Deploying Collaborative Machine Learning Systems in Edge with Multiple Cameras

Advancement in hardware capability has opened up the possibility of performing ML inference tasks at the edge using a large volume of sensory data generated from IoT devices such as cameras. As cameras become more pervasive, edge systems need to process streams from multiple sources with overlapping fields-of-view. In this position paper, we describe a collaborative sensing mechanism at the edge for such cases. We introduce a View Mapping Database (DB) that maps regions in a camera’s field of view to regions in other cameras’ view. We analyze characteristics of 5 video streams that capture an intersection from multiple angles, prototype a View Mapping DB, and present our preliminary results.

I. Introduction

The recent rapid advancement in hardware of edge accelerators has led to the deployment of cloud scale machine learning (ML) inferences to run locally, near sensory data sources such as cameras and microphones, without the need for sending large volumes of data to remote data centers. This leads to improved efficiency, latency, and throughput.