ST-Bayes: Robust Radar Odometry in Dynamic Environments via Student’s T-Distribution Temporal Bayesian Learning

Abstract

The problem of ego-motion estimation for automotive multiple-input multiple-output (MIMO) radar is studied in this article. Conventional methods for ego-motion estimation, such as random sample consensus (RANSAC) and its variants, fail when facing more dynamic driving scenarios, i.e., those where many and diverse moving targets are present in the field of view. The proposed method utilizes for the Student’s t-distribution Bayesian learning to distinguish between moving objects (which would count as outliers in the ego-motion estimation process) and static objects (which are instead inliers for ego-motion estimation). The method integrates temporal information across multiple frames within a Bayesian framework to enhance the estimation accuracy. The proposed method has been verified through numerical simulations and real-world experimental data. The accuracy of ego-motion estimation is improved by an average of over 50% in the real-world dataset compared to the RANSAC method.