改进YOLOX Tiny与DeepSort相结合的多目标跟踪算法

Research on Multiple Object Tracking Algorithm with the Combination of Improved YOLOX Tiny and DeepSort

针对多目标跟踪时有发生误检、漏检等情况,提出了CSD YOLOX Tiny的多目标跟踪算法。在骨干网络中搭建结合三卷积的跨阶段局部Swin Transformer Block结构,提升网络模型对全局和上下文信息的捕获能力。在网络中引入组归一化,加快网络模型收敛速度并提升跟踪精度;采用坐标注意力机制对不同通道特征之间的相关信息进行有效整合,提升网络模型对感兴趣区域特征的获取能力。实验结果表明:提出的多目标跟踪算法跟踪精度提升了2.36%,达到56.38%。在保证网络模型参数量较少、计算量较小的情况下,提出的跟踪算法较好地改善误检和漏检问题,相比于YOLOX Tiny DeepSort算法误检、漏检数量分别减少了811、1574个,能满足常规设备实时高效的多目标跟踪任务需求。

The CSD YOLOX Tiny multiple object tracking algorithm is proposed to solve the problems of false detection and missing detection.A cross stage partial Swin Transformer Block structure combined with triple convolution is built in the backbone network to improve its ability to capture global and contextual information.The group normalization is introduced into the network to speed up its convergence and improve its tracking accuracy.In order to increase the network's ability to acquire the features of the region of interest,a coordinate attention mechanism is used to effectively integrate the relevant information between different channel features.The results show that the tracking accuracy of the proposed multiple object tracking algorithm is improved by 2.36%up to 56.38%.The problems of false detection and missing detection are better solved with a smaller number of parameters and a smaller amount of computation.Compared with the YOLOX Tiny DeepSort algorithm,the proposed algorithm reduces the number of false detection by 811 and the number of missing detection by 1574.It can allow conventional equipment to implement real time and efficient multiple object tracking tasks.