Research And Implementation Of Reinforcement Learning Based Network Adapttive Realtime Video Transport System

Because of best-effort characteristic of current network, it is difficult to guarantee the quality of real-time video transportation and the way that can adapt video coding process to dynamic network conditions appropriately has been one current research focus. This dissertation proposes a novel video coding control algorithm that is based on reinforcement learning and the system reveals good network adaptive performance.This dissertation mainly focuses on the following works. Firstly, we carried out an in-depth study on the whole process of video codec, video transport, and quality analysis during which we chose H.264 as the codec and RTP as the transport protocol. Based on these tools, we built an optimal real-time video transport system in an ideal environment. Secondly, we analyzed the character of several typical networks in terms of delay, jitter and packet lost rate, and constructed corresponding simulation networks using NS2. With these simulation networks, we studied the influence of these different network characteristics on real-time video transport process and proposed some simple but advisable method to improve video transport quality. Thirdly, we conducted an in-depth research on the way of applying reinforcement learning algorithm to video transport control and implemented a network adaptive real-time video transport system. The system utilized the network state information which RTCP feedbacks and video complexity information to estimate current environment status and adjusted the bit rate of output video stream of H.264 codec accordingly. At last we conducted a series of test on different simulation networks and found that our system exhibited a good network adaptive property and produced better video quality than existing systems.