Research on an improved MOTR-GRU multi-target tracking and behavior analysis method

Localizing people and identifying their behaviors in videos is a challenging task in advanced video understanding. However, traditional behavior recognition algorithms only detect the behavior of a single target and use a two-stage method for feature extraction and behavior recognition, which leads to problems such as difficulty in distinguishing individual behaviors and low recognition accuracy. To address these problems, an end-to-end multi-target human tracking and behavior analysis method, MOTR-GRU, is proposed to solve the problems of difficulty in distinguishing individual behaviors and low accuracy caused by accumulated errors in multi-target tracking and behavior analysis. First, the model continuously tracks different human targets in the video by extending the detection query in MOTR as the tracking query, and obtains relatively accurate tracking results without post-processing. Then, the gating mechanism is used to process and fuse the human information captured by the tracking module, enhance the model’s ability to process time series data, and analyze the behaviors of multiple targets frame by frame. Finally, the focal loss is used to reduce the impact of target category imbalance. Experiments are carried out on the Volleyball dataset. The results show that the behavior recognition accuracy of the end-to-end MOTR-GRU method reaches $\mathbf{6 8. 6 \%}$. This shows that the MOTR-GRU method can improve the ability of multi-target behavior analysis while ensuring the real-time tracking of targets.