摘要
针对基于检测的两步多目标跟踪算法模型结构复杂、实时性差的问题,提出一种基于改进YOLOv4-Tiny和DeepSORT的实时多目标跟踪算法。在YOLOv4-Tiny算法中引入深度可分离卷积,压缩模型计算量;将检测分支增加至3个,并使用多尺度特征融合结构以降低对小目标的漏检率;利用改进的GC注意力模块,加强网络对全局上下文特征的提取能力。跟踪部分使用DeepSORT算法,使用匀加速卡尔曼滤波优化其行人运动模型,利用浅层分类网络重构其外观模型,最后在MOT16测试序列中实验。结果表明,改进算法的总参数量为4.2 M,较原算法减少52%且MOTA增加5.2%,GPU下处理时间加快,单CPU时能达到平均每秒11帧的跟踪速度,能满足低算力设备对跟踪任务精度和速度的要求。
To solve the issue of complicated structure and low real-time performance of the two-step multiple object tracking by detection algorithm, a real-time multiple object tracking algorithm founded on modified YOLOv4-Tiny and DeepSORT algorithm is proposed. The depthwise separable convolution is employed in YOLOv4-Tiny, to reduce the calculation of the model. The detection branches are increased to 3, and multi-scale feature fusion structure is built to decrease the missed ratio of tiny objects. The modified GC attention module is used to extract the global context features of the model. In the tracking part, the pedestrian motion model of DeepSORT is optimized and the appearance model is reconstructed, the detection and tracking algorithms are combined and experimented in MOT16 test sequences finally. The results show that the total parameters of the improved algorithm are 4.2 M, 52% less than the original algorithm and 5.2% more MOTA, faster processing speed under GPU, and the tracking speed of an average of 11 frames per second can be achieved under a single CPU, which can fulfill the requests of precision and speed for multiple object tracking mission in low-calculation devices.
作者
黄凯文
凌六一
王成军
吴起
李学松
Huang Kaiwen;Ling Liuyi;Wang Chengjun;Wu Qi;Li Xuesong(School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan 232001,China;School of Artificial Intelligence,Anhui University of Science and Technology,Huainan 232001,China)
出处
《电子测量技术》
北大核心
2022年第6期7-13,共7页
Electronic Measurement Technology
基金
安徽省高校协同创新项目(GXXT2019-018)
安徽省重点研发计划项目(201904a05020092)资助。
