Title: Enhancing Data Efficiency in Joint Tracking and Classification through Hybrid Adaptive Kalman Filtering

Abstract: The efficacy of Kalman filtering is often compromised by discrepancies between the assumed and actual models, as well as challenges in tuning noise covariances. While learning-based methods offer solutions to these issues, they frequently depend on supervised training requiring extensive datasets and often fail to yield reliable uncertainty estimates. To overcome these hurdles, this study introduces a self-supervised Hybrid Adaptive Kalman Filter. This approach autonomously learns structured adjustments to system dynamics and process noise covariance directly from measurement data, all while maintaining the filter’s inherent probabilistic framework. Consequently, the innovation likelihood can be calculated and leveraged for model classification through generalized Bayesian inference. Our experimental evaluations, conducted on both real-world and simulated datasets, reveal that the proposed method achieves superior estimation accuracy and statistical consistency, alongside robust classification capabilities in scenarios characterized by both limited and abundant data.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC