A Two-step Model for Multi-object Tracking

Multi-object tracking is widely used in video analysis. However, due to the limitation of detector performance, many multi-object tracking models have the problem of detecting two objects into one object in some occlusion scenes. In this paper, we propose a two-step model for handling this problem. In the first step model, the non-occlusion targets are detected and embeddings are extracted, while the occlusion areas are identified. The second step model processes the occlusion areas to obtain occlusion targets' accurate positions and embeddings. Finally, we integrate and optimize the output results of the two steps models. Experiments show that the number of false positives and missed positives in our model's object detection is significantly reduced. The multi-object tracking performance (MOTA metric) is improved by nearly 3% compared with other models.