Object-Based Resolution Selection for Efficient Edge-Assisted Multi-Task Video Analytics.

Chengzhi Wang,Peng Yang,Jie Lin,Wen Wu,Ning Zhang

DOI: https://doi.org/10.1109/globecom48099.2022.10000829

2022-01-01

Abstract:Camera-based monitoring is becoming increasingly popular, as multi-objective detection tasks can be enabled by video analytics over captured frames. Yet, video frames have to be delivered to computation-capable edge nodes for further processing, because the amount of required resources exceeds the capacity of built-in hardware of video cameras. In this paper, observing that video resolution directly determines the subsequent bandwidth and computing resource consumption, as well as the analytic accuracy, we propose an edge-assisted object-based resolution configuration algorithm to achieve efficient multi-task video analytics. The proposed algorithm harnesses the diversity of neural networks used for detecting different objects in one frame, which brings about two-fold possibility for bandwidth saving. On one hand, background information cannot be indiscriminately transmitted, as is unlikely to contribute to improving the analytics accuracy. On the other hand, fine-grained resolution selection allows object-level optimal resolution that minimizes the transmitted data volume under accuracy and latency constraints. Simulation results demonstrate that the proposed method can effectively reduce up to 50% of the transmitted data volume, compared to existing benchmarks.