Person re-identification on lightweight devices: end-to-end approach

In this work, an efficient and real-time person re-identification system based on an affordable hybrid framework was presented. The proposed pipeline consisting of human detecting, tracking and extracting features was developed based on lightweight deep neural models so that they could be computationally accelerated on limited hardware resources devices. A comprehensive and substantial dataset has been established aiming to facilitate the training and evaluation of a surveillance system implemented to monitor individuals in an indoor environment. The proposed processing pipeline was implemented on both low-cost devices as Nvidia Jetson Nano and Google Coral. The experimental results indicated that the system could achieve real-time performance with up to 29 FPS and 0.96 mAP for the person detection algorithm task via edge devices, whereas a comparable accuracy was reached on the proposed feature extraction model with 0.85 mAP.