Data-Driven 3D Digital Twin Generation for Transportation Applications

Abstract: Digital twins are gaining more tractions in transportation applications to replicate real-world driving scenarios under various traffic and environment conditions. These digital twins provide high-fidelity three-dimensional (3D) models that include buildings, road, and other environmental objects, which are critical for realistic vehicle and traffic simulation scenarios. These models are used to develop and validate emerging vehicle and infrastructure technologies that optimize energy efficiency and safety through connectivity, automation, and control. However, generating digital twin for such comprehensive scenario requires access to massive real-world data such as high-quality maps of the road network, road elevation and terrain, road markings and signages, building vegetations, signal timing plans and detector layouts, and traffic demand data . These real-world data can come from private or public data vendors, infrastructure sensors managed by local agencies or other entities or collected by probe vehicles with perception capabilities. Acquiring these data often involves coordination with multiple stakeholders, making the process laborious and time-consuming for researchers. Moreover, there is a lack of commercial or open-source solutions to streamline this process, leading to extended data processing and ingestion times. Furthermore, real-world data is often constrained by issues such as limited coverage, low sampling rates, noise, sensor inaccuracies, or communication loss leading to missing data. It can take months of efforts to process and integrate the data for digital twin creation in the simulation environment.

In this work, we introduce and explore a one-stop solution to tackle these challenges by developing a sharable, ready-to-use library of high-fidelity “standard” 3D digital twins for evaluation and testing of advanced vehicle and mobility technologies with a variety of inclusive traffic conditions .

To develop these digital twins, the proposed solution leverages sensing data from our own data collection platform, APaCK-V, as well as existing database (such as OpenStreetMap, USGS data). In addition, multiple 3D environment generation methods and pipelines utilizing different data sources with information fusion mechanism are developed to create digit twins based on the sensing data collected, these methods include 3D mesh reconstruction from colorized point clouds, deep-learning based 3D objects clustering-segmentation-localization-populating using lidar point clouds, and 3D building generation using footprints and attribute data from existing database. Next, with the aid of Real-Twin, a tool developed by the authors to generate realistic and calibrated traffic microsimulation model, high-fidelity digital twin is generated and deployed in simulation tools like SUMO and CARLA. The created digital twins are validated through co-simulation between micro-traffic and vehicle simulators with real vehicle in the loop to study CAV technology impacts.