A Review of Target Tracking Based on Deep Learning

Target tracking has always been a research hotspot in the field of computer vision. This article focuses on single-object tracking and introduces the application and development of deep learning in target tracking. Single-object tracking, as a technique that analyzes the motion state and provides localization for individual objects, has been widely applied in areas such as human-computer interaction, autonomous driving, and intelligent surveillance. However, practical applications have been slow to develop due to challenges such as lighting variations, object occlusion, and background interference during the tracking process. In recent years, with the significant success of deep learning in image classification, semantic segmentation, and object detection, an increasing number of researchers have applied deep learning frameworks to target tracking tasks. This paper provides a comprehensive review of popular deep learning-based target tracking algorithms from six perspectives: traditional tracking methods, correlation filters, convolutional neural networks, siamese neural networks, generative adversarial networks, and attention mechanism networks. Additionally, it summarizes the most representative datasets in the field of object detection, analyses the representative experimental results and explores the future trends in the field of single-object tracking.

I. Introduction

The target tracking algorithm is employed to determine the location of an object within a series of consecutive video frames. It utilizes the feature information extracted from the initial frame of the video to estimate and track the object's position in subsequent frames. As an important branch of computer vision, target tracking is widely applied in areas such as intelligent video surveillance, human-computer interaction, robot vision navigation, virtual reality, and medical diagnosis. With continuous research efforts, visual object tracking has made breakthrough progress in the past decade, going beyond traditional machine learning methods. It has incorporated deep learning (neural networks) and correlation filters, among other techniques, to achieve robust, accurate, and stable results.