Deep Complex-Valued Network for Ego-Velocity Estimation with Millimeter-Wave Radar

The estimation of the precise ego-velocity of a sensor is a crucial for advanced applications such as object tracking, localization, and autonomous driving. We have observed that the phase of complex-valued radar data contains abundant information, and that a delicate network design is required to fully incorporate the information. In this study, we propose an end-to-end neural network based method that is not affected by prior processes. Our approach is not subjected to the unambiguous maximum velocity problem of the radar. Furthermore, it is not necessary to accurately measure the mounting position or the viewing angle of the sensor. Experimental results based on extensive real-world dataset verify the advantages of the proposed method.