Transformer-based Motion Model for Robust Target Tracking under Intermittent and Noisy Measurements

Abstract

Target tracking under intermittent measurements is a fundamental challenge in autonomous systems. Traditional methods, including Kalman filters and deep learning-based models, often struggle when faced with sparse observations and high measurement noise. In this work, we present multiple transformer-based motion models desgined to learn target dynamics from noisy sensor measurements and occluded portions. By leveraging self-attention mechanisms, these models effectively capture temporal dependencies and infer motion trajectories under uncertainty. We evaluate various architectural formulations, including time-encoded position inputs to better handle occlusions. These learned motion models are then integrated with a particle filter for target estimation and with an information-driven planner to guide the tracking agent. Since the models influence the guidance logic through their predictions, we assess their effectiveness based on overall target tracking performance. Extensive simulation and hardware experiments demonstrate that our approach improves tracking accuracy and robustness compared to existing methods.