High Fidelity Physics Simulation-Based Convolutional Neural Network for Automotive Radar Target Classification Using Micro-Doppler

Detection and classification of vulnerable road users (VRUs) such as pedestrians cyclists is a key requirement for the realization fully autonomous vehicles. Radar-based VRUs can be achieved by exploiting differences in micro-Doppler signatures associated with VRUs. Specifically, machine learning (ML) algorithms trained to classify using spectral content radar signals. The performance these models depends on quality quantity data used during training process. Currently, collection typically done through measurements or low fidelity physics, primitive-based simulations. feasibility carrying out collect limited vast amounts required practicality issues when like animals. In this paper, we present computationally efficient, high physics-based simulation workflow that obtain large spectrograms from simulations are conducted full-scale VRU 77 GHz, frequency-modulated continuous-wave (FMCW) sensor model. Here, 4 targets; car, pedestrian, cyclist dog at different speeds angles-of-arrival. This then train 5-layer convolutional neural network (CNN) achieves nearly 100% accuracy after 5 epochs. Studies investigate impact size, velocity observation time window size CNN. Results study demonstrate how an 95% realized obtained over 0.2 s window.