Scale Adaptive Kernel Correlation Filter Tracker with Multi-Feature Fusion

Abstract

In order to solve the problems of illumination variation, scale variety and target occlusion during the process of tracking, the traditional kernel correlation filter tracking algorithm uses a single feature to track, which is easily affected by noise, so the performance is greatly limited. To this end, we propose a new multi-feature fusion visual tracking model, which combines the Color-naming (CN), Histogram of Oriented Gradient (HOG) and Lab spatial features after dimensionality reduction, and trains better classifier model to improve the tracking performance of targets in complex environments. In addition, we have proposed a simple and effective scale estimation method that can better adapt to scale change. In the face of occlusion, we use the depth difference between the two frames and the maximum response value of the classifier to judge the occlusion. Then we use the simplified random ferns detection algorithm to recover tracking. The experimental results on Princeton dataset show that the tracking accuracy and success rate of the proposed algorithm compared to other three tracking algorithms are greatly improved.