Enhancing Pedestrian Re-identification and Tracking System Using 3D Correlation Method
Chih‐Hsien Hsia, Justin Liao, Chao-Ming Chen · 2023
With the proliferation of surveillance cameras and the maturation of artificial intelligence (AI), the information security industry has witnessed rapid advancement. When multiple cameras capture non-overlapping areas, the re-identification (ReID) of pedestrians presents substantial technical hurdles. Thus, we introduced a pedestrian ReID and real-time crowd-tracking system for scenarios involving multiple camera viewpoints and the streaming of images from various angles. The system effectively detected and predicted the trajectories of monitored individuals as they traversed through different cameras. We employed YOLOX to detect the positions of pedestrians within the frames of individual cameras. Subsequently, the detection outcomes were processed with the BYTE-track algorithm for trajectory tracking. The ensuing tracking trajectories were transformed using calibrated matrices to project coordinates from screen space into spatial space. This integration amalgamated coordinate trajectories from various camera perspectives onto a global plane map. The trajectories of pedestrians detected by each camera were processed to use their temporal, spatial, and positional attributes. These 3-D features were employed for approximating distance computations. Through this process, the trajectories of the same individual captured by different cameras were integrated and displayed on a plane map. Consequently, a comprehensive tracking of the global movement paths, as captured by multiple cameras, was obtained. The results indicated that this method effectively enabled multi-camera tracking of numerous objects, outlining distinct trajectories for different tracked pedestrians on the plane map. By establishing global trajectories, notable enhancements in the performance of pedestrian ReID search tasks were observed.