TLD Target Tracking Fused with GMS Detection and Confidence Discrimination
Guo Wei,Yang Chen,Qu Haicheng,Xing Yuzhe
DOI: https://doi.org/10.3788/lop202158.0410009
2021-01-01
Abstract:In order to solve the problem of model drift in complex scenes such as occlusion and scale variation, this paper proposes a long-term target tracking algorithm based on TLD framework, which integrates GMS detection and confidence discrimination. First, in tracking module, the fast discriminating scale space correlation filter (fDSST) is used as the tracker, and the position filter and scale filter arc used to distinguish the position and scale of the target in the previous frame. According to the independence of the tracking module and the detection module in the TLD algorithm, the results of the tracking module arc input into the detection module, and the average peak-tocorrelation energy (APCE) is used to determine the template update to judge the confidence. In the detection module, GMS grid motion statistics is used as the detector to make the ORB algorithm with fast rotation invariance feature to match the target in the previous frame, and then the grid motion statistics is used to filter the matching results to achieve the rough positioning of the target position, and the target detection area is reduced dynamically according to the prediction position. Finally, the cascaded classifier is used to locate the target accurately. The results show that the tracking method proposed in this paper can greatly improve the tracking speed of the algorithm while effectively preventing model drift, and has better accuracy and robustness to challenging environments such as target occlusion, scale variation and rotation.