Background Modeling for HEVC Compressed Videos using Radial Basis Network

Abstract

High Efficiency Video Coding (HEVC) or H.265 is a successor of H.264/AVC, which is designed to gain performance, ease parallel processing and achieve better compression ratio over the latter. We present a compressed domain background modelling technique utilizing residual information from inter predicted motion compensated coding blocks at HEVC encoder. We employ machine learning techniques to model the dynamic background using training sequence and test it on target video sequences with highly dynamic background content. Rather than operating on a pixel level granularity, the proposed method operates on 8×8 data blocks from residual frames. The method classifies each 8×8 block of the input frame as foreground or background. After block level segmentation, conventional background subtraction methods can be used on the foreground blocks for pixel level segmentation, resulting in reduced computation time and effective utilization of resources.