Attention-Driven Lane Change Trajectory Prediction With Traffic Context in Urban Environments

Lane-changing is one of the most fundamental driving behaviors, yet it is also one of the most hazardous maneuvers, as interactions with adjacent lanes critically impact vehicle safety. Accurate prediction of lane-change trajectories is essential for enhancing road safety and the effectiveness of autonomous driving systems. This study proposes a model that leverages traffic context from drone footage to predict lane-change trajectories, incorporating interactions between surrounding vehicles. The proposed model analyzes drone footage collected from natural urban driving environments using computer vision techniques to extract traffic context and employs a Transformer-LSTM architecture to predict lane-change trajectories. Beyond trajectory prediction, this study enhances model interpretability by analyzing the key factors influencing lane-change predictions. Critical moments in the prediction process can be visually identified using attention heatmaps, and the model’s focus on essential input information, such as the approach of nearby vehicles or significant interactions, can be observed. This approach offers meaningful insights into the model’s predictive focus. The feasibility of the proposed model was validated through the implementation of a prototype and its application to real traffic data, confirming its effectiveness in accurately predicting lane-change trajectories across diverse traffic scenarios.