Moving segmentation in HEVC compressed domain based on logistic regression

Abstract

Moving segmentation comes first in the pipeline of video analysis system. Although pixel domain methods are fully exploited, compressed domain methods still remain few relatively, especially for HEVC. This paper proposes a novel moving segmentation method in HEVC compressed domain based on logistic regression which can handle with camera jitter situation. Firstly, distinction between HEVC syntax elements in moving foreground and background is fully analyzed. F-score evaluation is performed to confirm the discrimination capacity of these features. Then these raw features, including prediction mode, MV, etc., are pre-processed into more useful and robust features. Noise is reduced by spatio-temporal filter and global camera motion is removed in this step. Afterwards, moving segmentation is modeled as a binary classification problem which takes the pre-processed features as input. To solve the classification problem, logistic regression is utilized in this paper. Experimental results present that the proposed method achieves comparable performance with state-of-art pixel domain method and 16.78 times faster processing speed.