Collaborative Multi-Object Tracking as an Edge Service using Transfer Learning

Abstract

Video cameras have been pervasively deployed in Smart Cities and timely processing of the huge amount of video data brings up new challenges. Each individual camera is only able to cover a limited range due to the effective monitoring distances and the complicated layout of a building or an area. Once the object of interest moves away from the camera field of view, the system may lose object tracking. When the object appears and is captured by another camera, object detection, feature abstraction, and behavior classification require an association with other video data exploitation results. Hence, pragmatic approaches for object re-verification (or re-acquisition) are needed so as to not infer the object as a new track. In this paper, we propose a novel Collaborative Multi-Object Tracking system as an Edge service (CoMOTE) based on transfer learning technology. A transfer learning-based human detection and tracking method is introduced, which allocates most of the computing tasks on edge devices. An edge server is adopted to integrate information from edge cameras for decision-making to enable multi-view surveillance. The CoMOTE scheme not only reduces the dependence on a central server but also leads to a lower communication cost. Implemented on edge nodes (Raspberry Pi 4) and tested with WildTrack multi-view video database, extensive experimental results verify that the CoMOTE prototype achieved the design goal and outperforms multiple state-of-the-art methods.