An Adaptive Object Tracking Algorithm with Multi-Features Based on Correlation Filtering

Abstract

Aiming at the problem that visual object tracking algorithms based on correlation filtering using only Histograms of Oriented Gradients (HOG) feature has unsatisfactory tracking performance, an adaptive object tracking algorithm with multi-features is proposed based on background-aware correlation filters framework. An adaptive fusion module of HOG feature response and color feature response is constructed to improve the robustness of the algorithm in different tracking scenarios. A novel feature response evaluation index named significance of main peak is designed to enhance the accuracy of feature response discrimination and fusion. A model adaptive update module is presented to reduce the risk of model drifts and improve tracking performance. Through evaluating on OTB2015 dataset, experimental results show that the tracking algorithm has excellent comprehensive performance and can achieve more accurate real-time object tracking.