Recent studies of multi-object tracking have made remarkable progress in the performance of multi-object tracking algorithms. Nonetheless, little attention has been paid to the balance between accuracy and efficiency. Following the noteworthy trends of the end-to-end paradigm, to balance the accuracy and speed, we propose the Chained-Center-Tracker, a chained architecture jointly accomplishing multiple subtasks for tracking the central points. Specifically, the backbone feature from the previous frame is reused efficiently in our chained architecture. Additionally, the end-to-end framework can be globally trainable, aiming at fairer collaboration among multiple subtasks and less parameter-tuning efforts. To alleviate the negative effects of large-scale change, the generalisation ability of object detection is enhanced by embedding channel attention. For data association, 3D convolution is employed to extract spatiotemporal information across adjacent frames and predict the displacement of targets. Heatmaps, which highlight the regions where targets are located, serve as an attention map that can freely focus on the movement of multiple targets. Chained-Center-Tracker achieves 52.3%, 61.2%, 59.6%, and 54.8% multi-object tracking accuracy and 36.8, 33.8, 33.6, and 9.9 Hz on the 2DMOT15, MOT16, MOT17, and MOT20 challenge datasets, respectively.
展开▼
