Semantic Segmentation on Automotive Radar Maps

As radar sensors can measure an object's range and velocity with a high degree of precision, moving objects can be successfully classified, as well. Classifying stationary objects still needs a lot of research, however. In this paper, we use popular semantic segmentation networks in order to classify the vehicle's immediate infrastructure. To this end, a full 3D measurement is performed with a test vehicle equipped with four high resolution corner radar sensors. A preprocessed point cloud is transformed into various radar maps for input to a neural network. Simulations as well as real-world measurements show an overall intersection over union of 84 and 77%, respectively, as well as an overall accuracy of 95 and 90%, respectively, being a new benchmark for this young research field.