Machine learning-assisted network-aware bitrate prediction for remote driving

Abstract

Remote driving represents a promising technology with significant potential to provide more efficient transportation solutions. Its success relies on low-latency communication between remote operators and vehicles to provide real-time sensor information. In this paper, we propose a novel machine learning (ML)-assisted network-aware bitrate prediction framework to facilitate adaptive video streaming by adjusting its quality based on varying network conditions encountered during remote driving scenarios. First, we build a video streaming platform leveraging H.264 compression to gather data for training ML-based bit-rate prediction models. These models include Random Forest (RF), Gradient Boosting Machine (GBM), Light GBM (LGBM), AdaBoost, Deep Neural Network (DNN) and Extreme Gradient Boosting (XGBoost). Transmitted packet size, Packet Delivery Ratio (PDR), Received Signal Strength Indicator (RSSI), throughput, and successful packet transmission interval are used to predict bitrates. Furthermore, we compare the prediction accuracy of the models in terms of R-squared (R2 score), root-mean-square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE). Our results show that XGBoost is the most successful algorithm for bit-rate prediction. In terms of MAPE, XGBoost shows an enhancement of % 66.54, compared to AdaBoost.