Intelligent Dynamic Adaptive Video Streaming over HTTP using Smart Adaptation and Machine Learning Solutions

Abstract

It is expected that by 2021, around 82% of all internet traffic will be video. Moreover, Nielsen’s Law of Internet Bandwidth states that “a high-end user’s connection speed grows by 50% per year”. Thus, the users’ bandwidth will reach approximately 1 Gbps by 2021. Bearing in mind the limitations in bandwidth, growth in the video stock market and the issues regarding expansion of the Internet, video dynamic adaptive streaming over HTTP (DASH) has become a hot topic. This is due to the flexibility and tuning that can be given to DASH-Clients whilst streaming video. This thesis aims to provide uninterrupted and smooth video streaming over HTTP. To achieve this aim, an intensive study is carried out as well as the segment size impacts are investigated deeply. The testbed used in this thesis consists of six different segment sizes (i.e. 1, 2, 5, 10, 15 and 20 seconds), with the original resolutions, and the original frame rate. It objectively evaluates the impact on the perceived video quality at the DASH-Client when the network bandwidth is capped to be equal to the highest given/encoded average bitrates. The results show that the Peak Signal to Noise Ratio (PSNR) score would be noticeably increased with the increase of segment size; thanks to the proposed solution which dynamically resizes the buffer (at the DASH-Client) to match it with the requested segment size, and force DASH-Client to seek the highest possible video quality (representation) from the beginning of the streaming session. However, it is concluded that although segment size of 5 seconds is not showing the best performance in terms of the both PSNR score and absolute smooth video playback, it shows a very satisfactory PSNR and reduced initial delay; thus, it is recommended to be used to elaborate with unstable issues while video is streamed. Due to the issues arise while streaming such as the increased stalls and quality degradation, this thesis sheds light on the improvements that can be gained when using the machine learning solutions in terms of the both the traditional and the Artificial Neural Network (ANN) classifications for the video requesting by the DASH-Client taking into account the average encoded video bitrate, actual/available bandwidth and the DASH-Client’s buffer size. The results show that machine learning can be a credential when they are applied for further improvements. Lastly, this thesis studies and examine various video codecs and it carries out an intensive investigation into the integration of DASH with most commonly used video coding standards (i.e. AVC, HEVC, VP9 and AV1). The testbed used consists of five different segment sizes (i.e. 1, 2, 4, 5 and 10 seconds), with differing encoded spatial resolutions, and the original frame rate. A dataset is introduced as well as objectively evaluation is carried out to examine the impact on the perceived video quality at the DASH-Client when the network bandwidth is capped to be equal to the highest encoded average bitrates. The results show that HEVC performs very well under different network and client’s conditions.