Machine Learning Based Frame Classification for Videos Transmitted over Mobile Networks

Abstract

In wireless systems, content-aware MAC layer scheduling strategies contribute to supporting an adequate Quality of Service (QoS) and Quality of Experience (QoE) for the users. Content related information that can be used in such strategies include information about the video frame type (Intra-Predicted, Inter-Predicted or Backward-Predicted). Frame type identification and prediction in the MAC layer scheduler (as well as in other network locations) is hence a critical part of content-aware resource management and traffic engineering in video streaming. It is important to note that, when encryption is adopted at the upper layers, packet inspection is not possible. To address this issue, we propose an adaptive clustering and prediction algorithm. The algorithm uses unsupervised clustering combined with an adaptive classification approach to automatically group packets into frame clusters. Unlike conventional video traffic classifiers, our approach requires neither a-priori knowledge of video frames pattern nor a training set of frames. Instead, this approach automatically classifies packets into different frame types by employing simple statistical features. The proposed method continuously learns from the upcoming traffic thus the proper frame class labels can be easily discovered. Furthermore, the proposed method uses an autoregressive integrated moving average (ARIMA) algorithm to predict the upcoming traffic frame type. A large number of experiments has been carried out using data from several video samples. Results show the robustness and the effectiveness of our classification method, which is capable of achieving a detection rate of up to 97.3% for I frames and overall 2 identification accuracy of 91.3% (for all I, P and B frames).