Towards Pedestrian Detection in Radar Point Clouds with Pointnets

Pedestrian recognition with radar is a very sensitive task that can enable the deployment of several advanced driver assistance systems. Despite neural networks represent a powerful and effective tool for solving such task, their abilities as well as their limits have been scarcely addressed in the literature. In this paper, we investigate how point-wise processing architectures use the radar features to perform the task of pedestrian detection. We study the behavior of four different techniques to a perturbation of the radar point clouds. We show that radar Doppler represents the most critical feature for the detection of pedestrians. However, we find out that the context plays an important role. Finally, we prove that PointNet++ learns to use all the radar features in a proper, meaningful way, thus achieving superior performance to PointNet and Random Forest.