Point Cloud Features-Based Random Forest for Road User Classification via Millimeter Wave Radar

Abstract

Vehicle-mounted millimeter-wave(mmWave) radar sensors can meet the requirements of vehicles for the all-day and all-weather sensing of the surrounding environment. The mmWave radar can produce one to dozens of target points for ordinary road users, and such sparse point cloud data lead to difficulties in classification. When there are fewer or even only one point in the target-level point cloud, which usually occurs when the object is tiny or far from the radar, the classification accuracy of this type of object will become very poor. To improve the classification accuracy of radar-based road users, we propose a new feature extraction and hierarchical classification algorithm for mmWave radar point clouds. We extracted 6-dimensional and 125-dimensional features for single-point and multi-point objects based on the range, azimuth, velocity, and radar cross-section (RCS) of point clouds measured by the mmWave radar to characterize the motion, attitude, position, and material of the road user. To validate the effectiveness of the proposed method, we evaluate our method on the publicly available dataset. Compared with the traditional algorithm, the overall classification accuracy of the proposed method is higher. In addition, this divide-and-conquer strategy also improves the classification performance of single-point objects.