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