Lane changing decision prediction method for vehicles at signalized intersections based on spatiotemporal graph convolutional neural networks

ObjectiveTo accurately predict the lane changing decision behaviors of surrounding vehicles at signalized intersection areas, so as to enhance the driving safety and ride comfort of autonomous vehicles, the lane changing decision prediction method based on spatiotemporal graph convolutional neural networks was proposed.MethodFirst, the spatial interaction relations, between lane changing vehicles and their surrounding vehicles at intersection area, were modeled by using graph theory. Second, GCN was then employed to extract the spatial features of multi-vehicle interactions. Subsequently, LSTM neural network was utilized to capture the temporal evolution patterns of the spatial relations. Finally, the classification result of lane changing decisions was output through the fully connected network combined with Softmax function. A total of 2 875 valid samples extracted from unmanned aerial vehicle trajectory data were used to train and validate the proposed model.ResultThe model achieves optimal prediction performance with a feature extraction time window of 4 s and a prediction time of ―2 s, attaining an overall accuracy of 91.3%. It represents an average improvement of 4.3% compared with four baseline models, i.e., LSTM, GCN, SVM and XGBoost. The ablation tests further confirm that the spatiotemporal information fusion modeling outperforms single-dimensional feature modeling in terms of prediction accuracy and balance among different lane changing types.ConclusionThe findings provide an effective method for multi-vehicle interaction modeling and lane changing behavior prediction in connected environments, contributing to improving the safety decision-making capabilities of advanced driver assistance systems and autonomous driving systems.