A Pedestrian Multi-object Tracking Algorithm based on CenterTrack for Autonomous Driving

Multi-object tracking is a research hotspot in intelligent driving scenarios, and the trajectory association of tracked objects is one of the hotspots that needs to be solved urgently. The CenterTrack algorithm using the “joint detection and tracking” paradigm has high inference speed and accuracy. However, object overlap due to accidental occlusion and scene clutter often leads to missed detections and tracking failures. In response to the above problems, this paper proposes a multi-object tracking model SPCTrack, which combines the attention mechanism, multi-scale fusion and other ideas, and uses pyramid segmentation attention model to improve the detection effect of the network when the target is occluded. In addition, this paper also introduces the Swish activation function to effectively improve the correlation accuracy during tracking. This paper is validated on the MOT16 and MOT17 datasets, the results show that compared with CenterTrack and other mainstream tracking models, SPCTrack improves the detection accuracy, association accuracy, and overall tracking accuracy of multi-object tracking. At the same time, the speed can reach 19.1FPS, with high real-time performance.