Toward Land Vehicle Ego-Velocity Estimation Using Deep Learning and Automotive Radars

This paper presents a deep learning framework for the estimation of land vehicle ego-velocity using Frequency Modulated Continuous Wave (FMCW) automotive radars, addressing the challenges of data sparsity and noise without the need for extrinsic radar calibration. By structuring radar scans into image-based and voxel-based networks, our approach demonstrates robust ego-velocity estimation across multiple sensor configurations and orientations. Experimental results from three distinct datasets—RadarScenes, NavINST, and MSC-RAD4R—validate the framework’s effectiveness, showing superior performance over traditional methods. The models’ adaptability to various sensor specifications and their computational efficiency highlight their potential for real-time applications. We made our implementation open-source at: https://github.com/paaraujo/deep-ego-velocity .