EAVA: Adaptive and Fast Edge-assisted Video Analytics On Mobile Device

Mobile video analytics applications, such as smart driving, VR/AR, and video surveillance, have become increasingly popular due to the proliferation of mobile devices. These applications typically use compute-intensive Deep Neural Networks (DNNs) inference in real-time and require high accuracy. Recent studies have shown that edge computing can significantly improve the performance of these applications by offloading the computation, particularly neural network inference, from mobile devices to nearby edge servers. However, offloading continuous video streams to edge servers still faces the challenge of scarce and variable network bandwidth, resulting in high latency for mobile deep vision applications. Existing works often assume sufficient networks and powerful servers to offload all streaming computation to the edge, resulting in unsatisfactory performance in practical scenarios. In this paper, we propose EAVA, an adaptive Edge-Assisted framework on mobile devices designed for Video Analytics that considers a more practical edge situation with an unstable network environment and multiple DNN model choices. EAVA initially partitions video frame and combines mobile devices with powerful edge servers, allowing these frame partitions to parallel perform video analytics algorithms on local devices or edge servers. To handle the intricate network and inference model dynamics, EAVA trains a deep reinforcement learning model to optimize the Quality of Experience (QoE) for mobile deep vision applications, making adaptive configuration choices. Without relying on preconceived assumptions about the environment, EAVA makes optimal choices based on experiences. Finally, we implement and thoroughly evaluate the performance of EAVA using diverse real-world network traces, demonstrating its superior advantages over existing state-of-the-art solutions.