Learning Short-Term Spatial–Temporal Dependency for UAV 2-D Trajectory Forecasting

Trajectory forecasting for unmanned aerial vehicle (UAV) serves a crucial role in the detection and tracking of UAV. However, most existing trajectory sequence forecasting methods fail to excavate and characterize the short-term spatial variation feature in different UAV 2-D maneuvers. In this article, we propose a novel UAV 2-D trajectory forecasting mechanism and spatial-temporal learning aggregator (STLA) to enable accurate and robust short-term UAV 2-D trajectory forecasting in vision-based systems. Specifically, the short-term UAV trajectory correlation method is introduced to separate the azimuth and elevation dimensions, which exposes the spatial variation feature of consecutive UAV trajectory points and eases the burden of requiring vast trajectory data. Besides, STLA is employed to fully explore the spatial variation dependence across the temporal domain by utilizing the proposed spatial feature encoder. Furthermore, to precisely learn and predict the UAV 2-D trajectory under different UAV maneuvers, a spatial-temporal dependency module with a degenerate self-attention mechanism is proposed to adaptively learn influential spatial variation in the feature map. The evaluation of forecasting different short-term UAV 2-D maneuvers demonstrates the effectiveness and robustness of our method, which outperforms the state-of-the-art MLP-based methods as well as mainstream LSTM-based and Transformer-based methods.