Exploring Synthetic Radar Data and Deep Learning for Road User Classification in Autonomous Vehicles

Diogo Wachtel, Sinan Tasabat, Thomas Rothmeier, Letícia Cristófoli Duarte, Werner Huber

Abstract: This work investigates enhancing road safety and efficiency in autonomous vehicles through radar sensors and deep neural networks with synthetic data. It focuses on machine perception using image-based algorithms and Deep Convolutional Neural Networks (DCNNs) to detect and classify road users accurately with micro-Doppler images. It explores the effectiveness of ray tracing technology and the HFSS SBR+ tool from ANSYS for realistic radar simulations in dynamic environments. The research demonstrates the capability of these technologies in classifying pedestrians, cyclists, and vehicles using both simulated and real-world data. This work contributes to autonomous vehicle technology by showcasing advanced radar simulations and deep learning methods for improved road safety and vehicle efficiency. Augmenting the dataset with synthetic data significantly enhances car classification accuracy by nearly 20%, with minimal impact on the classification of other objects. Furthermore, the inclusion of synthetic data has shown to be highly beneficial, improving the model's overall accuracy by at least 10% across all training datasets. The trained CNN model with the combined dataset, synthetically generated data, and real data demonstrated a robust solution, achieving an overall accuracy of approximately 96% with the distinction that this study introduced the synthetic generation of objects.