Human Motion Training Data Generation for Radar Based Deep Learning Applications

Abstract

Abstract-Radar sensors are utilized for detection and classification purposes in various applications. In order to use deep learning techniques, lots of training data are required. Accordingly, lots of measurements and labelling tasks are then needed. For the purpose of pre-training or examining first ideas before bringing them into reality, synthetic radar data are of great help. In this paper, a workflow for automatically generating radar data of human gestures is presented, starting with creating the desired animations until synthesizing radar data and getting the final required dataset. The dataset could then be used for training deep learning models. A classification scenario applying this workflow is also introduced.

Automotive Radar — From First Efforts to Future Systems

“…In [138], radar measurements are enriched with GNSS ground-truth for ML-based VRU recognition. Furthermore, synthesizing VRU radar responses with radar simulators has been presented with the motion ground-truth obtained from kinematic models [139], animations [140], [141], or Kinect data [142], [143].…”

Learning on Multistatic Simulation Data for Radar-Based Automotive Gesture Recognition

“…Approaches can be distinguished with respect to the underlying human target model, which can be based on a set of char- [12] GAN - [13] GAN - [14] Simulation MoCap [16], [17] Simulation Kinect [15], [21] Simulation Kinect [18], [19] Simulation Blender [20] Simulation Blender [23] Domain Transfer Video (Mono) This work Simulation Video (Stereo) 1 DA: direct augmentation, 2 PT: pre-training, 3 CD: cross-domain training acteristic skeletal keypoints, whose positions are derived from motion capture data [14] or Kinect sensors [15], [16], [17], and models that derive scattering centers from 3D body models generated e.g. with help of computer graphics software [18], [19], [20]. Starting from these simulation approaches, some papers recently explored the potential of simulated radar data for augmentation.…”

Data Augmentation in Time and Doppler Frequency Domain for Radar-based Gesture Recognition

“…For that reason, researchers in radar-based gesture recognition have proposed more sophisticated methods to synthesize additional samples. Mainly, these methods are divided into ML-based techniques, such as the use of generative adversarial networks (GANs) [6], [7] or variational autoencoders [7], and simulation-based approaches [8], [9], [10]. However, these methods come with a high complexity and require a lot of effort in implementation and training, while their benefit is strongly problem-dependent and not guaranteed.…”