Enhanced traffic safety and efficiency of an accelerated LC decision via DNN-APF technique

Abstract

The decision-making module that operates in real-time is extensively valued for promoting traffic flow stability and throughput during autonomous driving. However, it may become problematic for autonomous vehicles to conduct Lane-changing (LC) decisions for acceleration in challenging urban scenarios, which poses certain risks to driving safety. This study develops a Deep Neural Network (DNN) and Artificial Potential Field (APF) model with appropriate feature selection for safe and efficient accelerating LC decisions. Firstly, a two-lane driving scenario with 5+1 vehicles is established. In the course of accelerating LC, 7 analysis methods are exploited to identify areas of vehicle concern with consideration of 30 different features. Secondly, the methods of DNN, Support Vector Machine (SVM), and K-nearest Neighbor (KNN) that commonly used in the prediction of decision classification are analysed and compared, based on which the DNN model is selected and obtained with an optimal decision effect and potential improvement extent. Furthermore, taking into account the security considerations of LC decision and the possible improvement of DNN model, the risk monitoring method of APF with extensive application background is introduced into the DNN results. Finally, the proposed DNN-APF approach is compared with pure DNN in terms of fitting degree, safety factor, and velocity enhancement efficiency. The results indicate that under the premise of reasonable feature extraction, the DNN-APF approach achieves the accelerated LC behaviour decision with high accuracy and safety performance in comparison with SVM, KNN, and the DNN without security considerations.