Multiobject Tracking Algorithm With Adaptive Noise Calculation and Information Fusion in Crowded Environments

In crowded environments, frequent occlusions and camera movements introduce noise to appearance features extraction and the computation for cosine similarity of appearance features, which will lead to identifier (ID) switches and tracking loss. To address the above issues, this article presents a novel tracker based on a detection and tracking framework, named adaptive noise and fusion information tracker (AFTrack). First, dynamic feature points are accurately removed through cascaded RANSAC and USAC algorithms to achieve motion compensation for the Kalman filter. Subsequently, intersection over union (IoU) distance, Mahalanobis distance, and cosine similarity of appearance features are calculated between the trajectory state estimated by the Kalman filter and the detection state, and these distances are fused to compute a cost matrix. The cost matrix is globally linearly matched for data association. Finally, the observation noise covariance matrix is adaptively computed based on the confidence of the matched detection and is used in the update stage of the Kalman filter to achieve accurate estimation of future frames. AFTrack is tested on the classic multiobject tracking (MOT) datasets MOT17 and MOT20. The test results indicate that AFTrack achieves the highest scores in all metrics on the MOT20 dataset, with the multiple object tracking accuracy (MOTA) metric achieving the highest score of 80.6% in the MOT17 dataset, while the higher-order tracking accuracy (HOTA) and IDF1 metrics achieve second place. It demonstrates that AFTrack has good positioning and tracking accuracy in crowded environments and exhibits robustness in motion scenes and traffic environments.