Abstract:
To address the challenges of non-uniform sampling, clutter interference, data packet loss, and multi-scale feature distribution of intentions observed from non-cooperative UAVs in low-altitude complex airspace, this paper proposes a low-altitude UAV intention prediction method with dual-stream Transformer fused with spatio-temporal physical priors (STPDS-Transformer). Continuous spatio-temporal relative position encoding is constructed, which introduces physical time intervals and spatial Euclidean distances into sequence modeling to achieve physically consistent representation of non-uniform trajectory sequences. A physical prior embedded self-attention mechanism is designed to softly filter abnormal data through physical limit constraints of velocity and acceleration. Furthermore, a local-global dual-stream attention encoder combined with an adaptive gating mechanism is developed to collaboratively model the long-term behavioral trends and short-term maneuver characteristics of UAVs, so as to dynamically fuse multi-scale features. Experimental results show that the intention prediction accuracy of the proposed STPDS-Transformer model reaches 97.3%, outperforming multiple mainstream models. Under strong interference conditions with Gaussian white noise (
σ=0.07) and a 40% packet loss rate, the prediction accuracy remains 94.3% and 93.8%, respectively, demonstrating excellent anti-noise robustness. The deep model integrated with spatio-temporal physical priors can effectively improve the UAV intention prediction capability under noisy and non-uniform sequences, providing critical technical support for forward decision-making in anti-UAV command and control systems.