Multi-Target Tracking with Trajectory Prediction and Re-Identification

Due to the complexity and clutter of real-world scenes, occlusion becomes a long-lasting difficulty in object tracking. Most existing tracking methods cannot effectively handle occlusion. In this paper, we propose a novel tracking framework that combines trajectory prediction and multi-cue appearance modeling to deal with the occlusion difficulty. When a target is completely occluded by background or other targets, it is unable to observe the target position. Therefore, we propose a Long Short-Term Memory (LSTM) model that merges attention mechanism and interaction module to predict the locations of all targets in the next frame. Considering that partial occlusion and inaccuracy of object bounding boxes often take place, we propose a multi-branch deep network architecture combining global and local features to realize accurate tracking and person re-identification (ReID). According to the experimental results on multiple benchmark datasets, our method achieves state-of the-art performance and outperforms many existing approaches.