Heterogeneity-aware risk-adaptive trajectory planning for UAVs in air-ground multi-agent interaction

Autonomous UAVs sharing low-altitude airspace with ground traffic must continuously interpret the heterogeneous and dynamically varying behaviors of ground agents, a challenge that conventional risk models, largely designed for static obstacles or aircraft-to-aircraft conflicts, are not equipped to address. This study formulates the air-ground interaction problem as a heterogeneous multi-agent risk perception task and proposes a three-stage computational intelligence framework for risk-adaptive UAV trajectory planning. First, a GMM-LightGBM model performs online identification of ground agent driving styles, which is coupled with a Full Velocity Difference (FVD) model to generate personalized agent trajectory predictions. Second, a fuzzy inference system fuses macroscopic traffic context, agent behavioral features, and air-ground geometric relationships into an interpretable, real-time risk score. Third, an MS-DGST-VNet model predicts UAV maneuver parameters, and a risk-adaptive cost function dynamically reweights safety, efficiency, and altitude constraints to generate adaptive trajectories. The proposed driving style classifier achieves 0.96 identification accuracy, and the resulting risk score correlates strongly with three-dimensional air-ground time-to-collision, supporting the framework's validity. Case studies across representative risk scenarios further demonstrate the framework's adaptive decision-making capability. By combining machine learning-based behavior identification, fuzzy reasoning, and learning-based trajectory generation, this work contributes a computational intelligence methodology for autonomous UAV decision-making in heterogeneous multi-agent environments.