一种面向遮挡行人检测的改进YOLOv3算法  

Occluded Pedestrian Detection Algorithm Based on Improved YOLOv3

作　　者：李翔[1,2,3,4] 何淼 罗海波[1,2,3] Li Xiang;He Miao;Luo Haibo(Key Laboratory of Opto-Electronic Information Processing,Chinese Academy of Sciences,Shenyang 110016,Liaoning,China;Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,Liaoning,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,Liaoning,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院光电信息处理重点实验室,辽宁沈阳110016 [2]中国科学院沈阳自动化研究所,辽宁沈阳110016 [3]中国科学院机器人与智能制造创新研究院,辽宁沈阳110169 [4]中国科学院大学,北京100049

出　　处：《光学学报》2022年第14期152-161,共10页Acta Optica Sinica

摘　　要：在密集行人检测场景中,目标间的相互遮挡重叠会造成YOLOv3模型的检测性能下降。针对造成YOLOv3性能下降的原因提出三点改进。一是提出了一种聚拢损失函数,通过优化预测框坐标的方差与均值,使得属于同一个目标的预测框更加紧致,进而降低假阳率。二是提出了一种高分辨率特征金字塔,通过上采样提高每层金字塔特征的分辨率,并引入浅层特征以增强相邻子特征的差异,从而为高重叠目标生成具有区分度的深度特征。三是提出了一种基于空间注意力机制的检测头,用以降低冗余预测框的数量,减少非极大值抑制(NMS)过程的计算负担。在密集行人数据集CrowdHuman上进行的实验的结果显示,所提算法在使用传统NMS方法的情况下使得YOLOv3检测的平均精度和召回率分别提高了2.91个百分点和3.20个百分点,丢失率降低了1.24个百分点,有效提升了对遮挡行人的检测性能。In crowded scenes,it is difficult for YOLOv3 to detect the objects that overlap each other heavily.Aiming at the reasons for the decline of YOLOv3 performance,three improvements are proposed.Firstly,a Tight Loss function is proposed,which optimizes the variance and mean of the coordinates of the prediction boxes to make the prediction boxes belonging to the same target more compact,thus reducing the false positive rate.Secondly,a high-resolution feature pyramid is proposed,in which the resolution of each pyramid feature is improved by upsampling,and shallow features are introduced to enhance the differences between adjacent sub-features,so as to generate distinguishing depth features for highly overlapped targets.Thirdly,a detection head based on spatial attention mechanism is proposed to reduce the number of redundant prediction boxes,so as to reduce the computational burden of the non-maximum suppression(NMS)process.The experimental results on the crowded dataset CrowdHuman show that the average accuracy and recall rate of YOLOv3 detection are improved by 2.91 percentage points and 3.20 percentage points,and the miss rate is reduced by1.24 percentage points by using the proposed algorithms under the condition of using the traditional NMS method,which demonstrates the effectiveness of the proposed algorithms in boosting the performance in occluded pedestrian detection.

关 键 词：机器视觉 目标检测 神经网络 行人检测 注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]