Towards Land Vehicle Forward Velocity Estimation using Deep Learning and Onboard Radars

Radars have been increasingly implemented in modern vehicles. Their presence opens opportunities for developing new positioning and collision avoidance solutions, to cite a few. However, radar scans are noisy and sparse, which challenges the robustness of the developed solutions. In this paper, we propose a novel method to structure radar scans to create point descriptors. The structured data is used in convolutional neural networks that explore orthogonal views of the radar point cloud. Experimental results demonstrate that the model performs well in estimating the forward velocity of the vehicle using only the radar scans, providing estimations at a higher data rate than odometers available in the vehicles.