An Adaptive Multi-Features Aware Correlation Filter for Visual Tracking

Xiangyue Zhang,Bin Hui,Zheng Chang,Qinghai Ding,Haibo Luo

doi:10.1109/access.2019.2942047

Abstract

In recent years, correlation filter (CF) based tracking methods have attracted more attention due to its low computational complexity and excellent performance. Most CF based tracking methods adopt CNN features of multiple layers to train the tracker for better performance. These methods fuse CNN features of multiple layers directly, and cannot make full use of the valuable information contained in the CNN features. In this paper, an adaptive multi-features aware correlation filter method is proposed. By extracting several basic features, different combinations of CNN features are formed. The proposed method can select an optimal feature combination for tracking adaptively according to the object appearance at the current frame. Experimental results show that the proposed method can track different challenging sequences robustly. By evaluating on the OTB-100 dataset, it can be found that the proposed method is advantageous compared with the state-of-the-art methods.

Highlights

Visual tracking is one of the most important and difficult tasks in computer vision
Visual tracking methods can be divided into two categories, one is based on the generative model and the other is based on the discriminative model
Different from the correlation filter (CF) based methods which directly fuse multiple features, we provide various combinations of different features

Summary

Introduction

Visual tracking is one of the most important and difficult tasks in computer vision. It is widely used in various fields such as video surveillance, automatic driving, robotic services and so on [1]–[5]. By giving the initial position and scale of the object, visual tracking can locate the object robustly [6]–[8]. Visual tracking methods can be divided into two categories, one is based on the generative model and the other is based on the discriminative model. The generative method models the object in the current frame, and finds a region most similar to the model as the predicted position of the object in the frame.

Objectives

Methods

Findings

Conclusion