C-BIOU tracker multi-target tracking paper notes
Hard to Track Objects with Irregular Motions and Similar Appearances? Make It Easier by Buffering the Matching Space
Paper link
C-BIou tracker : The main improvement point is to use the original sort, replace iou with Biou in the matching stage, set the scale to expand the size of the matching frame, and use the cascade matching mechanism to set small buffer & large through the parameter b (Buffer Scale) The buffer matches the activated tracker for the first time and the unmatched tracker for the second time respectively, and during the KF prediction process, if the tracker is not updated at this time, it will directly use the previous matching trajectory state to make an average prediction for the next Frame trajectory status, do not use KF prediction, because the author believes that KF is already inaccurate at this time, and using KF prediction will increase time-consuming!
1. The author guesses that the tracking failure is due to two reasons:
(1) The detection and tracking frames of the same object do not overlap, resulting in matching failure, which may be due to fast motion; (2),
inaccurate motion estimation or some nonlinear and irregular motions cause the detection and prediction frames to not match.
2. Improvement method:
(1) Use Biou to replace iou, and increase the matching space to alleviate the non-overlapping problem. The author proposes to use BIou to reduce the impact of irregular motion in two aspects: one is to directly match the same but non-overlapping detections in adjacent frames and trajectories, and the second is to compensate for motion estimation bias in the matching space.
(2) Use cascade matching, use a small buffer in the matching stage to match the predicted trajectory and detection, and use a large buffer in the unmatched upper trajectory stage to further improve the matching degree.
ii1: motion prediction
In the state where the trajectory is not updated, KF is not used for prediction, and the first n frames are used as an auxiliary linear average motion (the method is simpler, a bit similar to a sliding average)
where o=(x,y,w,h) indicates the detection result , s represents the estimated state, △ represents the number of mismatched frames, and n is a hyperparameter (used to calculate the average speed in a certain period). A target is updated in the tth frame, and then there is a continuous △frame mismatch, then the state s after the t+△frame is the detection result of the tth frame plus the number of mismatched frames△*the average speed of the previous n frames.
Note: In the non-updated state, only the mean is effective, and the prediction in the basic sort is also updated according to the 1x8 matrix. The above formula is only the average of the prediction trend recorded in the previous frames to do the speed weighting of the linear prediction. oIn fact, it can be used as c_x, c_y, ratio, h, that is, the first four dimensions in sort, and the last four dimensions are the average speed of update.
ii2: BIou calculation method
The flow chart of the tracking framework (roughly the same as sort)
where parameter b: small buffer is set to 0.3, and large buffer is set to 0.5.
Finally, look at the author's experiment: for multi-target and irregularly moving targets, the number of tracking ID-switches is significantly reduced, and there is a chance to change the source code and try it out!