Three dimensional reconstruction methods of vehicle driving scenarios for driving assistance

Abstract

Three-dimensional (3D) reconstruction plays a crucial role in the development of realistic driving scenarios for autonomous vehicle simulations. This work investigates and evaluates various methodologies for 3D reconstruction, aiming to enhance the fidelity and applicability of virtual environments in driving simulators. A comprehensive analysis of different reconstruction techniques was conducted, considering point cloud generation, mesh reconstruction, and postprocessing methods. The integration of reconstructed environments into simulation platforms, particularly CARLA, was explored. The study also presents an in-depth investigation into the challenges of integrating Unreal Engine maps with CARLA, identifying the need for multiple files for successful manual integration. Practical tests and evaluations were carried out to assess the accuracy and computational performance of selected methods through evaluations within Cloud Compare. Furthermore, different driving simulators were analyzed to determine their suitability for integrating reconstructed environments. CARLA was selected as the primary simulation tool due to its open-source nature and strong support for autonomous vehicle research. Other simulators, such as SUMO and LGSVL, were also considered but presented limitations regarding 3D scene reconstruction capabilities. The selection process involved evaluating factors like realism, ease of integration, and extensibility, ensuring an optimal environment for autonomous vehicle testing. To support the selection of methodologies, two systematic reviews were conducted following the PRISMA guidelines, utilizing NVivo for qualitative data analysis. This process allowed for the identification and categorization of relevant reconstruction techniques, simulation tools, and integration strategies. The review ensured a structured and comprehensive understanding of the current state of the art, facilitating an informed decision-making process throughout the research. This research contributes by advancing the understanding of 3D reconstruction techniques for driving simulations and providing a structured methodology for integrating virtual environments into autonomous vehicle testing frameworks. Future work includes further optimization of the integration process and the development of automated tools to streamline environment reconstruction for simulation purposes.