Research on Lane Change Decision Models for Autonomous Vehicles Based on Deep Learning

Automated lane changing is an important scenario in intelligent vehicle driving. Incorrect lane changes are often a significant cause of traffic congestion and accidents. In order to improve the generality and safety of the lane change method, this study addresses issues such as the low accuracy of proactive lane change decision-making in autonomous vehicles, the strong dependence of lane change trajectory planning algorithms on road shapes, and insufficient safety. In this paper, a Gated Recurrent Unit (GRU) lane change decision model was proposed based on the powerful extraction and fitting capabilities of deep learning algorithms for nonlinear feature extraction. And the test results of the GRU, Long Short-Term Memory (LSTM), Support Vector Machine (SVM), Random Forest (RF), and BP neural networks lane change decision model were compared. The results indicate that the GRU-based lane change decision model exhibits excellent performance in terms of accuracy and Receiver Operating Characteristic (ROC) metrics, achieving an accuracy of 96.85 % on the test set. The LSTM-based lane change decision model performs similarly to the GRU-based model, with an accuracy of 96.78 % on the test set, both surpassing the existing algorithms used for comparison.