Synthetic Radar Data Generation from RGB Images for Robotics and Autonomous Driving Applications

We propose a novel approach to enhance the re-liability of autonomous driving systems through the generation of synthetic Range-Azimuth (RA) heatmaps from RGB images. The method involves the generation of synthetic Range-Azimuth (RA) heatmaps from RGB images or relative depth images, facilitating the extraction of essential depth and azimuth angle information through supervised learning. This effectively bridges the gap between visual and radar data modalities, offering a more comprehensive and context-aware approach to mitigate noise and enhance the quality of radar-based perception. In summary, our method leverages a GAN-based architecture with tailored loss functions. These tailored loss functions play a pivotal role in refining the accuracy and realism of the generated heatmaps, ensuring a more robust and reliable perception system for autonomous driving. Our approach also highlights the potential for data augmentation, enabling synthetic data generation for non-existent or safety-critical scenarios. Additionally, our method serves as a secondary system, providing redundancy in situations where radar systems may fail, like degraded or cluttered reflections, enhancing the overall safety and reliability of autonomous driving systems. We demonstrate the results of the proposed methodology on benchmark datasets (CRUW) in comparison with SOTA methods using appropriate quantitative and qualitative metrics.