Multi-Object Tracking Algorithm of Fusing Trajectory Compensation

Abstract

Multi-object tracking (MOT) is an important research topic in the field of computer vision, including object detection and data association. However, problems such as missed detection and trajectory mismatch often lead to missing target information, thus resulting in missed target tracking and trajectory fragmentation. Uniform tracking confidence is also not conducive to the full utilization of detection results. Considering these problems, we first propose a threshold separation strategy, which sets different tracking thresholds for similarity matching and intersection over union (IoU) matching during association to make the distribution of detection information more reasonable. Then, the missing trajectories are screened and compensated with the predicted trajectories to improve the long-term tracking ability of the algorithm. When applied to different association algorithms or tracking algorithms, a better improvement effect can be obtained. It can achieve high tracking speed while achieving high tracking accuracy on the MOT Challenge dataset.

Multi-Camera Multi-Vehicle Tracking Guided by Highway Overlapping FoVs

“…Classical tracking methods are mainly based on probability theory, with Kalman filtering and particle filters laying a solid mathematical foundation for tracking [38] problems. Modern SCMT trackers can be divided into the tracking-by-detection [39][40][41][42][43][44][45] paradigm and jointdetection-tracking [46,47] paradigm, with the majority of SCMT algorithms following the tracking-by-detection paradigm. Tracking-by-detection methods, prevalent in SCMT tasks, involve obtaining multiple detection boxes in each frame using an effective detector and then associating them based on appearance and motion cues.…”