Full View Tracking: Multi-Camera People Tracking Based on Appearance and Spatio Temporal Feature

In modern unattended surveillance systems, multi-camera people tracking (MCPT) technology is of great significance for achieving continuous tracking of individuals within extensive surveillance areas. Although the target detection and tracking technology for a single camera has become relatively mature, multi-camera collaborative tracking still faces many challenges, including how to maintain the consistency of the identity ID of individuals across different camera perspectives, handling target occlusions, disappearances and reappearances, and the diversity and variability of appearance features. In response to these challenges, this paper proposes a multi-camera collaborative tracking system based on appearance and spatio-temporal information (which we call Full-View Tracking, FVT). The system integrates an improved BoT-SORT algorithm, TransReID feature extraction technology, and a global ID matching method to achieve stability and identity consistency within single-camera tracking. At the same time, through trajectory priority matching, appearance feature fusion, and spatio-temporal consistent ID reallocation, it effectively solves the issue of identity ID consistency across multiple cameras. Experimental results on the AI CITY 2023 dataset show that this method significantly improves MCPT performance, with an IDF1 score reaching 94.48%, ranking among the top contenders.