Visual object tracking has drawn great attentions and made great achievements during recent years. We systematically review the representative algorithms and evaluate the effectiveness of different trackers.

The advantages of this work are:

(1) We categorize recently proposed trackers into CF-based trackers and DLN-based trackers at the top level, as the trend of tracking methods is changing from correlation filters to deep learning networks.

(2) In order to provide a clearer classification and more valuable information for future exploration, we creatively summarize the DLN-based trackers from three aspects: network architecture, network procedure and network training.

(3) 25 state-of-the-art trackers are evaluated on four widely accepted benchmarks, and we further analyze the performance of trackers in terms of different attributes and conduct contrastive experiments based on intuitive visuals and features.

(4) We summarize our investigation and shed light on future trends in the field of visual object tracking.



To provide a clear overview of current progress, we divide the tracking methods into two top-level categories, Correlation Filter(CF)-based trackers and Deep Learning Network(DLN)-based trackers. We present a very detailed summary with an in-depth analysis of each category, especially for DLN-based methods. The taxonomy of trackers are showed in TrackerDetail .



We have chosen 24 representative visual object tracking methods and carried out comprehensive experiments on 4 publicly available object tracking benchmarks, including OTB100, VOT2018, UAV123 and LaSOT.

The detailed experiment results are showed as:

(1) Experimental introduction

(2) Quantitative Evaluation

(3) Attribute-based Evaluation

(4) Contrastive Analysis

(5) Speed Analysis

(6) Raw results


Laboratory of Intelligent Media Computing and Network Security,HUST. ©  2021. All rights reserved.