Real-Time Person Re-Identification Using Omni-Scale Feature Learning Network and Yolov5: A Comparative Study

Abstract

Video-based person re-identification seeks to match video footage of an individual across non-overlapping multi-camera systems in real-time, probing for instances of the same identity appearing at different locations and times. The critical process in video-based person re-identification involves feature aggregation from the video track. This study introduces a method utilizing a convolutional neural network model named Omni-Scale Feature Learning Network (OSNet) for video-based re-identification. The performance of this method is evaluated on the large-scale MARS dataset and compared with other network models. Furthermore, a novel approach using You Only Look Once version 5 (Yolov5) is proposed for the first time for image, video, and real-time person detection and re-identification. This approach was trained on a custom-created dataset, gathered from two cameras capturing multiple identities of Computer Engineering students at the University of Basrah. The proposed method yielded promising results, with a re-identification accuracy of 80%. The aim of this work is to establish a real-time person re-identification system using the Yolov5 algorithm, and to contrast its performance with that of the OSNet.