TVV: Real-Time Visual Identity and Tracking with Edge Computing.

Video surveillance today has become pervasive, making visual identification and tracking technology attractive to a broad class of applications like traffic counting, crime tracking, and Blockchain. However, visual tracking is also a victim of the ubiquity of surveillance camera: a huge amount of data that generated by the cameras leads to severe congestion problem, which decreases the frame rate and in turn affects the tracking accuracy. In this paper, we present TVV, a real-time visual tracking system that leverages edge computing to support accurate and continuous tracking in large scale areas. The design of TVV is based on a insight that almost 80% of frames in a video stream exhibit high quality, and such frames can be processed on the edge nodes using a lightweight filtering method named KCF. Based on this insight, TVV adaptively load the visual tacking task on the edge or the server, based on the quality of the currently generated frame. In this way, the traffic load is largely decreased, without sacrificing the tracking accuracy. Our experimental result show that the average frame rate of TVV achieves 45.75 fps, outperforming most state-of-the-art visual tracking approaches.