4DEgo: ego-velocity estimation from high-resolution radar data

Automotive radars allow for perception of the environment in adverse visibility and weather conditions. New high-resolution sensors have demonstrated potential for tasks beyond obstacle detection and velocity adjustment, such as mapping or target tracking. This paper proposes an end-to-end method for ego-velocity estimation based on radar scan registration. Our architecture includes a 3D convolution over all three channels of the heatmap, capturing features associated with motion, and an attention mechanism for selecting significant features for regression. To the best of our knowledge, this is the first work utilizing the full 3D radar heatmap for ego-velocity estimation. We verify the efficacy of our approach using the publicly available ColoRadar dataset and study the effect of architectural choices and distributional shifts on performance.