Region Of Interest (ROI) Based Rate Control For H.263 Compatible Video Conferecing

This dissertation presents a region of interest (ROI) based H.263 [10] [13] compatible video codec, which integrates the idea of object-based coding from MPEG-4 Visual [9] [15] into the traditional block-based H.263 codec. A face detection and tracking scheme with very low complexity is proposed to segment human face region from video conferencing sequences in real-time. For intra frames, the proposed detection method is a hybrid skin color and mosaic-rule based face detection and for inter frames the face tracking method uses motion vectors only. With the segmentation information, the ROI based codec and its associated rate control schemes are designed. By analyzing quadratic rate models in frame layer, VOP layer and macroblock layer extracted from the test data, a quadratic rate model at macroblock layer with a modified physical meaning is proposed to improve the model accuracy. The basic idea is to use a group of un-coded macroblocks in the current VOP instead of individual macroblocks to achieve the rate model and update model parameters. Based on this proposed rate model, some new features are designed to use both average statistics of a VOP and individual statistics of MB at the same time in order to achieve more efficient rate control performance. For CBR video, a joint VOP layer and MB layer rate control algorithm is proposed. The performance is compared with conventional TMN8 rate control as a baseline, and object based VM8 rate control as well. The comparisons show that the proposed algorithm can achieve more accurate rate control and better PSNR for ROI based on several video sequences. The proposed rate control algorithm is applied for variable bit rate (VBR) video case also, together with a frame layer rate control scheme. The efficiency of the proposed rate control is proved by simulation. TMN8 [6] is adopted as the platform, and the Modified Quantization Mode in Annex T of H.263 [69] is adopted to achieve flexibility in assigning quantization parameters among different macroblocks.