ByteTrack
Main methods:
1. For all detection frames in the current frame (with confidence higher than the minimum threshold), divide them into high-confidence detection frames and low-confidence detection frames based on the intermediate confidence threshold. (Almost all detection frames are retained for trajectory matching)
2 For high-confidence detection frames, it is easy to associate them with previous trajectories (pseudocode lines 17-19)
(2.1) Use the detection frame of the current frame and the Kalman filter prediction (or estimation) result of the previous frame, and use IOU to calculate the similarity (or if ReID is used, the distance between features will be calculated to calculate the similarity).
(2.2) The Hungarian algorithm is used for matching based on similarity.
(2.3) It is necessary to retain the high-confidence detection frames that do not match the trajectory (i.e., the remaining detection frames of the current frame) and the trajectories that existed in the previous frame that did not match the detection frame (i.e., the remaining trajectories of the previous frame).
3 For low-confidence detection frames (pseudocode lines 20-21)
(3.1) is first associated with the remaining trajectories in (2.3). In fact, it is the association between the current low-confidence detection frame and the remaining trajectories in (2.3). It is worth mentioning here that the association here uses IOU to calculate the similarity, because low-confidence detection frames are usually blurred or blocked, and ReID is not reliable at this time.
(3.2) To retain the trajectories that still do not match the detection frame in (3.1)
(3.3) Delete low-confidence detection frames that did not match the trajectory in (3.1). These low-confidence detection boxes are considered not to contain the target.
4. High-confidence detection frames that do not match the trajectory are saved as new trajectories. (pseudocode 23-27)
BYTE pseudocode diagram:
ref
ref1: https://zhuanlan.zhihu.com/p/613223235