Lane-Change Intention Recognition in Roundabouts Based on Explainable Deep Learning

As complex traffic nodes, roundabouts rely on drivers to independently adjust their driving direction and speed to maintain traffic flow. The frequent lane changes and multidirectional interactions significantly increase the risk of accidents and congestion. However, lane-change intention recognition in roundabouts remains underexplored, raising collision risks for autonomous vehicles. This study proposes a novel framework for recognizing lane-change intentions in roundabouts, offering a new modeling perspective to identify key factors influencing these intentions. A deep learning model integrating convolutional neural networks and bidirectional long short-term memory networks is presented for detecting lane-change intentions. To improve feature capture, a multi-head attention mechanism is added. The SHapley Additive exPlanations (SHAP) methodology is applied to interpret model predictions and identify critical factors influencing lane-change decisions. The experimental evaluation demonstrates that the developed framework achieves superior performance metrics compared with existing baseline approaches with regard to lane-change intentions, recognition accuracy, and reliability. Through the analysis of the SHAP model, it is found that the key features affecting the top ranking of the three types of intentions are all the vehicle’s own driving state. Reintegrating the key nine features into the recognition model, the results show that the recognition accuracy remains almost unchanged but can reduce the iteration time of the model.