Classification of Automotive Targets Using Inverse Synthetic Aperture Radar Images

Traditionally, point cloud representations or Doppler spectrograms have been generated from short-range automotive radars for dynamic object detection and classification. In this work, we propose using inverse synthetic aperture radar (ISAR) images obtained from range compensated turning targets for the classification of different types of vehicles. We experimentally demonstrate that ISAR images of automotive targets provide rich features such as the dimensions, trajectory, and the number of wheels of the vehicles for classification. Additionally, we present a simulation framework for generating large volumes of realistic ISAR images of automotive targets at millimeter-wave frequencies for training classifiers. The model incorporates radar scattering phenomenology of commonly found vehicles along with range-Doppler-based clutter and receiver noise. The model is experimentally validated with measurement data gathered from an automotive radar. The images from the simulation database are subsequently classified using traditional machine learning techniques and deep neural networks based on transfer learning. We show that the ISAR images offer a classification accuracy above 90% and are robust to both noise and clutter.