基于多特征匹配的轻量化行人跟踪算法研究

Research on lightweight pedestrian tracking algorithm based on multi-feature matching

行人跟踪是深度学习研究中的热点内容。目前的跟踪算法存在无法满足实时性和因跟踪目标相似度太高、目标间的遮挡、运动不规律造成ID频繁转换的问题。为了提高运行速度,在目标检测阶段使用CNN和transformer相结合的轻量化网络,采用联合检测的方式,共享特征权重,并行计算检测、重识别、人体姿态估计分支,同时调整各个分支卷积通道数。跟踪部分则利用卡尔曼滤波预测的目标运动信息,目标重识别信息,和目标姿态的各个关键点位置信息共同完成目标身份匹配,减少了同一ID的频繁转换。实验部分采用MOT16数据集训练和测试。本算法的多目标跟踪准确度(MOTA)为48.5%,多目标跟踪精确度(MOTP)为78.17%,FPS为20,模型大小为18.4 M。实验表明,提出的跟踪算法提高了整体的跟踪性能,实时性和准确性达到了预期要求。

Pedestrian tracking is a hot topic in deep learning research. The current tracking algorithm has the problems that it cannot meet the real-time performance and frequent ID conversion due to the high similarity of the tracking targets, the occlusion between the targets, and the irregular motion. In order to improve the running speed, a lightweight network combining CNN and transformer is used in the target detection stage, and a joint detection method is adopted to share feature weights, calculate detection, re-identification, and human pose estimation branches in parallel, and adjust the number of convolution channels of each branch at the same time.. The tracking part uses the target motion information predicted by Kalman filtering, the target re-identification information, and the position information of each key point of the target pose to complete the target identity matching, which reduces the frequent conversion of the same ID. The experimental part uses the MOT16 dataset for training and testing. The multi-target tracking accuracy(MOTA) of this algorithm is 48.5%, the multi-target tracking accuracy(MOTP) is 78.17%, the FPS is 20, and the model size is 18.4 M. Experiments show that the proposed tracking algorithm improves the overall tracking performance, and the real-time performance and accuracy meet the expected requirements.