AdaFusionNet: Adaptive Trajectory Prediction of Group Two-wheelers in Interactive Conflict Scenarios for Autonomous Vehicles

The accurate prediction of two-wheeler motion trajectories is essential for ensuring the safety of autonomous vehicles (AVs), particularly at crowded intersections in China. Existing trajectory prediction approaches face limitations in representing the diverse motion patterns, dynamic constraints, and group-level behavioral characteristics of two-wheeler riders. This paper proposes an adaptive trajectory prediction framework for two-wheelers by integrating the revised social force model (RSFM) and a spatiotemporal Transformer (STTransformer) module. The RSFM models group-level interactions among two-wheelers, capturing self-organization, leader-follower dynamics, and herding behaviors. The STTransformer extracts temporal dependencies and spatial interaction patterns using multi-head spatiotemporal attention mechanisms to capture group motion patterns. A gated recurrent unit (GRU) serves as the fusion module, recurrently integrating the trajectory predictions generated by the RSFM and STTransformer together with a local traffic-scene complexity representation constructed from the numbers and mean speeds of nearby pedestrians and vehicles. Experimental results on the Signalized INtersection Dataset (SIND) show that AdaFusionNet achieves an accuracy improvement of over 11% at the 3 s prediction horizon compared with the selected baselines. The error distributions further indicate fewer large displacement errors, while the leave-one-intersection-out evaluation provides additional evidence of cross-intersection transferability, with AdaFusionNet maintaining competitive prediction performance when applied to intersections excluded from training.