Multi Visual Target Tracking Technology for USV Based on DeepSORT

In response to the poor continuity of visual multi target tracking trajectories for unmanned surface vehicle, this paper analyzes the impact of DeepSORT detection post-processing on target tracking loss, designs a hierarchical matching strategy, and effectively recalls low confidence targets. The experiment showed that the hierarchical matching tracking strategy achieved a 3.8 percentage point increase in IDF1, while the MOTA remained basically unchanged. In response to the problem of target underreporting in DeepSORT tracking under unmanned boat motion conditions, this paper analyzes the reasons for the insufficient predictive ability of Kalman filtering for nonlinear changes in inter frame target trajectories in motion tracking. It eliminates the influence of Mahalanobis distance between Kalman filtering predicted trajectories and target detection results on the tracking cost matrix, thereby optimizing the update method of the cost matrix. The experiment shows that this method achieves a 5.8% and 1.2% increase in MOTA under motion and static tracking conditions, respectively. The paper proposes the MDeepSORT target tracking algorithm framework based on a comprehensive hierarchical matching strategy and cost matrix update optimization method. This framework achieves a MOTA improvement of 8.2% and 2.2% under both motion and static tracking conditions, respectively, verifying the effectiveness of the hierarchical matching strategy and cost matrix update optimization method.