Digital Shadow & Virtual Worlds for VR Driving in CARLA

Evaluating rare but critical driving situations, such as hazardous road events or complex traffic interactions, is challenging yet essential. Physical tests for every possible scenario are expensive, difficult to manage reliably, and potentially dangerous. Consequently, virtual simulations have become fundamental resources for research in autonomous driving, traffic modeling, and safety analysis. Researchers take advantage of these environments to replicate diverse driving scenarios. Sensor responses allow them to analyze vehicle behavior in a controlled and highly adaptable setting. Among the virtual simulation tools, CARLA (Car Learning to Act) stands out as an open-source driving simulator widely used for autonomous vehicle research and other multidisciplinary studies. Thanks to its realistic physics engine, customizable sensors and high-quality rendering, CARLA is also used in areas such as computer vision for neural network training, robotics through integration with the Robot Operating System (ROS), human-machine interaction to analyze human behavior in autonomous driving scenarios and urban planning. CARLA uses physically based rendering (PBR) techniques in its rendering pipeline to achieve photorealistic visuals while still maintaining real-time performance. Achieving the right balance between realism and efficiency becomes essential to create meaningful and immersive simulations. This thesis explores methodologies for generating maps that are compatible with the CARLA environment, focusing on two key objectives: achieving a precise reconstruction of real-world environments through geospatial data and enhancing visual realism to improve simulation immersion. The first approach focuses on 'Digital Shadows', aiming to achieve a high-fidelity representation of real-world environments by leveraging geospatial data sources such as LiDAR scans and OpenStreetMap (OSM). These datasets provide precise topographic and infrastructural details, ensuring that the generated maps closely match real-world road networks and layouts. The second approach prioritizes visual realism and immersion, utilizing Unreal Engine assets to enhance graphical quality and the overall simulation experience. More generally, achieving high spatial accuracy is based on geospatial data acquisition techniques, such as LiDAR scanning, to capture real-world structures with precision. On the other hand, enhancing the visual fidelity of virtual assets requires procedural generation and optimization methods to achieve both realism and performance.