Adaptive Kalman Filter with power transformation for online multi-object tracking

• Published in: Multimedia systems 2023-06, Vol.29 (3), p.1231-1244
• Main Authors: Liu, Youyu, Li, Yi, Xu, Dezhang, Yang, Qingyan, Tao, Wanbao
• Format: Article
• Language: English
• Subjects: Ablation; Accuracy; Algorithms; Computer Communication Networks; Computer Graphics; Computer Science; Covariance matrix; Cryptology; Data Storage Representation; Kalman filters; Multimedia Information Systems; Multiple target tracking; Operating Systems; Regular Paper
• ISSN: 0942-4962; 1432-1882
• DOI: 10.1007/s00530-023-01052-7

By introducing a low-score detection box association stage, the full-detection association can effectively enhance the accuracy and robustness of online multi-object tracking. However, this association would lead to a decline in tracking precision, the key point of which is that the fixed noise setting of Kalman Filter is difficult to balance the system requirements for high-score and low-score detection boxes. A Power-Adaptive Kalman Filter (PAKF) was proposed in this article. Taking the motion matching cost and confidence score as process and observation noise scale parameters, respectively, and combined with the power transformation, two adaptive factors were constructed to adjust the process and observation covariance matrices, respectively. Sufficient ablation experiments were conducted on the full validation set of MOT17. After introducing the PAKF into the ByteTrack and SORT, the High-Order Tracking Accuracy, Multi-Object Tracking Precision (MOTP) and ID F1 score of them were improved by about 1%, and their improvements were more obvious in complex scenarios. On the challenging HiEve benchmark dataset, after introducing the PAKF, the Multi-Object Tracking Accuracy and MOTP of the ByteTrack were improved by 0.53% and 0.28%, respectively. It is more advantageous than other state-of-the-art online methods. The proposed PAKF can effectively improve the performances of the multi-object tracking algorithms based on the Kalman Filter and tracking-by-detection. The codes are available at https://github.com/LiYi199983/PAKF .