New AI Predicts Multi-Vehicle Lane Changes and Trajectories for ADAS

Accurate prediction of surrounding traffic behavior is essential for the safety and effectiveness of advanced driver-assistance systems (ADAS) and autonomous vehicles. Many current methods for predicting lane changes often focus on a single target vehicle, while multi-agent forecasting approaches typically only provide future positions without explicit information about the specific maneuvers each vehicle intends to perform. This limits the depth of understanding required for complex driving scenarios. Researchers propose a dynamic scene graph attention framework, DSiGAT, designed to predict both lane-change intentions and future trajectories for all relevant vehicles in a local traffic scene. The core of the framework is a time-varying interaction graph where vehicles are represented as nodes, and their spatial and kinematic relationships are encoded as explicit edge features. This graph structure allows for capturing evolving inter-vehicle dependencies and subtle pre-maneuver cues. DSiGAT employs temporal graph-attention message passing to process these dependencies, and an intention-guided decoder links predicted maneuvers to corresponding future motions. A scene-level consistency objective further ensures that the multi-vehicle predictions are coherent and compatible. Experiments on real-world datasets (NGSIM I-80, US-101, highD) show significant improvements, with intention prediction accuracies over 90% and trajectory RMSE reductions of up to 52.94% compared to leading baselines. The framework also yields lower inter-agent collision rates and joint displacement errors, indicating more reliable and coherent scene-level forecasts.