A learning-based interacting multiple model filter for trajectory prediction of small multirotor drones considering differential sequences

Reliable trajectory prediction is a cornerstone component for supporting various higher-level applications in uncrewed traffic management (UTM). The high flexibility of multirotor drones presents significant challenges to achieving accurate trajectory prediction. In scenarios involving the tracking of non-cooperative targets, insufficient information hinders the effectiveness of machine learning-based methods for forecasting future trajectories. To address these limitations, we conceptualise the problem as a multivariate time series prediction task and propose an innovative integrated flight trajectory prediction approach, including two distinct modules. The initial module employs three modified Transformer models to forecast future sequences of position, velocity, and acceleration in parallel for capturing diverse flight patterns. The subsequent module features a learning-based interacting multiple model (IMM) filter designed to fuse the three predicted sequences, regarded as pseudo measurements, by adaptively learning time-invariant transition probabilities. We conducted two experiments using 17 multirotor drone trajectory datasets collected from industrial and academic applications. The results demonstrate: i) integrated position sequence and discrete velocity approach can significantly enhance trajectory prediction accuracy; ii) the modified Transformer architecture shows substantial potential compared to baselines; iii) the learning-based IMM method yields superior prediction results on 15 new trajectory datasets, effectively simulating scenarios of managing unidentified drones in real-world contexts.