📝 SummarXiv

Abstract

LiDAR-based perception in intelligent transportation systems (ITS), for tasks such as object detection, tracking, and semantic and instance segmentation, is predominantly solved by deep neural network models which often require large-scale labeled datasets during training to achieve generalization. However, creating these datasets is costly. time consuming and require human labor before the datasets are ready for training models. This hinders scalability of the LiDAR-based perception systems in ITS. Sim2Real learning offers scalable alternative, however, its effectiveness is dependent on the fidelity of the source simulation(s) to real-world, in terms of environment structure, actor dynamics, and sensor emulations. In response, this paper introduces a rigorous and reproducible methodology for creating large-scale, high-quality synthetic datasets using High-Fidelity Digital Twins (HiFi DTs). The proposed workflow outlines the steps, tools, and best practices for digitally replicating real-world environments, encompassing static geometry modeling, road infrastructure replication, and dynamic traffic scenario generation. Leveraging open-source and readily available resources such as satellite imagery and OpenStreetMap data, alongside specific sensor configurations, this paper provides practical, detailed guidance for constructing robust synthetic environments. These environments subsequently facilitate scalable, cost-effective, and diverse dataset generation, forming a reliable foundation for robust Sim2Real learning.