An Edge-Cloud Collaborative Object Detection System

Edge computing system usually consists of the lightweight neural network to preprocess the video stream, and then transmits the intermediate data to the cloud for video analysis, which not only ensures the real-time performance of video processing but also greatly reduces the WAN bandwidth consumption. However, many existing edge processing systems sacrifice video processing accuracy to reduce intermediate transmission volume or reduce processing delay. Therefore, the leveraging of accuracy and latency places a challenge on how to deploy the network on the edge device and set the pre-processing parameters. This paper builds a real-time video stream processing system, then tries to achieve the balance between the cost and benefit of edge preprocessing by designing a dynamic configuration algorithm for optimal preprocessing deployment to achieve low latency, low transmission, and high precision real-time video processing.