CuesTrack:Multi-Object Tracking Based on Weak Clues and Trajectory Prediction

Multi-Object Tracking (MOT) is a crucial problem in computer vision, focusing on detecting and associating all required targets across frames. Despite the rapid development of MOT, challenges such as occlusions, similar objects, and complex scenes remain unresolved. This paper introduces CuesTrack, an innovative object tracker that leverages multiple cues to effectively address these challenges by combining weak cues-such as height state and trajectory trends-with strong cues like position and appearance features. Firstly, the TPM (Trajectory Prediction Module) utilizes the historical trajectory information of targets. The PV-LSTM module within TPM predicts the target's next position and movement direction, thereby improving identity determination and reducing the impact of occlusion on identity recognition. Secondly, the MH-IoU (Mixed-Height IoU) is designed to reasonably utilize height information during object association, making the association more robust. Lastly, the ReID-CCM (ReID-Cost Management Module) extracts target appearance features and balances the costs of position cues, appearance cues, trajectory cues, and height cues through the CCM cost management module. Despite its excellent performance, our method remains simple, online, and real-time (SORT). Furthermore, CuesTrack achieves outstanding performance across various benchmarks, including MOT17 and MOT20. Notably, CuesTrack outperforms BYTETrack with HOTA scores of 0.9 and 1.9 on the MOT17 and MOT20 datasets, respectively. The code is available at: https://github.com/zhengpangzi/CuesTrack.