Adaptive ROI Encoding and Caching for Video Surveillance Streaming

Conventional video streaming techniques in security surveillance systems often utilize uniform bit rate strategies, leading to suboptimal resource allocation. We propose a novel streaming system that dynamically assigns variable bit rates to regions of interest (ROIs) within surveillance frames. By intelligently enhancing the quality of critical areas while reducing the quality of peripheral areas, the system significantly reduces network transmission and storage costs without compromising ROI clarity. The proposed system leverages a deep learning-based ROI detection module to effectively identify regions demanding intensive monitoring. An adaptive encoding scheme then assigns lower quantization parameters to ROIs, yielding higher quality, while applying higher quantization to non-ROIs to conserve bitrate. A dynamic quality enhancement module is integrated, significantly improving the quality of both foreground and background regions, thereby enhancing recognizability for surveillance personnel or machine analysis. A key innovation is a caching mechanism that exploits the high redundancy in surveilled background scenes across frames. By reusing enhanced background blocks from preceding frames, the caching mechanism accelerates the quality enhancement module with negligible quality loss. Extensive experiments validate the framework's superior rate-distortion performance. The proposed system's improved quality, optimized resource usage, and reduced storage make it a promising solution for advancing video coding in security and surveillance domains.