People Detection Using Radar

This thesis aims to research the applicability of deep learning methods on point clouds generated by millimeter-wave radars, as a solution for people detection, and 3D scene understanding in general. Radar is a system that uses radio waves to determine the distance, azimuth, and velocity of surrounding objects. For each point of detection, Cartesian coordinates can be calculated, to produce a set of points in 3D space called a point cloud. Deep neural network architectures designed to operate on sparse 3D point clouds can be trained for point-wise segmentation, object detection, classification, and tracking. This can be used to greatly advance the 3D scene understanding by machines. A model based on the state-of-the-art methods for object detection and classification on sparse point clouds was trained as a part of this thesis, for the purpose of people detection. To showcase the robustness of the trained model and the straightforwardness of its applicability to solve prominent real-world tasks of scene understanding, a people counting application was developed. The employed methods were thoroughly evaluated on a dataset created as a part of this thesis, consisting of over 19,500 labels on 3D radar point clouds.

millimeter-wave radar, radar point cloud, 3D point cloud annotation, people detection, people counting, 3D scene understanding, PointNet++, PointRCNN