论文地址：Real-time part-based visual trackingvia adaptive correlation filters
Paper summary
The method in this article is called RPAC .
In this paper, a KCF-based part-based tracking method is proposed. By tracking each part, the final position and tracking target tracking size are synthesized. At the same time, when learning strategies, when the target is known to have been occluded, the iterative learner can be stopped and features can be tracked to avoid the introduction of "noise".
The introduction of multiple parts allows RPAC to dynamically change the size of the tracking frame.
Paper content
The traditional correlation filter tracker cannot solve the occlusion problem. If there is occlusion, it is easy to lose the tracking target. The part-based tracking method is proposed to better solve the problem of partial occlusion. The main theme of the Part-based method is that when the target part is occluded, the remaining part is still reliable.
RPAC's solution uses a kcf tracker for each part to track, and then uses a solution to merge each part. According to the PSR of each tracker, each KCF tracker can assign different learning rates to iterators based on the PSR.
Each part has its own weight coefficient, so that the more confident part contributes more to the target position . The weight formula is derived as shown in the figure below.
Since the tracking problem is long-term tracking, some methods are needed to solve the problem of iterator error learning caused by occlusion. The strategy used by RPAC is to use PSR and SCCM as the basis for occlusion judgment (through experiments, the author found that the tracking effect is unreliable when the PSR value is about 5, and the eye occlusion may occur when the SCCM is greater than 0.5), and when occlusion is detected , Then stop iterating the classifier and features; in this article, the threshold is set to 7. (Deduced by two unreliable tracking indicators and formula 5)
Finally, the confidence map obtained by formula 4 is used to track and judge the target; but I am not interested in this part for the time being, in the 3.4 part of the article.
Experimental results
KCF is 360 frames per second, and RPAC method is 30 frames per second.
