A Cascade Collaborative Offloading Framework for Video Analytics Based on Online Learning

Abstract

Offloading has already become an essential component of the video analytics system due to its low latency and energy saving. The generic traffic offloading strategy allocates video tasks according to computation resources without considering video information, which leads to incorrect decisions that can not guarantee bandwidth reduction and accuracy in video analytics. Therefore, to alleviate the problem, we propose a Cascade Collaborative Offloading Framework (CCOF) in the paper. CCOF is a cross-edge framework, and it coordinates mobile devices, edge servers, and cloud servers to complete video analytics tasks. First, we model the traffic offloading problem in a video analytics scenario as an optimization problem and propose the CCOF as a solution. Then, we propose a periodic online learning algorithm that can update the model parameters with stream data. Afterwards, we adopt the online algorithm and the sharing-based classifier on the mobile devices and edge servers to filter out unnecessary frames and reduce the bandwidth cost. Furthermore, we feed the recognition results back to the edge and mobile devices for online learning. Evaluation of dataset Jackson shows that the proposed method reduces the bandwidth cost by an order of magnitude and achieves 10x speed-ups on video analysis tasks.