Transformer-based autonomous driving vehicle ramp merging decision algorithm in multi-vehicle interaction scenarios

To address traffic problems during vehicle ramp merging, this paper studies a decision-making algorithm for autonomous vehicles merging on ramps in a multi-vehicle interaction scenario. A decision-making algorithm is built based on a Transformer encoder to improve the feature processing capability of vehicle trajectories. An LSTM-Transformer bi-branch trajectory prediction model is constructed, and longitudinal and lateral motion features are processed separately through a feature-decoupling mechanism to achieve collaborative prediction of multimodal trajectories. Based on this, a planning algorithm integrating safety constraints and multi-objective cost functions is designed, and a decision-making model considering lane-change conflict risks is established. Experimental results show that compared with other comparative algorithms, this model is safer and more efficient in decision making. In comparison with all other models, the model in this study achieved the highest True Positive Rate of 95.201%, True Negative Rate of 93.192%, and Accuracy of 96.310%. Considering the vehicle ramp merging decision-making performance under different traffic densities, the model effectively facilitated vehicle merging under all three traffic densities, with merging success rates of 100%, 98%, and 96%, respectively, demonstrating its effectiveness in ramp merging.