A Bayesian inference based adaptive lane change prediction model

Abstract

Abstract Predicting lane change behavior of surrounding vehicles is critical for autonomous vehicles as the lane change may cause conflict between vehicles. Most of the existing lane change prediction models cannot update their model to adapt to the change of road and traffic environment, such as free and congested traffic environments and roads with different levels. Thus, those models cannot be robust. This paper aims to break this limitation and builds a lane change prediction model for surrounding vehicles based on machine learning methods. The lane change prediction model contains a basic model and an adaptive model: the basic model is a long short-term memory (LSTM) based prediction model which reflects the decision-making mode for drivers; the adaptive prediction model embeds an adaptive decision threshold on the basic model, and the threshold updates by Bayesian Inference method on time. We prove the performance of the adaptive model based on the HighD dataset, and the results are inspiring that the model achieves 93.64–97.52% accuracies for target-vehicle in the left adjacent lane and 94.30–98.01% accuracies for target-vehicle in the right adjacent lane. Besides, our proposed model is capable of different driving environments, especially when the traditional LSTM method cannot capture drivers’ lane change decision well. Because of the timeliness and transferability nature of our proposed model, this model could be applied to the Advanced Driver Assistance System (ADAS) and autonomous vehicles in the future to reduce driving perception errors.