An Analysis-Oriented ROI Based Coding Approach on Surveillance Video Data

Abstract

Driven by the growing amount of surveillance video data, intelligent video analysis has been applied to manipulate the stored videos automatically. As the most important method of video storage, the conventional coding method with high compression ratio severely degrades the video quality, which restrict the performance of intelligent analysis. In this paper, we propose an analysis-oriented region of interest (ROI) based coding approach to relieve this problem. We qualitatively analyze the effect of video compression on the performance of video analysis, such as feature similarity and object detection. Based on the analysis, we generate the ROI by the prior knowledge of interest objects rather than considering the characteristics of Human Visual System (HVS). Then, a weight-based rate control scheme is proposed to protect the quality of ROI by assigning more bits to encode it. Experimental results show that the proposed approach can reach 5.52% and 4.39% gains average over HEVC on the performance of feature similarity and object detection respectively under the same bitrate.