Abstract: To address the high false-alarm rate in collision risk assessment for non-cooperative unmanned aerial vehicles (UAVs) within airport control areas, a collision risk assessment method based on long short-term memory (LSTM) and Bayesian filtering is proposed. First, a UAV motion model is established, from which temporal features including radial approach rate, heading stability index, and normalized approach probability are extracted for three typical intentional trajectories: accidental intrusion, malicious intrusion, and active evasion. Second, an integrated LSTM–Bayesian filtering intention recognition model is constructed to deeply capture temporal dependencies in UAV trajectories, suppress noise interference via Bayesian filtering, and output a target intention probability distribution matrix. Finally, airspace grid-based modeling is adopted, where dynamically adjusted weights are designed for collision probability calculation based on intention recognition results, enabling intention-driven collision risk assessment. Simulation results demonstrate that the proposed model achieves an intention recognition accuracy of over 97% and a collision risk assessment accuracy of 84.8%, representing improvements of 14.35%, 8.57%, and 5.86% compared with the conventional static Alert Zone method, Worst-case Intention (WCI) method, and Direction Retention (DR) method, respectively. The method also significantly reduces false alarms while maintaining a low missed-alarm rate, providing technical support for refined risk management of non-cooperative UAVs at airports.

Key words: Unmanned aerial vehicle (UAV), LSTM, Bayesian filtering, Collision risk