Scale-Adaptive Regression Position Prediction Tracking

Abstract

Traditional kernelized correlation filter tracking methods use the target position in the current frame to estimate the moving target initial position in the next frame. For fast moving target, these methods lose the target easily. To cope with this problem, a novel scale-adaptive regression position prediction tracking approach is proposed. This algorithm employs regression prediction method to predict the initial position in the next frame. Then the kernelized correlation filter method is utilized to obtain the final target position. For further improving the accuracy and robustness, we exploit a scale pyramid model to estimate the target scale. Experimental results over 10 benchmark sequences demonstrate the proposed approach performs favorably against the state-of-the-art tracking methods.