Machine Learning-Based Target Classification for MMW Radar in Autonomous Driving

Millimeter-wave (MMW) radar sensors are considered key components of autonomous vehicles. Because of the performance degeneration of cameras and lidars under inclement weather conditions, robust autonomy must rely on radar sensors to perform target detection and classification. Unlike basic target classifier methods in literature that make use of target velocity, the proposed approach is far more comprehensive and can be applied to targets with zero-Doppler. Depending on the radar type and target range, this paper presents four target classification models based on four different types of radar data: statistical RCS, distributed (time-domain) RCS, range-azimuth angle radar images and 3D radar images. The classification models are implemented by machine learning approaches artificial neural network (ANN) and convolutional neural network (CNN) with a comprehensive simulated dataset. Good classification accuracies are demonstrated, and the proposed model is validated with measured data. Different radar target classification approaches are compared, which clearly reveals the trade-off between classification performance and system complexity. The proposed radar target classification methods can be applied effectively to both static and dynamic targets, at near or far ranges, using traditional or imaging radars, resulting in improved safety for autonomous vehicles in a wide variety of complex environments.