VBaR: Content-sensitive variable GOP length scheme

Abstract

The present work proposes a new strategy to analyze videos which incorporates viewer's interest to the studies of video content in the selection for intra-coded frames. This new approach studies the low-level video features to extract commonly browsed frames to enable random access while avoiding the overhead in transmitting periodic intra-coded frames. A tri-step process is proposed, where the first step discriminates of the original video stream by detecting the shot boundary and checking adjacent frame characteristics using low level features. The second step refines the initial frame selection by eliminating non-commonly browsed frames. The final step verifies the remaining frames if they are within a time stamp apart to maintain the efficiency of video delivery over low bit rate channel. A framework, named the Video Browsing and Retrieval (VBaR), is developed to analyze the original video stream and allows viewers to analyze the videos at their full control. Experimental results show the ability to effectively select intra-coded frames for video delivery, which could maximizes the resources available while fitting the viewer's browsing requirements.