Simulation-Based Radar Gesture Recognition Using Domain Adversarial Training

The classification of human hand gestures using UWB-Radar sensors is becoming increasingly important because, unlike optical systems, they can be used regardless of light and weather conditions. However, training deep neural networks to classify such gestures requires a large amount of training data. Collecting this data in real-world scenarios is often challenging, especially when rare or safety-critical events need to be considered. This paper presents a simulation-based approach for generating synthetic radar data based on the open-source Blender framework. In combination with a ray-tracing-based radar simulation, realistic radar measurements based on 3D hand movements are generated. Real-world radar data is used to evaluate the synthetic data. Domain-Adversarial training is used to improve classification and to adapt the real-world and synthetic data. The results demonstrate that the use of adversarial training can lead to improved generalization to real radar data. Thus, this work demonstrates the potential of 3D simulated radar data for training neural networks for hand gesture classification.