Interactive Multiple Model Filter Algorithm for Multi-Object Tracking of Unmanned Surface Vessels

Multi-object tracking algorithm based on deep learning has made rapid development in the field of pedestrian tracking. However, there are few tracking studies of unmanned surface vessels in the marine field, some research institutions have attempted to integrate YOLOv5 DeepSORT for unmanned surface vessels tracking. The motion of unmanned surface vessels exhibits nonlinear characteristics. However, the Kalman filter algorithm assumes linear motion of objects. The frame rate of the videos in the MOT datasets is 25 FPS, and the pedestrian movement is slow. The frame rate of unmanned surface vessels videos is low, and unmanned surface vessels movement faster compared to pedestrians. The position of the same object changes greatly in the consecutive frames. Consequently, traditional Kalman filter algorithms and IOU for association match results in poor performance in multiobject tracking for unmanned surface vessels. To address this issue, we propose an interactive multiple model (IMM) filter algorithm, and the position metric employ Euclidean distance. This method demonstrates stronger robustness and accuracy on a self-built datasets for unmanned surface vessels tracking. Experimental results demonstrate that the proposed method significantly improves tracking performance: MOTA increases by 6.2%, HOTA by 7.4%, IDF1 by 7.5%, and IDSW decreases by 100 instances. The method shows strong performance in tracking unmanned surface vessels.