基于深度学习的行人和车辆检测与跟踪研究  

Research on Pedestrian and Vehicle Detection and Tracking Based on Deep Learning

作　　者：袁旻颉 罗荣芳[1] 陈静[1] 苏成悦[1,2] YUAN Minjie;LUO Rongfang;CHEN Jing;SU Chengyue(School of Physics and Optoelectronic Engineering,Guangdong University of Technoology,Guangzhou 510006,China;School of Advanced Manufacturing,Guangdong University of Technoology,Jieyang 515548,China)

机构地区：[1]广东工业大学物理与光电工程学院,广东广州510006 [2]广东工业大学先进制造学院,广东揭阳515548

出　　处：《现代信息科技》2024年第1期121-124,129,共5页Modern Information Technology

摘　　要：针对行人及车辆的多目标检测和跟踪中检测精度不足及跟踪目标丢失和身份切换问题,文章提出一种改进YOLOv5与改进Deep SORT相结合的多目标检测跟踪算法。检测阶段使用Varifocal Loss替换二元交叉熵损失函数结合CA注意力机制和DIoU_NMS算法。跟踪阶段将Deep SORT的REID模块特征提取网络替换为EfficientNetV2-S。在COCO数据集检测上,map@0.5达到78%,比原始模型提升4.5%,在MOT16数据集跟踪上,MOTA达到58.1,比原始模型提升5.7,IDswitch减少了516次相当于减少了55.1%,测试结果表明该算法有较好的实际应用价值。This paper proposes a multi-objective detection and tracking algorithm combining improved YOLOv5 and improved Deep SORT to address the issues of insufficient detection accuracy,lost tracking targets,and identity switching in pedestrian and vehicle's multi-target detection and tracking.Replacing binary cross entropy loss function with Varifocal Loss in the detection phase,combined with CA attention mechanism and DIoU_NMS algorithm.During the tracking phase,replace the feature extraction network of the REID module of Deep SORT with EfficientNetV2-S.In COCO dataset detection,map@0.5 reaches 78%,an improvement of 4.5%compared to the original model.On the MOT16 dataset tracking,the MOTA reaches 58.1,an improvement of 5.7 compared to the original model.The IDswitch is reduced by 516 times,which is equivalent to a reduction of 55.1%.The test results show that the algorithm has good practical application value.

关 键 词：深度学习 目标检测 目标跟踪 计算机视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]