Multi-object tracking with Siamese-RPN and adaptive matching strategy

The multiple object tracking (MOT) task has always been a research hot point in computer vision. However, most current MOT algorithms do not pay enough attention to the prediction module. Also, in data association, they use manual debugging to determine the matching threshold. In this paper, we propose a new MOT algorithm. By introducing the Siamese RPN network as a predictor in the advanced detection module, the algorithm greatly enhances the adaptability to complex and diverse application scenarios while improving accuracy. Simultaneously, by analyzing the distance matrix in the data association module, we design a simple adaptive threshold determination method, which saves a lot of redundant experiments in the debugging process and avoids manual intervention. Combined with the self-designed matching strategy, the MOT algorithm with high accuracy and adaptability to more complex and diverse application scenarios such as nonlinear and high-speed is realized. Finally, the effectiveness and advantages of each module are verified on the MOT16, MOT17, and MOT20 benchmarks.