An Adaptive Multi-scale Tracking Method Based on Kernelized Correlation Filter

Abstract

Visual tracking is a fundamental computer vision task with a wide range of applications. Kernelized Correlation Filter (KCF) is an excellent algorithm with high tracking speed. However, the target tracking scale in the KCF algorithm is a fixed value which might cause tracking failure or target drifting problem when the target scale changes significantly. In this paper, we present an adaptive multi-scale tracking algorithm based on the KCF algorithm by estimating the scale of the target. Our method builds upon the correlation filter with a Gaussian kernel and reasonable prediction of the target size. In order to verify the effectiveness of the proposed algorithm, 9 sets of complex video sequences of a commonly used tracking benchmark were selected and the results were compared with other tracking methods (KCF, CSK, CT, TLD, Struck, CNN-SVM and MDNet). The results show that the proposed method has high accuracy. The method in this paper has strong robustness in the complex scenes with challenges of scale variation, illumination variation, occlusion, in-plane rotation, out-of plane rotation and deformation.