Neural Network-Based Adaptive Multimedia Streaming for Vehicular Networks

The Adaptive Bitrate (ABR) algorithm is a fundamental technique employed in video streaming over the Internet. Its primary objective is to predict the current available bandwidth and select the optimal bitrate that the network can sustain, based on this prediction. However, most existing ABR algorithms, whether rule-based or deep learning-based, rely on application layer information for decision-making. This limitation confines the bandwidth estimation to the application layer, restricting the use of lower-layer network metrics for more accurate decision-making. Therefore, these algorithms may not be able to fully optimize performance under dynamic network conditions, especially when users are in an environment such as the Internet of Vehicles where bandwidth frequently fluctuates.To address this limitation, we propose CABR, a novel model that integrates transport layer information and video playback metrics into a reinforcement learning framework for ABR decision-making. We conducted simulations in a real network environment characterized by frequent bandwidth variations to evaluate the performance of CABR. The results were compared with those of the most commonly used ABR algorithms. Across all test scenarios, CABR demonstrated superior performance, achieving an average QoE (Quality of Experience) improvement of 7% to 10% over existing algorithms.