Polarimetric Information Representation for Radar based Road User Detection with Deep Learning

Automated driving requires a complete and robust perception of the car’s environment. A crucial part in the perception module is the reliable detection and classification of road users encountered in road traffic. With its robustness to adverse weather conditions and instantaneous radial velocity estimation capabilities the radar is one of the main sensors employed to ensure robust perception. While a conventional automotive radar measures the range, azimuth angle, received energy and radial velocity of an object, a newly available automotive polarimetric radar can provide additional information about the electromagnetic wave scattering properties, like shapes and orientations of geometries on that object.In this work, the impact of additional polarimetric features in radar point clouds on a deep learning model’s semantic and instance segmentation performance is evaluated. In experiments on a large real-world data set, multiple polarimetric information representations are compared and a significant improvement in semantic and instance segmentation performance with polarimetric information is shown.