Radar target simulations are vital to automotive research as they make it possible to reproduce real scenarios with reasonable accuracy. However, this process comes with high computational costs and limitations. The generated targets are ideal and thus far from reality. Recent studies have led to an alternative to generated synthetic radar data: Generative Adversarial Networks (GAN), a neural network architecture designed to generate realistic data to synthesize images, texts, and more, which is being studied for radar data applications. Until now, only a few studies have explored the approach, especially when considering automotive radars, but showing promising results in its applications. This work proposes a Conditional GAN (CGAN) that synthesizes FMCW radar Range-Doppler targets with Micro-Doppler given a selected object input. The results show that the generated samples are realistic enough to be classified with an accuracy of 82% in a pre-trained classifier, proving that the synthetic data seems to be similar to the real ones but not representing ideal results, as simulation targets do, thus fulfilling an important gap of knowledge for simulation purposes.
A novel Conditional Generative Adversarial Networks for Automotive Radar Range-Doppler Targets Synthetic Generation
2023-09-24
6806141 byte
Conference paper
Electronic Resource
English
REALISTIC ULTRASONIC ENVIRONMENT SIMULATION USING CONDITIONAL GENERATIVE ADVERSARIAL NETWORKS
British Library Conference Proceedings | 2019
|