PointNeXt Adapter Model for Low Altitude Route Conflict Detection in UAV Trajectory Intent Streams

Low-altitude UAV routes are increasingly organized as time-dependent corridors, but conflict detection remains vulnerable when vehicles change speed, altitude, or maneuver intention before a formal route update is broadcast. This paper proposes a PointNeXt Adapter model for low-altitude route conflict detection in UAV trajectory intent streams. Each traffic scene is encoded as a spatiotemporal point set containing position, velocity, altitude layer, maneuver probability, and route- permission attributes. The proposed model couples a PointNeXt geometric backbone with lightweight adapter blocks that modulate intermediate point features by temporal intent embeddings, rather than simply appending intent variables at the final classifier. A mixed-source evaluation corpus containing 62,050 trajectories and 9,928 conflict windows was constructed from urban corridor replay, vertiport approach simulation, inspection logs, emergency rerouting scenarios, and weather-perturbed augmentation. Compared with PointNet++, DGCNN, Point Transformer, and recurrent or graph-based baselines, the proposed model achieved 95.8% accuracy, 94.6% F1 score, 0.963 AUROC, and 37 ms median inference latency per 128-trajectory scene. The results indicate that intent-conditioned point adaptation can improve early warning reliability under controlled mixed-source evaluation while preserving a computational profile suitable for engineering traffic-monitoring studies.