Image Classification on the Edge for Fast Multi-Camera Object Tracking

Abstract This paper introduces a stochastic model for testing a low-latency method of tracking an object as it moves throughout an area observed by a dense network of video surveillance cameras. This new method utilizes the computing power of edge device to run lightweight image classifiers closer to the source of the video data. The sensor redundancy in wide camera networks allows us to increase the accuracy of local lightweight image classifiers to provide for a real-time estimate of a target's location in the sensing region. Running image classifiers on the edge eliminates the need to offload all video data to the cloud and would improve the latency issues inherent to offloading solutions.
Authors
  • Nick Nordlund (Yale)
  • Geeth de Mel (IBM UK)
  • Heesung Kwon (ARL)
  • Leandros Tassiulas (Yale)
Date Sep-2018
Venue 2nd Annual Fall Meeting of the DAIS ITA, 2018
Variants