Correlation Filter Tracking Algorithm Based on Filter Space Constraint and Object Re-Detection

fast Discriminative Scale Space Tracking (fDSST) is an excellent object tracking algorithm with a good balance between tracking precision and running speed. In order to improve its long-term and stable tracking performance, especially its ability to recapture object, the fDSST_SrRed algorithm is proposed in this paper. The improvement details include: in order to reduce the negative impact of boundary effect on fDSST, a spatial constraint method of position filter is proposed; Aiming at the defect that the fDSST cannot relocate the object after the object is occluded and disappeared, an object re-detection method based on zero-filling of position filter was proposed. In order to judge when to start object detection, a tracking confidence evaluation index is proposed. Based on this confidence evaluation index, a dynamic filter update method is designed to make the filter learn more accurate object information and background information. Experiments show that compared with the baseline fDSST, proposed algorithm has a great improvement in the tracking precision of occluded and disappeared scenes. Compared with other correlation filter based long-term tracking algorithms (LCT, MUSTER and FuCoLoT), the proposed algorithm not only has good tracking precision, but also has the fastest running speed. The experimental results show that the improvement is effective.