A cross-camera multi-target tracking algorithm for substation scenarios

With the rapid evolution of smart grids, the need for unattended operation of substations has grown increasingly urgent, and cross-camera multi-target tracking technology has emergedas a crucial element in guaranteeing reliability and security of intelligent monitoring systems. In this paper, target detection based on improved You Only Look Once Version 8 (YOLOv8) for target detection and feature matching based on improved re-identification algorithm are integrated into the Bag-of-Tricks-SORT (BOT-SORT) multi-target tracking algorithm to tackle the difficulties of detecting small targets and frequent target occlusions in substation scenarios, single-camera multi-target tracking is achieved. The global feature pool is introduced for feature matching of tracked targets, and the time–space association strategy is designed based on the relationship between the motion pattern of targets and the scene topology to achieve cross-camera multi-target tracking in the substation scene. The outcomes of the experiments demonstrate that the algorithm put forward surpasses existing methods: the enhanced YOLOv8 achieves a mean average precision (mAP50) of 97.9%, the optimized re-identification model reaches a Rank-1 accuracy of 72.5%, and cross-camera tracking obtains a multiple object tracking accuracy (MOTA) of 82.8%.This research effectively enhances the intelligent monitoring capabilities of substations, thereby providing critical support for the development of unattended smart grids. An improved YOLOv8 framework integrating a small object detection layer and Global Attention Mechanism (GAM) is proposed, significantly enhancing small target detection accuracy in complex substation environments. The domain-general re-identification (DG-ReID) algorithm is optimized with hybrid data augmentation, domain-adversarial neural networks (DANN), and hierarchical learning rates, improving robustness against cross-camera variations. A global feature pool with spatio-temporal association is developed to realize cross-camera identity matching, effectively reducing false matches through motion and substation layout constraints. The proposed cross-camera multi-target tracking algorithm achieves superior performance in substation scenarios, with 82.8% MOTA and reduced ID switches compared to existing methods. Comprehensive ablation and comparison experiments validate the effectiveness of each improvement, demonstrating the algorithm's suitability for intelligent substation monitoring.