Region of Interest Encoding in Video Conference Systems

In this paper, we present a region of interest encoding system for video conference applications. We will utilize the fact that the main focus in a typical video conference lies upon the participating persons in order to save bit-rate in less interesting parts of the video. A Viola-Jones face detector will be used to detect the regions of interest. Once a region of interest has been detected it will get tracked across consecutive frames. In order to represent the detected region of interests we use a quality map on the level of macro-blocks. This map allows the encoder to choose its quantization parameter individual for each macro-block. Furthermore, we propose a scene composition concept that is merely based upon the detected regions of interest. The visual quantization artifacts introduced by the encoder thus get irrelevant. Experiments on recorded conference sequences demonstrate the bitrate savings that can be achieved with the proposed system. Keywords-region of interest coding; object detection; object tracking; scene composition; video-conferencing