Scene Understanding With Automotive Radar

Extracting semantic information solely from automotive radar data is a relatively new topic in the radar community. We present a complete pipeline to obtain semantic information for each target measured by a network of radar sensors. Static and dynamic objects are treated in two separate branches: In the first branch, a convolutional neural network performs semantic segmentation on radar grid maps of the static environment. In the second branch, a novel neural network architecture is used for recurrent instance segmentation on radar point clouds of moving objects. The class probabilities assigned to each cell in the grid map are mapped back to the radar targets in this spatial region so that in a merging step the results from the two classifiers can be combined into one point cloud. In addition to a novel network structure for recurrent instance segmentation of point clouds, we present a new set of layers for radar grid maps which are beneficial for semantic segmentation tasks and we also develop a weighting scheme for the network's loss function to account for the data integration process in grid maps. We evaluate our approaches on large data sets and we display that they outperform previously proposed methods.