红外交通场景下遮挡行人目标检测算法研究  

Research on the detection algorithm of obscured pedestrian targets in infrared traffic scenes

作　　者：李明益 贺敬良[1,2] 陈勇 赵理[1,2] 龙震海 LI Ming-yi;HE Jing-liang;CHEN Yong;ZHAO Li;LONG Zhen-hai(School of Electromechanical Engineering,Beijing Information Science and Technology University,Beijing 100192,China;Collaborative Innovation Center of Electric Vechicles in Beijing,Beijing 100192,China;School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京信息科技大学机电工程学院,北京100192 [2]北京电动车辆协同创新中心,北京100192 [3]北京理工大学机械与车辆学院,北京100081

出　　处：《激光与红外》2022年第9期1417-1424,共8页Laser & Infrared

基　　金：国家自然科学基金项目(No.520770077);科技创新服务能力建设-科研基地建设-新能源汽车北京实验室项目(No.PXM2017-014224-000005-00249684-FCG)资助。

摘　　要：针对交通十字路口等视野盲区往来行人间存在遮挡情况,如何高效准确地检测复杂道路中目标行人具有实际意义。为了实现夜间交汇路口场景行人检测,提出一种基于改进YOLOv5的行人目标检测算法,采用Non-local和PSA模块对YOLOv5原网络的Bottleneck CSP进行改进,能够有效弥补遮挡中行人特征的帧间信息交互过程,增强长程范围通道特征依赖关系。设计更深的160×160检测层和自适应anthor,提升夜间行人检测的边界回归精确度。实验结果表明,针对夜间下交通路口场景,压缩改进后模型对行人检测鲁棒性高,相较于原始算法mAP_0.5和mAP_0.5:0.95值分别提升了14.2%和12.7%,说明所提算法对夜间行人检测的有效性。In view of the occlusion between pedestrians in blind field of vision such as traffic intersections, it is of practical significance to detect target pedestrians in complex roads efficiently and accurately.In order to achieve pedestrian detection at nighttime intersection scenes, a pedestrian target detection algorithm based on improved YOLOv5 is proposed, using Non-local and PSA modules to improve the Bottleneck CSP of the original YOLOv5 network, which can effectively make up for the inter-frame information exchange process of pedestrian features and enhance the feature dependence of long-range channels.A deeper 160×160 detection layer and adaptive anthor are designed to improve the boundary regression accuracy of night pedestrian detection.The experimental results show that for the traffic intersection scene under nighttime, the compressed and improved model has high robustness for pedestrian detection, with 14.2 % and 12.7 % improvement compared to the original algorithm mAP_0.5 and mAP_0.5:0.95 values, respectively, indicating the effectiveness of the proposed algorithm for nighttime pedestrian detection.

关 键 词：深度学习 行人目标检测 YOLOv5 NON-LOCAL PSA Model 

分 类 号：TN219[电子电信—物理电子学]