End-to-end person re-identification: Real-time video surveillance over edge-cloud environment

Abstract

In this work, a real-time end-to-end person re-identification (Re-ID) system is developed for a realistic unconstrained open-world environment. The raw surveillance videos are analyzed to provide the ranked list of person images matched with the query at the user-end. Two frameworks based on distributed-computing and edge-based computing that utilize the NVIDIA Jetson Nano and TX2, respectively have been developed to realize the objective. Various practical aspects of deploying an end-to-end Re-ID system over edge-cloud environment, such as resource constraints at the edge, hardware allocation to individual tasks and their acceleration, communication costs and scalability are dealt with. An improvised omni-scale network (iOSNet) with a powerful feature aggregation scheme has been proposed to handle more complex variations in the appearance of a person. The performance of the system has been evaluated on our own dataset and compared with other existing datasets, where state-of-the-art accuracy is achieved on all the datasets.