Transformer-Based State Estimation for Tracking: Maneuvering Target and Multi-Target Capabilities

An exploration of a Transformer's behavior is proposed in the context of object tracking, with a particular emphasis on maneuvering targets. We compare the performance of our Transformer-based approach against established tracking methods, including the Kalman filter, Extended Kalman Filter, and Unscented Kalman Filter. The first experiment examines its performance in single-target maneuvering scenarios, revealing heightened reactivity without compromising accuracy during abrupt maneuvers and straight-line trajectory. In a second experiment, we showcase the method's multi-target associative capability. By leveraging the attention mechanisms inherent in Transformers, we capitalize on both spatial and temporal dependencies for accurate tracking. We introduce specific training strategies and modifications to the original Transformer architecture. Our proposed method underscores the Transformer's potential in maneuvering and multi-target scenarios, providing valuable insights into its efficacy for dynamic object motion estimation. We discuss the comparative results, highlighting performance gains over traditional approaches and addressing potential limitations.