mVideo: Edge Computing Based Mobile Video Processing Systems

Abstract

Computer vision is widely used to detect anomalies in video processing systems for public safety. Applying Deep Neural Networks (i.e., DNNs) in computer vision can achieve a high detection accuracy but it requires a huge amount of computing power, storage space, and video data. Thus, DNNs-based video analytics is mostly deployed in the cloud with video data steaming from a set of stationary cameras. There are mainly three issues in this setting. First, steaming a huge amount of video data from cameras to cloud leads to high bandwidth consumption and latency. Second, when DNNs are deployed on resource-limited devices like edge nodes to reduce communication costs, it is hard to achieve a high detection accuracy. Third, stationary cameras can only collect a limited amount of video data that covers a small area, so it barely satisfies the needs of the real-time analytics in applications like public safety. We propose a mobile edge computing-based video stream processing platform, mVideo, which conducts video analytics making full use of resources at the collaborative edge and cloud nodes. On the mVideo, a mechanism is designed to partition a video analysis task based on available resources on the mobile edge node. Then, the edge nodes pre-process video data using a lightweight DNN model and upload the results to cloud nodes for further analysis. Thus mVideo not only collects video data that covers a large area, but also reduces the communication costs. To validate the proposed platform, a face recognition application is deployed on the mVideo prototype. Experimental results reveal that compared with the existing cloud computing model, mVideo reduces video data volume transmitted to the cloud nodes and power consumption by up to 99.5% and 96.2%, respectively. mVideo also improves the execution time by 90.0% to optimize mobile video analytics performance.