Infrastructure Systems: Digital Twins for Smart Cities

REU Scholar: Mohammed Chowdhury

REU Scholar Home Institution: Lehman College

REU Mentor: Dr. Jinwoo Jang

Semantic Segmentation for Street-level LiDAR scans

This project focuses on the semantic segmentation of urban LiDAR scans to extract and classify street-level elements, including vegetation, buildings, vehicles, and pedestrians, supporting the development of urban digital twins. An end-to-end pipeline is implemented to manipulate multi-source point clouds and object labeling data, enabling training of a PointNet neural network for LiDAR segmentation. The PointNet segmentation leverages per-point, multi-layer perceptrons and symmetric max-pooling to learn robust, permutation-invariant shape descriptors directly from raw point clouds. The architecture is enriched with feature transformation layers and class-balanced loss functions to sharpen boundary delineation in dense urban scenes. Furthermore, interactive 3D visualizations are implemented to enable rapid qualitative assessment of segmentation performance. The resulting labeled and segmented point clouds can be integrated within digital twin and situational awareness environments, offering detailed street-level semantic information for applications in autonomous navigation, urban planning, and smart-city monitoring. Future work will validate the framework on additional real-world scans, extend it across diverse urban datasets, and produce side-by-side visual comparisons that deliver the depth of results and visualizations needed to better understand the functioning of vegetation, buildings, agents, and vehicles in the streetscape.

Watch video presentation
REU Scholar: Mohammed Chowdhury