mrow><mi>k</mi></mrow></msub></mrow></munder><mrow><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">i</mi><msub><mrow><mi mathvariant="normal">r</mi></mrow><mrow><mi>i</mi></mrow></msub></mrow></msub></mrow><msub><mrow><mi>I</mi></mrow><mrow><msub><mrow><mi mathvariant="normal">b</mi></mrow><mrow><mi>i</mi></mrow></msub></mrow></msub><mo>-</mo><msub><mrow><mi>μ</mi></mrow><mrow><mi>k</mi></mrow></msub><mover><mrow><msub><mrow><mi>I</mi></mrow><mrow><msub><mrow><mi mathvariant="normal">b</mi></mrow><mrow><mi>k</mi></mrow></msub></mrow></msub></mrow><mo>¯</mo></mover></mrow><mrow><msubsup><mrow><mi>σ</mi></mrow><mrow><mi>k</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msubsup><mo>+</mo><mi>ε</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037365&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037363&type= 38.18466568 13.97000027 ， <math display="block" id="M43"><msub><mrow><mi>b</mi></mrow><mrow><mi>k</mi></mrow></msub><mo>=</mo><mover><mrow><msub><mrow><mi>I</mi></mrow><mrow><msub><mrow><mi mathvariant="normal">b</mi></mrow><mrow><mi>k</mi></mrow></msub></mrow></msub></mrow><mo>¯</mo></mover><mo>-</mo><msub><mrow><mi>a</mi></mrow><mrow><mi>k</mi></mrow></msub><msub><mrow><mi>μ</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037367&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037366&type= 22.01333237 4.23333359 ，式中： <math id="M44"><msub><mrow><mi>μ</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037369&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037368&type= 3.21733332 4.57200003 和 <math id="M45"><msubsup><mrow><mi>σ</mi></mrow><mrow><mi>k</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037371&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037370&type= 3.21733332 3.97933316 分别表示引导图像 <math id="M46"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037373&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037372&type= 2.70933342 3.72533321 在 <math id="M47"><msub><mrow><mi>w</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037375&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037374&type= 3.97933316 3.72533321 中的像素均值和方差； <math id="M48"><mover><mrow><msub><mrow><mi>I</mi></mrow><mrow><msub><mrow><mi mathvariant="normal">b</mi></mrow><mrow><mi>k</mi></mrow></msub></mrow></msub></mrow><mo>¯</mo></mover></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037377&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037376&type= 3.04800010 4.23333359 代表 <math id="M49"><msub><mrow><mi>I</mi></mrow><mrow><msub><mrow><mi mathvariant="normal">b</mi></mrow><mrow><mi>k</mi></mrow></msub></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037379&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037378&type= 3.13266683 3.64066648 的像素均值； <math id="M50"><mfenced open="|" close="|" separators="|"><mrow><mi>w</mi></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037382&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037381&type= 5.16466665 3.97933316 为 <math id="M51"><msub><mrow><mi>w</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037384&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037383&type= 3.97933316 3.72533321 图像块内包含的像素点个数； <math id="M52"><mi>ε</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037386&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037385&type= 1.43933344 2.79399991 表示正则化参数，以避免 <math id="M53"><msub><mrow><mi>a</mi></mrow><mrow><mi>k</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037388&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037387&type= 2.96333337 3.72533321 取值过大。最后，利用线性相减算子可提取红外亮度特征图像 <math id="M54"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">L</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037391&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037389&type= 3.97933316 3.72533321 ，如式（8）式（8）所示： <math display="block" id="M55"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">L</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>=</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>-</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037394&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037392&type= 20.06599998 3.72533321 . LPM模块结构如图2 图2 所示。 3.2　EG-Chimp优化模型 2 为抑制复杂背景对行人目标准确检测的影响，本文提出EG-Chimp优化模型，通过迭代寻优生成具有局部高对比度的动态特征优化图像 <math id="M56"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037405&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037403&type= 4.14866638 3.72533321 ，如式（9）式（9）所示： <math display="block" id="M57"><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>=</mo><mi>α</mi><mo>×</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">L</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>+</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037409&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037407&type= 28.27866554 3.80999994 ，式中 <math id="M58"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037451&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037449&type= 1.94733346 2.79399991 为寻优参数。 3 . 2 . 1　目标函数构建 3 本文构建的目标函数旨在兼顾动态特征优化图像的梯度变化和能量损失，分别用空间频率 <math id="M59"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">S</mi><mi mathvariant="normal">F</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037417&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037415&type= 5.07999992 3.72533321 和内容损失 <math id="M60"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037420&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037418&type= 7.02733374 3.72533321 表示。空间频率（Spatial Frequency， SF）反映图像空间突变如边缘的变化情况，用于度量图像的梯度分布，图像SF值越高，图像中人物目标越清晰［ 20 20 ］。其计算公式如下： <math display="block" id="M61"><mtable><mtr><mtd><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">S</mi><mi mathvariant="normal">F</mi></mrow></msub><mo>=</mo><mo>-</mo><mo stretchy="false">(</mo><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>i</mi><mo>=</mo><mn mathvariant="normal">1</mn></mrow><mrow><mi mathvariant="normal">M</mi></mrow></munderover></mstyle><mrow/><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>j</mi><mo>=</mo><mn mathvariant="normal">1</mn></mrow><mrow><mi>N</mi></mrow></munderover></mstyle><mrow><mo stretchy="false">(</mo></mrow><msup><mrow><mfenced open="|" close="|" separators="|"><mrow><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>-</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>-</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo></mrow></mfenced></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup><mo>+</mo></mtd></mtr><mtr><mtd><msup><mrow><mfenced open="|" close="|" separators="|"><mrow><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>-</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo stretchy="false">(</mo><mi>i</mi><mo>-</mo><mn mathvariant="normal">1</mn><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo></mrow></mfenced></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup><msup><mrow><mo stretchy="false">)</mo><mo stretchy="false">)</mo></mrow><mrow><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mn mathvariant="normal">2</mn></mrow></mfrac></mrow></msup><mo>,</mo></mtd></mtr></mtable></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037424&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037422&type= 68.07199860 17.27199936 式中： <math id="M62"><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037427&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037426&type= 7.11199999 3.97933316 表示图像的横纵坐标。为避免在优化过程中由于行人目标亮度特征过度增强而导致其细节信息无法有效保留的问题，利用 <math id="M63"><msub><mrow><mi>l</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037431&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037429&type= 2.37066650 3.72533321 函数构建内容损失函数，即： <math display="block" id="M64"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi></mrow></msub><mo>=</mo><msubsup><mrow><mfenced open="‖" close="‖" separators="|"><mrow><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">O</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub><mo>-</mo><msub><mrow><mi>I</mi></mrow><mrow><mi mathvariant="normal">i</mi><mi mathvariant="normal">r</mi></mrow></msub></mrow></mfenced></mrow><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037434&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037433&type= 28.44799995 5.24933338 . 综上所述，为了增强原始红外图像中行人亮度特征的同时兼顾细节信息，构造的目标函数为： <math display="block" id="M65"><mi>F</mi><mo>=</mo><mi mathvariant="normal">m</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">n</mi><mfenced separators="|"><mrow><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">S</mi><mi mathvariant="normal">F</mi></mrow></msub><mo>+</mo><mi>λ</mi><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi></mrow></msub></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037439&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037438&type= 35.64466476 4.06400013 ，式中： <math id="M66"><mi>λ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037443&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037441&type= 1.77800000 2.79399991 为保持两项平衡的正则化参数，这里 <math id="M67"><mi>λ</mi><mo>=</mo><mrow><mrow><mn mathvariant="normal">1</mn></mrow><mo>/</mo><mrow><mn mathvariant="normal">500</mn></mrow></mrow></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037447&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037445&type= 15.15533447 3.30200005 。 3 . 2 . 2　目标函数自适应优化 3 元启发式优化算法由于其简单、灵活、高效等特点，已成为解决全局优化问题的主要方法［ 21 21 ］。然而，随着问题复杂性的增加，该算法易出现收敛速度较慢，陷入局部最优解等问题［ 22 22 ］。因此，针对目标函数式（12）式（12），本文引入人猿优化算法［ 23 23 ］，通过模拟人猿捕食和混乱抢食2种状态，加快算法收敛，进而求解得到寻优参数 <math id="M68"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037451&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037449&type= 1.94733346 2.79399991 ，在保证动态特征优化图像空间频率的同时抑制其内容损失。该算法将人猿划分为攻击型 <math id="M69"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">A</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">k</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037457&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037454&type= 9.82133293 3.72533321 、驱赶型 <math id="M70"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037461&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037459&type= 8.46666718 3.72533321 、拦截型 <math id="M71"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">B</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037466&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037464&type= 8.46666718 3.72533321 和追逐型 <math id="M72"><msub><mrow><mi>X</mi></mrow><mrow><mi mathvariant="normal">D</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037471&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037468&type= 8.04333401 3.72533321 4类。狩猎过程中，人猿可根据猎物的位置相应地改变自身位置，如式（13）式（13）所示： <math display="block" id="M73"><mtable><mtr><mtd><mi mathvariant="bold-italic">D</mi><mo>=</mo><mfenced open="|" close="|" separators="|"><mrow><mi mathvariant="bold-italic">c</mi><mo>⋅</mo><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">p</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">y</mi></mrow></msub><mo stretchy="false">(</mo><mi>t</mi><mo stretchy="false">)</mo><mo>-</mo><mi>m</mi><mo>⋅</mo><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">c</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">m</mi><mi mathvariant="normal">p</mi></mrow></msub><mo stretchy="false">(</mo><mi>t</mi><mo stretchy="false">)</mo></mrow></mfenced></mtd></mtr><mtr><mtd><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">c</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">m</mi><mi mathvariant="normal">p</mi></mrow></msub><mo stretchy="false">(</mo><mi>t</mi><mo>+</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo><mo>=</mo><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">p</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">y</mi></mrow></msub><mo stretchy="false">(</mo><mi>t</mi><mo stretchy="false">)</mo><mo>-</mo><mi mathvariant="bold-italic">a</mi><mo>⋅</mo><mi mathvariant="bold-italic">D</mi></mtd></mtr></mtable></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037477&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037474&type= 47.24399948 10.58333302 ，式中： <math id="M74"><mi>t</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037482&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037480&type= 1.26999998 2.79399991 表示当前的迭代次数， <math id="M75"><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">p</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">y</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037489&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037486&type= 6.51933336 3.72533321 为猎物的位置向量， <math id="M76"><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">c</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">m</mi><mi mathvariant="normal">p</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037494&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037492&type= 7.87400007 3.72533321 为人猿的位置向量， <math id="M77"><mi mathvariant="bold-italic">a</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037500&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037498&type= 1.94733346 2.79399991 ， <math id="M78"><mi mathvariant="bold-italic">c</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037508&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037505&type= 1.60866666 2.79399991 和 <math id="M79"><mi mathvariant="bold-italic">m</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037513&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037510&type= 3.04800010 2.79399991 为系数向量，具体取值如下： <math display="block" id="M80"><mfenced open="{" close="" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><mi mathvariant="bold-italic">a</mi><mo>=</mo><mi>f</mi><mo stretchy="false">(</mo><mn mathvariant="normal">2</mn><msub><mrow><mi mathvariant="bold-italic">r</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>-</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo></mtd></mtr><mtr><mtd><mi mathvariant="bold-italic">c</mi><mo>=</mo><mn mathvariant="normal">2</mn><mo>⋅</mo><msub><mrow><mi mathvariant="bold-italic">r</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub></mtd></mtr><mtr><mtd><mi>m</mi><mo>=</mo><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037519&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037517&type= 25.14599991 14.64733315 ，式中： <math id="M81"><mi>f</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037525&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037522&type= 2.45533323 3.64066648 在迭代过程中由2.5非线性递减至0， <math id="M82"><msub><mrow><mi mathvariant="bold-italic">r</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037531&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037528&type= 2.79399991 3.72533321 和 <math id="M83"><msub><mrow><mi mathvariant="bold-italic">r</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037536&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037534&type= 2.79399991 3.72533321 是 <math id="M84"><mo stretchy="false">[</mo><mn mathvariant="normal">0,1</mn><mo stretchy="false">]</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037582&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037580&type= 8.97466660 3.55599999 内的随机向量； <math id="M85"><mi>m</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037548&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037544&type= 2.87866688 2.79399991 为一个基于高斯混沌映射得到的混沌向量 <math id="M86"><mi mathvariant="normal">c</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037554&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037551&type= 8.46666718 2.79399991 ，代表了人猿在狩猎过程中混乱抢食的社会现象，具体取值如下： <math display="block" id="M87"><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi><mo stretchy="false">(</mo><mi>θ</mi><mo>+</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo><mo>=</mo><mi mathvariant="normal">m</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">d</mi><mfenced separators="|"><mrow><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo></mrow></mfrac><mo>,</mo><mn mathvariant="normal">1</mn></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037560&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037557&type= 55.45666885 9.65200043 ，式中： <math id="M88"><mi mathvariant="normal">m</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">d</mi><mo stretchy="false">(</mo><mo>·</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037565&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037562&type= 11.34533310 3.55599999 为求余算子，且 <math id="M89"><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo><mo>=</mo><mn mathvariant="normal">0.7</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037570&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037568&type= 24.72266769 3.55599999 。猎物被包围后，由驱赶者、拦截者、追逐者辅助攻击者完成最终狩猎任务。为提高算法收敛速度，模型在迭代后期模拟了人猿混乱抢食过程，如下： <math display="block" id="M90"><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">c</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">m</mi><mi mathvariant="normal">p</mi></mrow></msub><mo stretchy="false">(</mo><mi>t</mi><mo>+</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo><mo>=</mo><mfenced open="{" close="" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><msub><mrow><mi mathvariant="bold-italic">X</mi></mrow><mrow><mi mathvariant="normal">p</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">y</mi></mrow></msub><mo stretchy="false">(</mo><mi>t</mi><mo stretchy="false">)</mo><mo>-</mo><mi>a</mi><mo>⋅</mo><mi mathvariant="bold-italic">D</mi><mo>,</mo><mtext> </mtext><mi>χ</mi><mo>≤</mo><mn mathvariant="normal">0.5</mn></mtd></mtr><mtr><mtd><mi mathvariant="normal">C</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi><mo>,</mo><mtext> </mtext><mi>χ</mi><mo>></mo><mn mathvariant="normal">0.5</mn></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037576&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037573&type= 65.02400208 10.49866676 ，式中 <math id="M91"><mi>χ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037578&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037577&type= 2.03200006 3.64066648 为 <math id="M92"><mo stretchy="false">[</mo><mn mathvariant="normal">0,1</mn><mo stretchy="false">]</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037582&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037580&type= 8.97466660 3.55599999 内的随机参数。狩猎过程中，一方面人猿根据攻击者、驱赶者、拦截者和追逐者位置更新位置，并攻击猎物；另一方面人猿通过分散活动寻找猎物完成全局搜索任务。最终，通过多次迭代可得到动态特征优化图像，如图3 图3 所示。通过动态特征优化机制可以有效增强红外图像亮度信息，相较于原始图像，优化后的图像具有更高的清晰度，为行人目标的准确检测奠定了基础。 4 基于CSFF-BiFPN的颈部网络 1 4.1　CSFF-BiFPN金字塔结构 2 原始Yolov5网络为提高不同尺度间特征聚合能力，利用FPN+PAN结构在自底向上传递强语义信息的同时，自顶向下传递强定位信息。然而，该结构将不同尺度特征图进行尺寸调整后直接相加聚合，无法充分利用输入端的跨尺度信息，进而影响最终的检测精度。为解决此问题，本文构建CSFF-BiFPN特征金字塔结构，通过尺度跳跃连接方式提高网络特征聚合的能力，并设计跨尺度特征融合模块，以利用可学习权重参数对多尺度特征进行权值分配，进而增强检测网络对多尺度及部分遮挡行人目标的检测性能，特征金字塔结构对比如图4 图4 所示。本文所构建的CSFF-BiFPN特征金字塔结构，通过删除仅有一条输入边的节点来减少网络参数量。同时，借鉴FPN-PAN 结构，从两个方向分别传递强语义和强定位信息。为进一步利用输入特征的跨尺度信息，通过尺度跳跃连接的方式提高网络特征聚合能力并结合跨尺度特征融合模块CSFF_2和CSFF_3对来自不同输入端的特征信息进行权重分配。 4.2　跨尺度特征融合模块 2 FPN-PAN结构对来自不同尺度的输入特征进行尺度调整后直接进行相加运算，易导致特征融合不足的问题。因此，本文设计了跨尺度特征融合模块CSFF_2和CSFF_3，自适应学习尺度不同的输入特征权重参数，模型结构如图5 图5 所示。将 <math id="M93"><mi mathvariant="normal">l</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">l</mi><mtext> </mtext><mi>i</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037651&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037650&type= 9.05933285 2.87866688 的输入特征记作 <math id="M94"><msub><mrow><mi>I</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037613&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037610&type= 2.11666679 3.72533321 ，其中 <math id="M95"><mi>i</mi><mo>∈</mo><mfenced open="[" close="]" separators="|"><mrow><mn mathvariant="normal">1,2</mn><mo>,</mo><mn mathvariant="normal">3</mn></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037617&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037615&type= 16.84866714 3.97933316 。首先，对不同尺度特征进行尺度调整，在上采样过程中，利用 <math id="M96"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037704&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037703&type= 8.63599968 2.79399991 的卷积调整通道数并引入最近邻插值提高分辨率，在下采样过程中，通过步长为2的 <math id="M97"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037625&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037623&type= 8.63599968 2.79399991 卷积完成 <math id="M98"><mrow><mrow><mn mathvariant="normal">1</mn></mrow><mo>/</mo><mrow><mn mathvariant="normal">2</mn></mrow></mrow></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037628&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037626&type= 4.91066647 3.21733332 比例的下采样，并通过在2步卷积之前添加步长为2的最大池化层完成 <math id="M99"><mrow><mrow><mn mathvariant="normal">1</mn></mrow><mo>/</mo><mrow><mn mathvariant="normal">4</mn></mrow></mrow></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037632&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037631&type= 4.91066647 3.21733332 比例的下采样；接着，利用 <math id="M100"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037636&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037634&type= 8.63599968 2.79399991 的卷积预估各输入特征的权重参数，并在保持权重参数之间相对关系的同时，加快算法收敛，利用softmax函数对各预估权重进行归一化处理；最后，对尺度调整后的输入特征进行权重分配得到融合后的输出特征，分别如下： <math display="block" id="M101"><mi mathvariant="normal">C</mi><mi mathvariant="normal">S</mi><mi mathvariant="normal">F</mi><mi mathvariant="normal">F</mi><mo>_</mo><mn mathvariant="normal">2</mn><mo stretchy="false">(</mo><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>,</mo><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo stretchy="false">)</mo><mo>=</mo><mi>φ</mi><mo>⋅</mo><msubsup><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mn mathvariant="normal">1</mn><mo>→</mo><mi>l</mi></mrow></msubsup><mo>+</mo><mi>γ</mi><mo>⋅</mo><msubsup><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mn mathvariant="normal">2</mn><mo>→</mo><mi>l</mi></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037640&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037638&type= 53.76333237 5.16466665 ， <math display="block" id="M102"><mi mathvariant="normal">C</mi><mi mathvariant="normal">S</mi><mi mathvariant="normal">F</mi><mi mathvariant="normal">F</mi><mo>_</mo><mn mathvariant="normal">3</mn><mo stretchy="false">(</mo><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>,</mo><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo>,</mo><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msub><mo stretchy="false">)</mo><mo>=</mo><mi>ψ</mi><mo>⋅</mo><msup><mrow><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub></mrow><mrow><mn mathvariant="normal">1</mn><mo>→</mo><mi>l</mi></mrow></msup><mo>+</mo><mi>δ</mi><mo>⋅</mo><msup><mrow><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub></mrow><mrow><mn mathvariant="normal">2</mn><mo>→</mo><mi>l</mi></mrow></msup><mo>+</mo><mi>η</mi><mo>⋅</mo><msup><mrow><msub><mrow><mi>I</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msub></mrow><mrow><mn mathvariant="normal">3</mn><mo>→</mo><mi>l</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037644&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037641&type= 71.20466614 5.07999992 ，式中： <math id="M103"><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi></mrow><mrow><mi>i</mi><mo>→</mo><mi>l</mi></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037648&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037646&type= 5.75733376 3.97933316 表示 <math id="M104"><mi mathvariant="normal">l</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">l</mi><mtext> </mtext><mi>i</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037651&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037650&type= 9.05933285 2.87866688 调整到 <math id="M105"><mi mathvariant="normal">l</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">l</mi><mtext> </mtext><mi>l</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037655&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037653&type= 9.22866726 2.87866688 后的特征信息， <math id="M106"><mi>φ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037659&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037657&type= 2.37066650 3.72533321 ， <math id="M107"><mi>γ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037663&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037661&type= 1.94733346 3.64066648 ， <math id="M108"><mi>ψ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037667&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037665&type= 2.11666679 3.55599999 ， <math id="M109"><mi>δ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037671&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037669&type= 1.94733346 2.79399991 和 <math id="M110"><mi>η</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037675&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037673&type= 1.86266661 3.64066648 表示各跨尺度特征通过soft-max函数归一化后得到的可学习权重参数，取值分别为： <math display="block" id="M111"><mi>φ</mi><mo>+</mo><mi>γ</mi><mo>=</mo><mn mathvariant="normal">1</mn><mo>,</mo><mtext> </mtext><mi>φ</mi><mo>=</mo><mfrac><mrow><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>φ</mi></mrow></msub></mrow></msup></mrow><mrow><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>φ</mi></mrow></msub></mrow></msup><mo>+</mo><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>γ</mi></mrow></msub></mrow></msup></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037679&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037678&type= 37.84600067 8.89000034 ， <math display="block" id="M112"><mi>ψ</mi><mo>+</mo><mi>δ</mi><mo>+</mo><mi>η</mi><mo>=</mo><mn mathvariant="normal">1</mn><mo>,</mo><mtext> </mtext><mi>ψ</mi><mo>=</mo><mfrac><mrow><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>ψ</mi></mrow></msub></mrow></msup></mrow><mrow><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>ψ</mi></mrow></msub></mrow></msup><mo>+</mo><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>δ</mi></mrow></msub></mrow></msup><mo>+</mo><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>λ</mi></mrow><mrow><mi>η</mi></mrow></msub></mrow></msup></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037684&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037682&type= 52.06999969 8.80533314 ，式中： <math id="M113"><msub><mrow><mi>λ</mi></mrow><mrow><mi>φ</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037688&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037686&type= 3.21733332 3.72533321 ， <math id="M114"><msub><mrow><mi>λ</mi></mrow><mrow><mi>γ</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037692&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037690&type= 3.04800010 3.72533321 ， <math id="M115"><msub><mrow><mi>λ</mi></mrow><mrow><mi>ψ</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037695&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037694&type= 3.04800010 3.72533321 ， <math id="M116"><msub><mrow><mi>λ</mi></mrow><mrow><mi>δ</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037699&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037697&type= 2.96333337 3.72533321 和 <math id="M117"><msub><mrow><mi>λ</mi></mrow><mrow><mi>η</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037702&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037701&type= 2.96333337 3.72533321 表示利用 <math id="M118"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037704&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037703&type= 8.63599968 2.79399991 卷积得到的预估权重参数。 5 实验结果与分析 1 实验中软硬件平台配置如表1 表1 所示。 Configuration Version parameter Operating system Microsoft Windows 10 GPU NVIDIA GeForce GTX 1660 Ti CPU Intel Core i5-10400F@2.90 GHz×6 CPUs CUDA 11.1 Deep learning framework Pytorch 5.1　实验数据集和网络模型训练 2 为验证本文所提出算法的优势，从KAIST数据集［ 24 24 ］中随机选取1 000张样本数据进行实验，行人目标共5 384个。其中，训练样本为700张，剩余300张图片作为测试数据集。在网络模型训练过程中，将输入图像分辨率统一调整为 <math id="M119"><mn mathvariant="normal">640</mn><mo>×</mo><mn mathvariant="normal">640</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037708&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037707&type= 16.00200081 2.79399991 ，并采用动量项为0.937的异步随机梯度下降法进行训练，在每一批训练的batch中包含32张图片，分29次送入训练网络。在前200轮将训练中的学习率设置为0.01，随着迭代轮数的增加，在后100轮迭代中将训练的学习率降到0.001。为了防止模型过拟合，将权重衰减正则项设置为5×10 - 3 ，同时选用CIOU损失函数［ 16 16 ］通过计算两框的重叠面积、中心点距离及宽高比提高预测框的回归速度，如下： <math display="block" id="M120"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">U</mi></mrow></msub><mo>=</mo><mi mathvariant="normal">I</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">U</mi><mo>-</mo><mfrac><mrow><msup><mrow><mi>ρ</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup><mo stretchy="false">(</mo><mi>b</mi><mo>,</mo><msup><mrow><mi>b</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup><mo stretchy="false">)</mo></mrow><mrow><msup><mrow><mi>C</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></mrow></mfrac><mo>-</mo><mi>ς</mi><mi>v</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037710&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037709&type= 47.24399948 8.46666718 ，式中： <math id="M121"><mi>b</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037714&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037712&type= 1.86266661 2.79399991 和 <math id="M122"><msup><mrow><mi>b</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037717&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037716&type= 3.47133350 3.04800010 分别表示预测框和目标框的中心点， <math id="M123"><mi>ρ</mi><mo stretchy="false">(</mo><mo>⋅</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037721&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037720&type= 5.16466665 3.89466691 为两框中心点间的欧式距离， <math id="M124"><mi>C</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037723&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037722&type= 2.62466669 2.79399991 表示覆盖预测框与目标框之间的最小矩形的斜距， <math id="M125"><mi>ς</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037729&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037728&type= 1.69333339 3.64066648 为限制预测框长宽比的权重参数， <math id="M126"><mi>v</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037731&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037730&type= 1.86266661 2.79399991 表示衡量预测框与真实框宽高比的一致性参数。 <math id="M127"><mi>ς</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037729&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037728&type= 1.69333339 3.64066648 和 <math id="M128"><mi>v</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037731&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037730&type= 1.86266661 2.79399991 分别为： <math display="block" id="M129"><mi>ς</mi><mo>=</mo><mfrac><mrow><mi>v</mi></mrow><mrow><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><mi mathvariant="normal">I</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">U</mi><mo stretchy="false">(</mo><mi>b</mi><mo>,</mo><msup><mrow><mi>b</mi></mrow><mrow><mi>g</mi><mi>t</mi></mrow></msup><mo stretchy="false">)</mo><mo stretchy="false">)</mo><mo>+</mo><mi>v</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037733&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037732&type= 40.89400101 8.63599968 ， <math display="block" id="M130"><mi>v</mi><mo>=</mo><mfrac><mrow><mn mathvariant="normal">4</mn></mrow><mrow><msup><mrow><mi mathvariant="normal">π</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></mrow></mfrac><msup><mrow><mfenced separators="|"><mrow><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">n</mi><mfrac><mrow><msup><mrow><mi>w</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></mrow><mrow><msup><mrow><mi>h</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></mrow></mfrac><mo>-</mo><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">n</mi><mfrac><mrow><mi>w</mi></mrow><mrow><mi>h</mi></mrow></mfrac></mrow></mfenced></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037735&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037734&type= 49.36066818 10.32933331 ，式中： <math id="M131"><mi>w</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037737&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037736&type= 2.96333337 2.79399991 ， <math id="M132"><mi>h</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037739&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037738&type= 1.94733346 2.79399991 ， <math id="M133"><msup><mrow><mi>w</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037741&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037740&type= 4.65666676 3.04800010 ， <math id="M134"><msup><mrow><mi>h</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037743&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=36037742&type= 3.64066648 3.04800010 分别表示预测框和目标框的宽和高。 5.2　图像对比度分析 2 为验证本文所提出动态特征优化机制的有效性，选取熵（Entropy）［ 25 25 ］，Brenner梯度函数［ 26 26 ］，离散余弦变换函数（Discrete Cosine Transform， DCT）［ 27 27 ］和方差函数（Variance）［ 26 26 ］作为客观评价指标。其中，Entropy表征了图像所含信息量的丰富程度，Brenner通过计算图像二阶梯度估计其清晰程度，DCT突出图像频率分布信息，Variance则从图像灰度变化的角度评估图像的聚焦程度。Entropy，Brenner，DCT和Variance数值越大，动态特征优化后图像对比度越高，特征表达能力越强。对KAIST数据集中随机选取1 000张样本数据进行测试验证，其定量分析结果如表2 表2 所示。 Image Entropy Brenner DCT Variance IR 5.65 4.72×10 6 4.69×10 3 1.06×10 8 Optimized IR 6.04 2.84×10 7 4.87×10 3 1.92×10 8 由表2 表2 可知，本文提出的动态特征优化算法的4种评价指标均为最优，从而证明原始图像经过动态特征优化后不仅包含更丰富的信息量，而且具有最强的对比度和最好的特征表达能力。 5.3　消融实验 2 为客观评估本文算法的检测性能，将DFOM-CSNet检测网络与原始YOLOv5检测框架进行对比，结果如图6 图6 所示。从图6 图6 （a）可以看出，YOLOv5检测网络约在75轮后达到平衡且最终稳定于0.018；而本文提出的DFOM-CSNet检测算法仅训练20轮损失值即下降至0.02，并最终稳定于0.009左右，取得了较好的训练效果。从图6 图6 （b）可以看出，YOLOv5检测网络最终的平均准确率大致为0.895，而本文算法的检测精度可达0.913。为了进一步验证各改进模块的作用，本文在原始YOLOv5算法框架基础上通过加入不同改进策略对测试集进行消融实验，所有实验均使用相同的数据样本和参数设置，对比结果如表3 表3 所示。 Component 8 Method GIOU √ √ √ √ CIOU √ √ √ √ DFOM √ √ √ √ YOLOv5 √ √ √ √ CSFF-BiFPN √ √ √ √ mAP@.5/% 88.3 88.8 89.4 89.1 90.2 89.8 89.5 90.7 从表3 表3 可知，相比于原始YOLOv5检测网络（损失函数为GIOU），引入CIOU损失函数后精度提升了0.5%，加入DFOM动态特征优化机制后精度提升了1.1%，构建CSFF-BiFPN跨尺度特征融合模块后精度提升了0.8%。为验证不同模块协同训练的优势，对CSFF-BiFPN网络下结合GIOU和DFOM的检测算法，YOLOv5网络下结合CIOU和DFOM的检测网络，及CSFF-BiFPN网络下基于CIOU的检测算法进行消融实验，结果表明，检测精度相较于原始网络分别提升了1.9%，1.5%及0.9%。最后，本文算法通过融合多种改进模块检测精度可达90.7%，较原始检测网络提升了2.4%。 5.4　对比实验分析 2 为客观评价本文算法的检测性能，选取Faster-RCNN ［ 7 7 ］、SSD ［ 28 28 ］、RetinaNet ［ 29 29 ］、Sparse R-CNN ［ 30 30 ］、VarifocalNet ［ 31 31 ］、TOOD ［ 32 32 ］、I-YOLO ［ 33 33 ］、YOLOv4-CLAHE ［ 13 13 ］和TC-Det ［34］ 9种检测算法，在部分行人遮挡场景、多尺度密集多目标场景、亮度过曝场景下与本文CSNet及DFOM-CSNet检测网络进行对比测试，结果如图7 图7 所示（彩图见期刊电子版）。其中，真实红外行人目标如原始红外图像中黄色实线框所示，为了便于观察和后续分析，各算法的漏检和误检区域已用红色虚线框进行标注。 1;2; 从第1组实验可以看出，除DFOM-CSNet检测算法外，其余几种对比算法在处理部分行人遮挡场景时均存在不同程度的漏检。从第2组实验可以看出，在复杂环境下Faster-RCNN、SSD、RetinaNet，及Sparse R-CNN算法存在较为严重的漏检问题，而其余5种检测框架对不同尺度间特征信息的聚合能力不足，难以实现多尺度行人目标的准确检测。本文提出的DFOM-CSNet检测算法通过设计CSFF-BiFPN特征金字塔结构，以增强网络特征聚合能力并构建跨尺度融合模块对不同尺度特征进行权重分配，有效提高了多尺度密集多目标场景下行人目标的检测性能。从第3组实验可以看出，在亮度过曝场景下，受日间过度光照影响，红外图像无法准确表达行人目标的细节信息，导致各种算法在检测过程中均出现较多漏检。本文提出的DFOM-CSNet检测算法通过设计动态特征优化机制可以有效增强红外图像局部对比度，相较于CSNet算法具有更好的检测性能。为定量评估本文检测网络的优势，对KAIST数据测试集进行对比实验，所有目标检测算法均使用相同的数据样本和参数设置，检测结果如表4 表4 所示。可以看出，本文算法在检测精度上明显优于其他9种对比算法。虽然引入动态特征优化机制后网络运行时间变长，但相比对比算法仍具有明显优势。 Method Backbone mAP@.5/% mAP@.5：.95/% Time/s Faster-RCNN ResNet50 72.9 — 0.163 SSD VGG 75.1 — 0.087 RetinaNet ResNet50 74.8 36.3 0.140 Sparse R-CNN ResNet50 84.3 42.9 0.148 VarifocalNet ResNet50 86.2 45.3 0.172 TOOD ResNet50 88.9 46.1 0.156 YOLOv4-CLAHE CSPDarkNet 87.9 45.9 0.042 I-YOLO DarkNet 88.9 45.8 0.024 TC-Det DarkNet 89.2 46.2 0.149 CSNet（Ours） Focus+ CSPDarkNet 89.5 46.2 0.011 DFOM-CSNet（Ours） Focus+ CSPDarkNet 90.7 46.8 0.013 6 结论 1 针对红外行人图像分辨率低、待检测目标存在多尺度及部分遮挡特性导致传统算法难以准确检测的问题，本文在CSPDarkNet特征提取网络的基础上提出了一种动态特征优化机制下的跨尺度红外行人检测算法。在KAIST数据集上进行行人目标检测实验，结果表明，本文设计的动态特征优化机制通过构建LPM模块和EG-Chimp优化模型，在抑制红外背景信息的同时能够增强行人目标的对比度和细节表达能力。然后，提出了CSFF-BiFPN特征金字塔结构，采用尺度跳跃连接方式有效增强了颈部网络对部分遮挡及多尺度行人目标的检测能力；并通过构建跨尺度特征融合模块，利用可学习权重参数对多尺度特征进行权值分配，有效解决了行人目标在多尺度密集多目标场景下由于特征表达能力不足而造成的漏检和误检问题。与9种经典检测算法相比，本文提出的DFOM-CSNet网络检测精度可达90.7%，在各类复杂场景下具有更好的检测效果，且能够满足实时性需求。但该算法对小尺度目标仍然存在漏检，下一阶段将针对这一问题展开研究。参考文献陈莹，朱宇 . 模态自适应权值学习机制下的多光谱行人检测网络［J］. 光学精密工程， 2020 ， 28 （ 12 ）： 2700 - 2709 . doi: 10.37188/OPE.20202812.2700 http://dx.doi.org/10.37188/OPE.20202812.2700 模态自适应权值学习机制下的多光谱行人检测网络 CHEN Y ， ZHU Y . Multispectral pedestrian detection network under modal adaptive weight learning mechanism ［J］. Optics and Precision Engineering ， 2020 ， 28 （ 12 ）： 2700 - 2709 . （in Chinese） . doi: 10.37188/OPE.20202812.2700 http://dx.doi.org/10.37188/OPE.20202812.2700 Multispectral pedestrian detection network under modal adaptive weight learning mechanism 杨琦，车进，张良，等 . GAN网络混合编码的行人再识别［J］. 液晶与显示， 2021 ， 36 （ 2 ）： 334 - 342 . doi: 10.37188/CJLCD.2020-0167 http://dx.doi.org/10.37188/CJLCD.2020-0167 GAN网络混合编码的行人再识别 YANG Q ， CHE J ， ZHANG L ， et al . Person re-identification of GAN network hybrid coding ［J］. Chinese Journal of Liquid Crystals and Displays ， 2021 ， 36 （ 2 ）： 334 - 342 . （in Chinese） . doi: 10.37188/CJLCD.2020-0167 http://dx.doi.org/10.37188/CJLCD.2020-0167 Person re-identification of GAN network hybrid coding LI T ， MA Y T ， SHEN H ， et al . FPGA implementation of real-time pedestrian detection using normalization-based validation of adaptive features clustering ［J］. IEEE Transactions on Vehicular Technology ， 2020 ， 69 （ 9 ）： 9330 - 9341 . doi: 10.1109/tvt.2020.2976958 http://dx.doi.org/10.1109/tvt.2020.2976958 FPGA implementation of real-time pedestrian detection using normalization-based validation of adaptive features clustering KANAZAWA A ， BLACK M J ， JACOBS D W ， et al . End-to-end recovery of human shape and pose ［C］. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition . Salt Lake City， UT， USA . IEEE ， 2018 ： 7122 - 7131 . doi: 10.1109/cvpr.2018.00744 http://dx.doi.org/10.1109/cvpr.2018.00744 End-to-end recovery of human shape and pose 穆靖，李伟华，饶俊民，等 . 采用三层模板局部差异度量的红外弱小目标检测［J］. 光学精密工程， 2022 ， 30 （ 7 ）： 869 - 882 . doi: 10.37188/OPE.20223007.0869 http://dx.doi.org/10.37188/OPE.20223007.0869 采用三层模板局部差异度量的红外弱小目标检测 MU J ， LI W H ， RAO J M ， et al . Infrared small target detection using tri-layer template local difference measure ［J］. Optics and Precision Engineering ， 2022 ， 30 （ 7 ）： 869 - 882 . （in Chinese） . doi: 10.37188/OPE.20223007.0869 http://dx.doi.org/10.37188/OPE.20223007.0869 Infrared small target detection using tri-layer template local difference measure BISWAS S K ， MILANFAR P . Linear support tensor machine with LSK channels： pedestrian detection in thermal infrared images ［J］. IEEE Transactions on Image Processing ， 2017 ， 26 （ 9 ）： 4229 - 4242 . doi: 10.1109/tip.2017.2705426 http://dx.doi.org/10.1109/tip.2017.2705426 Linear support tensor machine with LSK channels： pedestrian detection in thermal infrared images REN S Q ， HE K M ， GIRSHICK R ， et al . Faster R-CNN： towards real-time object detection with region proposal networks ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence ， 2017 ， 39 （ 6 ）： 1137 - 1149 . doi: 10.1109/tpami.2016.2577031 http://dx.doi.org/10.1109/tpami.2016.2577031 Faster R-CNN： towards real-time object detection with region proposal networks 刘琼，罗晴，彭绍武 . 车载热成像行人检测RoI提取方法［J］. 东北大学学报（自然科学版）， 2020 ， 41 （ 8 ）： 1083 - 1090 . doi: 10.12068/j.issn.1005-3026.2020.08.004 http://dx.doi.org/10.12068/j.issn.1005-3026.2020.08.004 车载热成像行人检测RoI提取方法 LIU Q ， LUO Q ， PENG SH W . RoI extraction for vehicular thermal infrared pedestrian detection ［J］. Journal of Northeastern University （Natural Science）， 2020 ， 41 （ 8 ）： 1083 - 1090 . （in Chinese） . doi: 10.12068/j.issn.1005-3026.2020.08.004 http://dx.doi.org/10.12068/j.issn.1005-3026.2020.08.004 RoI extraction for vehicular thermal infrared pedestrian detection CHEN Y Y ， JHONG S Y ， LI G Y ， et al . Thermal-based pedestrian detection using faster R-CNN and region decomposition branch ［C］. 2019 International Symposium on Intelligent Signal Processing and Communication Systems （ISPACS）. Taipei ， China . IEEE ， 2019 ： 1 - 2 . doi: 10.1109/ispacs48206.2019.8986298 http://dx.doi.org/10.1109/ispacs48206.2019.8986298 Thermal-based pedestrian detection using faster R-CNN and region decomposition branch XU Z W ， VONG C M ， WONG C C ， et al . Ground plane context aggregation network for day-and-night on vehicular pedestrian detection ［J］. IEEE Transactions on Intelligent Transportation Systems ， 2021 ， 22 （ 10 ）： 6395 - 6406 . doi: 10.1109/tits.2020.2991848 http://dx.doi.org/10.1109/tits.2020.2991848 Ground plane context aggregation network for day-and-night on vehicular pedestrian detection PEI D S ， JING M X ， LIU H P ， et al . A fast RetinaNet fusion framework for multi-spectral pedestrian detection ［J］. Infrared Physics & Technology ， 2020 ， 105 ： 103178 . doi: 10.1016/j.infrared.2019.103178 http://dx.doi.org/10.1016/j.infrared.2019.103178 A fast RetinaNet fusion framework for multi-spectral pedestrian detection WANG D ， LAN J H . PPDet： a novel infrared pedestrian detection network in a per-pixel prediction fashion ［J］. Infrared Physics & Technology ， 2021 ， 119 ： 103965 . doi: 10.1016/j.infrared.2021.103965 http://dx.doi.org/10.1016/j.infrared.2021.103965 PPDet： a novel infrared pedestrian detection network in a per-pixel prediction fashion 李经宇，杨静，孔斌，等 . 基于注意力机制的多尺度车辆行人检测算法［J］. 光学精密工程， 2021 ， 29 （ 6 ）： 1448 - 1458 . doi: 10.37188/OPE.20212906.1448 http://dx.doi.org/10.37188/OPE.20212906.1448 基于注意力机制的多尺度车辆行人检测算法 LI J Y ， YANG J ， KONG B ， et al . Multi-scale vehicle and pedestrian detection algorithm based on attention mechanism ［J］. Optics and Precision Engineering ， 2021 ， 29 （ 6 ）： 1448 - 1458 . （in Chinese） . doi: 10.37188/OPE.20212906.1448 http://dx.doi.org/10.37188/OPE.20212906.1448 Multi-scale vehicle and pedestrian detection algorithm based on attention mechanism 刘怡帆，王旭飞，周鹏，等 . 基于YOLOv4神经网络的红外图像道路行人检测［J］. 数字技术与应用， 2021 ， 39 （ 10 ）： 19 - 22 . doi: 10.19695/j.cnki.cn12-1369.2021.10.07 http://dx.doi.org/10.19695/j.cnki.cn12-1369.2021.10.07 基于YOLOv4神经网络的红外图像道路行人检测 LIU Y F ， WANG X F ， ZHOU P ， et al . Road pedestrian detection based on YOLOv4 neural network in infrared image ［J］. Digital Technology ＆ Application ， 2021 ， 39 （ 10 ）： 19 - 22 . （in Chinese） . doi: 10.19695/j.cnki.cn12-1369.2021.10.07 http://dx.doi.org/10.19695/j.cnki.cn12-1369.2021.10.07 Road pedestrian detection based on YOLOv4 neural network in infrared image LIU Y ， SU H L ， ZENG C ， et al . A robust thermal infrared vehicle and pedestrian detection method in complex scenes ［J］. Sensors （Basel， Switzerland）， 2021 ， 21 （ 4 ）： 1240 . doi: 10.3390/s21041240 http://dx.doi.org/10.3390/s21041240 A robust thermal infrared vehicle and pedestrian detection method in complex scenes ZHENG Z H ， WANG P ， REN D W ， et al . Enhancing geometric factors in model learning and inference for object detection and instance segmentation ［J］. IEEE Transactions on Cybernetics ， 2022 ， 52 （ 8 ）： 8574 - 8586 . doi: 10.1109/tcyb.2021.3095305 http://dx.doi.org/10.1109/tcyb.2021.3095305 Enhancing geometric factors in model learning and inference for object detection and instance segmentation 贺愉婷，车进，吴金蔓 . 基于YOLOv5和重识别的行人多目标跟踪方法［J］. 液晶与显示， 2022 ， 37 （ 7 ）： 880 - 890 . doi: 10.37188/CJLCD.2022-0025 http://dx.doi.org/10.37188/CJLCD.2022-0025 基于YOLOv5和重识别的行人多目标跟踪方法 HE Y T ， CHE J ， WU J M . Pedestrian multi-target tracking method based on YOLOv5 and person re-identification ［J］. Chinese Journal of Liquid Crystals and Displays ， 2022 ， 37 （ 7 ）： 880 - 890 . （in Chinese） . doi: 10.37188/CJLCD.2022-0025 http://dx.doi.org/10.37188/CJLCD.2022-0025 Pedestrian multi-target tracking method based on YOLOv5 and person re-identification FAN H Y ， MA J S ， FAN H H ， et al . Iterative quadtree decomposition based automatic selection of the seed point for ultrasound breast tumor images ［J］. Multimedia Tools and Applications ， 2017 ， 76 （ 3 ）： 3505 - 3517 . doi: 10.1007/s11042-016-3761-z http://dx.doi.org/10.1007/s11042-016-3761-z Iterative quadtree decomposition based automatic selection of the seed point for ultrasound breast tumor images MUN J ， JANG Y ， NAM Y ， et al . Edge-enhancing bi-histogram equalisation using guided image filter ［J］. Journal of Visual Communication and Image Representation ， 2019 ， 58 ： 688 - 700 . doi: 10.1016/j.jvcir.2018.12.037 http://dx.doi.org/10.1016/j.jvcir.2018.12.037 Edge-enhancing bi-histogram equalisation using guided image filter MA J Y ， MA Y ， LI C . Infrared and visible image fusion methods and applications： a survey ［J］. Information Fusion ， 2019 ， 45 ： 153 - 178 . doi: 10.1016/j.inffus.2018.02.004 http://dx.doi.org/10.1016/j.inffus.2018.02.004 Infrared and visible image fusion methods and applications： a survey EMARY E ， ZAWBAA H M ， GROSAN C . Experienced gray wolf optimization through reinforcement learning and neural networks ［J］. IEEE Transactions on Neural Networks and Learning Systems ， 2018 ， 29 （ 3 ）： 681 - 694 . doi: 10.1109/tnnls.2016.2634548 http://dx.doi.org/10.1109/tnnls.2016.2634548 Experienced gray wolf optimization through reinforcement learning and neural networks BASHI-AZGHADI S N ， AFSHAR A ， AFSHAR M H . Multi-period response management to contaminated water distribution networks： dynamic programming versus genetic algorithms ［J］. Engineering Optimization ， 2018 ， 50 （ 3 ）： 415 - 429 . doi: 10.1080/0305215x.2017.1318871 http://dx.doi.org/10.1080/0305215x.2017.1318871 Multi-period response management to contaminated water distribution networks： dynamic programming versus genetic algorithms KHISHE M ， MOSAVI M R . Chimp optimization algorithm ［J］. Expert Systems With Applications ， 2020 ， 149 ： 113338 . doi: 10.1016/j.eswa.2020.113338 http://dx.doi.org/10.1016/j.eswa.2020.113338 Chimp optimization algorithm HWANG S ， PARK J ， KIM N ， et al . Multispectral pedestrian detection： Benchmark dataset and baseline ［C］. 2015 IEEE Conference on Computer Vision and Pattern Recognition . Boston， MA， USA . IEEE ， 2015 ： 1037 - 1045 . doi: 10.1109/cvpr.2015.7298706 http://dx.doi.org/10.1109/cvpr.2015.7298706 Multispectral pedestrian detection： Benchmark dataset and baseline 王灿芳，崔良玉，阎兵 . 基于Variance-Brenner函数的显微图像清晰度评价算法研究［J］. 装备制造技术， 2020 （ 10 ）： 78 - 82 . doi: 10.3969/j.issn.1672-545X.2020.10.020 http://dx.doi.org/10.3969/j.issn.1672-545X.2020.10.020 基于Variance-Brenner函数的显微图像清晰度评价算法研究 WANG C F ， CUI L Y ， YAN B . Study on the microscopic image sharpness evaluation algorithm based on variance-Brenner function ［J］. Equipment Manufacturing Technology ， 2020 （ 10 ）： 78 - 82 . （in Chinese） . doi: 10.3969/j.issn.1672-545X.2020.10.020 http://dx.doi.org/10.3969/j.issn.1672-545X.2020.10.020 Study on the microscopic image sharpness evaluation algorithm based on variance-Brenner function 孙红利，冯旗，董峰 . 图像清晰度评价算法研究［J］. 传感器与微系统， 2017 ， 36 （ 2 ）： 67 - 70 . doi: 10.13873/J.1000-9787(2017)02-0067-04 http://dx.doi.org/10.13873/J.1000-9787(2017)02-0067-04 图像清晰度评价算法研究 SUN H L ， FENG Q ， DONG F . Research on algorithm for image clarity evaluation ［J］. Transducer and Microsystem Technologies ， 2017 ， 36 （ 2 ）： 67 - 70 . （in Chinese） . doi: 10.13873/J.1000-9787(2017)02-0067-04 http://dx.doi.org/10.13873/J.1000-9787(2017)02-0067-04 Research on algorithm for image clarity evaluation LIU W ， ANGUELOV D ， ERHAN D ， et al . SSD ： Single Shot MultiBox Detector ［M］. Computer Vision - ECCV 2016 . Cham ： Springer International Publishing ， 2016 ： 21 - 37 . doi: 10.1007/978-3-319-46448-0_2 http://dx.doi.org/10.1007/978-3-319-46448-0_2 SSD： Single Shot MultiBox Detector LIN T Y ， GOYAL P ， GIRSHICK R ， et al . Focal loss for dense object detection ［C］. 2017 IEEE International Conference on Computer Vision . Venice， Italy . IEEE ， 2017 ： 2999 - 3007 . doi: 10.1109/iccv.2017.324 http://dx.doi.org/10.1109/iccv.2017.324 Focal loss for dense object detection SUN P Z ， ZHANG R F ， JIANG Y ， et al . Sparse R-CNN： end-to-end object detection with learnable proposals ［C］. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Nashville ， TN ， USA . IEEE ， 2021 ： 14449 - 14458 . doi: 10.1109/cvpr46437.2021.01422 http://dx.doi.org/10.1109/cvpr46437.2021.01422 Sparse R-CNN： end-to-end object detection with learnable proposals ZHANG H Y ， WANG Y ， DAYOUB F ， et al . VarifocalNet： an IoU-aware dense object detector ［C］. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Nashville ， TN ， USA . IEEE ， 2021 ： 8510 - 8519 . doi: 10.1109/cvpr46437.2021.00841 http://dx.doi.org/10.1109/cvpr46437.2021.00841 VarifocalNet： an IoU-aware dense object detector FENG C J ， ZHONG Y J ， GAO Y ， et al . TOOD： task-aligned one-stage object detection ［C］. 2021 IEEE/CVF International Conference on Computer Vision （ICCV）. Montreal ， QC， Canada . IEEE ， 2021 ： 3490 - 3499 . doi: 10.1109/iccv48922.2021.00349 http://dx.doi.org/10.1109/iccv48922.2021.00349 TOOD： task-aligned one-stage object detection SUN M Y ， ZHANG H C ， HUANG Z L ， et al . Road infrared target detection with I‐YOLO ［J］. IET Image Processing ， 2022 ， 16 （ 1 ）： 92 - 101 . doi: 10.1049/ipr2.12331 http://dx.doi.org/10.1049/ipr2.12331 Road infrared target detection with I‐YOLO Kieu M ， Bagdanov A D ， Bertini M ， et al . Task-conditioned domain adaptation for pedestrian detection in thermal imagery ［C］. European conference on computer vision. Springer， Cham ： ECCV ， 2020 ： 546 - 562 . doi: 10.1007/978-3-030-58542-6_33 http://dx.doi.org/10.1007/978-3-030-58542-6_33 Task-conditioned domain adaptation for pedestrian detection in thermal imagery 10.37188/OPE.20223019.2390 TP391.41 TN219 A 1004-924X（2022）19-2390-14 郝帅,何田,马旭等.动态特征优化机制下的跨尺度红外行人检测[J].光学精密工程,2022,30(19):2390-2403. 郝帅,何田,马旭等.动态特征优化机制下的跨尺度红外行人检测[J].光学精密工程,2022,30(19):2390-2403. HAO Shuai,HE Tian,MA Xu,et al.Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism[J].Optics and Precision Engineering,2022,30(19):2390-2403. HAO Shuai,HE Tian,MA Xu,et al.Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism[J].Optics and Precision Engineering,2022,30(19):2390-2403. 国家自然科学基金资助项目（No.51804250）；中国博士后科学基金资助项目（No.2019M653874XB， No. 2020M683522）；陕西省科技计划项目（No.2021JQ-572， No.2020JQ-757）；陕西省教育厅科研计划项目（No.18JK0512， No.21JK0769）；西安市碑林区科技计划项目（No.GX2116） 2022-06-17 2022-07-05 2022-10-10 2022-10-24 光学精密工程 19 30

浏览量

下载量

CSCD

文章被引用时，请邮件提醒。

提交

工具集

关联资源

融合分形几何特征Resnet遥感图像建筑物分割

引入注意力机制的轻量级小目标检测网络

基于N3D_DIOU的图像与点云融合目标检测算法

采用多尺度特征融合SSD的遥感图像小目标检测

光谱色差与主分量特征融合的石窟表面风化智能评估