AdaptTrack: Multi-Object Tracking by Adaptive Correlation
Kunpeng Li, Lei Wang, Hongjun Ren, Yangjie Cao · 2024
The main idea of Multi-Object tracking is to accurately identify and track the position and motion status of multiple objects in a video sequence in real time. The current mainstream tracking algorithm is to associate the detection boxes with the trajectory in a comprehensive and violent way, as a way to accurately determine and maintain the unique identity mark of the moving object. However, implementing a unified association process for detection boxes with different confidence levels, this suffers from non-negligible low information utilization, long association time, and high resource consumption. To solve this problem, we propose a faster, simple and effective association algorithm ADAPT, dividing the queue of detection boxes and trajectories by confidence level to achieve adaptive association of detection boxes and trajectories in the queue, so as to realize tracking, shorten the association time, and improve tracking effect. When applied to six different trackers, our method achieved significant improvements in both MOTA and IDF1 scores. In order to improve the MOT performance, we designed a simple and powerful tracker, AdaptTrack, and evaluated it on 3090GPU for VisDrone-MOT datasets, achieving 41.3 MOTA, 53.6 IDF1, and 42.9 HOTA. AdaptTrack on UAVDT datasets also shows the same excellent results of 69.8 MOTA, 78.9 IDF1, 66.3 HOTA, and achieves a tracking speed of 18.2 FPS.