Low-speed UAV trajectory prediction under short-term data loss: a preliminary feasibility study for vehicle-mounted electro-optical tracking.

Short-term data loss occurs when mobile platforms track unmanned aerial vehicle (UAV) targets, severely impairing the continuity and stability of electro-optical (EO) tracking. A UAV trajectory prediction method integrating flight-state recognition and an attention mechanism is proposed for short-term data-loss scenarios. Based on UAV trajectory data collected by EO tracking systems, an improved support vector machine (SVM) is adopted to identify UAV flight states. A joint architecture combining convolutional feature extraction units and long short-term memory (LSTM) networks is established to mine temporal trajectory features, with an adaptive attention mechanism introduced to strengthen key feature learning. Experimental results show that the improved SVM achieves a recognition accuracy of no less than 90% for all basic flight states. Under 2 s of observation loss, the proposed method outperforms LSTM and Interacting Multiple Model Extended Kalman Filter (IMM-EKF) in root mean square error (RMSE) for turning flight prediction; its accuracy for non-turning states matches Temporal Convolutional Network (TCN) and IMM-EKF. Meanwhile, it features fewer parameters and lower computational cost than lightweight TCN. Ablation experiments verify that flight-state recognition and state-guided attention effectively boost trajectory prediction precision. Extended tests covering multi-gradient velocities, sampling frequencies, prediction horizons and data-loss positions demonstrate the favorable overall performance of the proposed method and define its application scope. In conclusion, the algorithm preliminarily proves its feasibility in addressing short-term data loss for vehicle-mounted EO tracking of low-speed mildly maneuverable UAVs.