Bandwidth-Efficient Edge Video Analytics via Frame Partitioning and Quantization Optimization

Chuqin Zhou, Peng Yang, Zeyu Zhang, Chengzhi Wang, Ning Zhang

Abstract: The surging penetration of video cameras drives the rapid growth of video frames processed on the mobile edge. However, the scarce bandwidth and limited edge computing resources hinder edge video analytics at scale. Observing the non-uniform distribution of objects in one frame, we find the quality requirements and importance for video analytics of different regions vary across one frame. Hence, we propose Abate, a content-aware video coding and analytics scheme, to achieve bandwidth-efficient video analytics. This scheme consists of two phases, i.e., frame partitioning and quantization optimization. Taking structural features of frames into account, frames are subtly divided into small blocks in the first phase. Then, the quality of each of the blocks is properly controlled by quantization parameters based on contents in the block, taking into account the balance between video data volume and analytic accuracy. Extensive experimental results on real-world datasets show that, compared to existing benchmarks, the proposed system and algorithm can effectively save 50% bandwidth while achieving 10% higher accuracy.