Smart Visual Surveillance: Proactive Person Re-identification instead of Impulsive Person Search

Abstract

Simultaneous person detection and re-identification (re-id) on live CCTV camera feed can handle surveillance issues in a proactive manner instead of applying person search on CCTV video archives after occurrence of some incident. Current person re-id solutions do not deal with person detection and depend on off-the-shelf detectors for their practical use. Whereas person search approaches implicitly detect persons but these work for a given set of pre-cropped persons (query set) instead of keeping track of all on-the-spot persons. In this work we propose an autonomous solution for simultaneous person detection and re-id for all people appearing in a surveillance cameras network. A deep backbone network integrates Region proposal network and region of interest pooling to detect locations of persons in given surveillance images. Concurrently, proposed re-id stream emphasizes on mapping step-wise associations among respective local parts of a person and embeds this information into global person representations. The proposed method is evaluated on a public dataset PRW**PRW: http://zheng-lab.cecs.anu.edu.au/Project/project_prw.html, which is the only public dataset available to evaluate simultaneous person detection and re-identification. The proposed method shows superior performance as compared to the existing methods for particular query set in PRW. Additionally, the proposed method spots and identifies the rest of people (visible in given images and not part of the standard query set), with a great success.