Rate Adaptation For Low Latency Real-Time Video Streaming

We study video rate control as a means to minimize latency for a real-time video stream while ensuring good video quality. When the channel capacity changes, we react by changing the video rate appropriately such that latency does not degrade significantly and, at the same time, capacity is not severely underutilized. We setup the underlying control problem as a Markov Decision Process and obtain two different policies using a greedy approach and the Stochastic Gradient Descent (SGD) method. We compare the two policies by implementing variants of them over a testbed comprising FFmpeg video codec transmitting over WiFi and wired networks. Our policies require estimates of bottleneck channel capacity and queue size at the video source. We devise a network-agnostic technique to estimate these parameters. We evaluate the performance of the two policies first in a controlled setting over a wired network and then in an uncontrolled setting over WiFi 802.11n network. For a 30 fps video, we conclude that when the capacity is fluctuating over short intervals of time, which is the case in real wireless networks, though the greedy approach achieves slightly higher capacity utilization, but SGD performs significantly better in achieving lower latency while still maintaining a good video bit rate.