Radar Image-Based Pedestrian Detection Scheme in Mobile Radar Platform for Millimeter-Wave Automotive Radar

An accurate static pedestrian detection scheme for mobile radar platforms is presented for vehicle-mounted millimeter-wave (MMW) radar sensing scenarios. While there are several studies focusing on the target identification in line-of-sight (LOS) or non-LOS (NLOS) scenarios, exploiting a unique signature of pedestrians in complex-valued radar echoes, most of them assume a static radar platform. In an actual sensing scenario, a vehicle-mounted radar platform has a natural motion, which significantly affects the classification accuracy using the scattered data, including through the Doppler shifts or other phase variation factors, such as vibration, and this issue becomes more crucial in classifying static pedestrians and other clutter, such as a metallic pole. Thus, this study focuses on target identification for a radar platform moving along a straight line, where phase variations due to Doppler shifts are compensated in signal postprocessing. As the wavelength of a 79-GHz MMW radar is much shorter than the motion within a single pulse repetition interval (PRI), a signature caused by respiration or small body motion is difficult to extract from the reflection data obtained by a mobile radar platform. To address this issue, this article introduces a radar image-based feature extraction scheme in which the temporal variations of radar images are utilized as feature vectors for post-machine learning classification [e.g., support vector machine (SVM) and long short-term memory (LSTM)]. Experiments conducted using a 79-GHz MMW frequency-modulated continuous wave (FMCW) radar demonstrate that the proposed scheme achieves substantially higher accuracy in distinguishing between artificial objects and static pedestrians compared with conventional approaches.