Graph-Based Cross-Camera Continual Tracking: Improvements to Graph Structure and Feature Extraction
Yujia Pu, Sijie Fan, Yang Yuzhen, Ruichun Tang · 2024
Target tracking is a significant issue in computer vision. Cross-camera continuous target tracking, on the other hand, is a challenging task that aims to achieve continuous tracking of the same target among multiple cameras. However, most existing methods focus primarily on improving target detection performance and less on local relationships between objects. To address these restrictions, this paper proposes a graph neural network-based target tracking method as the core solution for cross-camera scenes. In this method, a more complex graph model is utilized for graph construction of detection results, considering historical target information. Additionally, the feature extraction network is improved by incorporating a two-channel attention mechanism, enhancing the network's representation capability. Furthermore, constraints are added in the post-processing stage to cluster tracking results from multiple cameras, improving the accuracy of tracking outcomes. Extensive experiments have been conducted on public datasets to evaluate the proposed method. The results demonstrate that our approach surpasses traditional methods in terms of accuracy and robustness, successfully achieving continuous tracking of targets in complex cross-camera scenarios.