Kernelized Correlation Filter Tracking with Scale Adaptive Filter and Feature Integration

Abstract

Recently, kernelized correlation filter (KCF) has been a popular tracker for high accuracy and robustness with high speed. However, KCF tracks objects with a fixed size template without scale estimation, causing tracking failure during target scale changes because of learning background or local appearance of the target. In this paper, we incorporate a separate scale filter into KCF tracker with feature integration. Experiments have shown that our tracker outperforms KCF and other scale adaptive trackers on distance and overlap precision while attaining relatively high speed.