Long-term Tracking Based on Kernelized Correlation Filtering

Abstract

Aiming at the problems of kernelized correlation filter tracking algorithm such as scale change, severe occlusion and tracking failure under similar target tracking, a long-term target tracking algorithm based on KCF is proposed. The algorithm introduces color information into the tracker, and the scale information makes the algorithm adapt to the more complex situation when tracking. It is worth noting that the trackers train two models, one to represent coordinates and the other to characterize confidence. In addition, training online support vector machine classifiers is also crucial. This article uses a multi-expert tracking strategy to enable longterm tracking to be re-detected in the event of a failed tracking condition. The validity of the proposed algorithm is verified by the OTB-2013 [1] evaluation sequence. Compared with several other classical algorithms, this algorithm has been significantly improved. In the target occurrence scale, occlusion, deformation and other interference cases with strong robustness.