CAC: Content-aware Adaptive Configuration for Edge Video Analytics

Edge video analytics, which offloads video stream processing tasks to edge servers, is a key enabling technology for deep neural networks (DNNs)-based intelligent video applications. Unfortunately, it often suffers from dynamic network conditions that incur long transmission latency. Recent efforts focusing on adaptive configuration, however, rarely consider the content dynamics, resulting in suboptimal configuration decisions. In this paper, we present CAC, a Content-aware Adaptive Configuration for edge video analytics supporting both network and content adaptation. CAC employs deep reinforcement learning (DRL) techniques to learn the control policy for configuration tunning that optimizes application performance while accommodating fluctuations in network conditions and video content. Specifically, to adapt to the content dynamic, we quantify the relationship between content dynamics and the configuration settings in terms of impact coefficients to instruct the configuration decision. We compare CAC to state-of-the-art methods on a variety of real-world network traces. Results show that CAC outperforms the best state-of-the-art scheme, with improvements in average Quality of Experience (QoE) of 3.56%, 40.6%, and 59.6% in three real-world applications.

I. Introduction

The proliferation of video sensors and the significant advancement in deep neural networks (DNNs) have fostered a plethora of video analytics applications, such as pedestrian detection [1], traffic surveillance [2], and object tracking [3]. Advanced DNNs usually require intensive computation, which can easily drain the computing, storage, and battery resources on end devices that capture the video. Conversely, conventional cloud offloading may introduce prolonged latency and privacy concerns due to the long-distance data transmission. As a result, edge video analytics emerges as an effective solution [4]–[6], that offloads the streaming data to nearby edge servers for real-time video analysis.