Pedestrian-Accessible Infrastructure Inventory: Enabling and Assessing Zero-Shot Segmentation on Multi-Mode...

Area of Science:

• Geospatial Science
• Computer Vision
• Urban Planning

Background:

• Traditional pedestrian infrastructure inventories often omit accessibility-critical street furniture.
• Mobile LiDAR and satellite imagery offer rich data for infrastructure mapping.
• The Segment Anything Model (SAM) shows promise for zero-shot segmentation tasks.

Purpose of the Study:

• To develop and optimize a SAM-based workflow for segmenting pedestrian infrastructure from multi-source geospatial data.
• To investigate the use of mobile LiDAR and street-view imagery for comprehensive inventory creation.
• To evaluate SAM's performance in segmenting diverse pedestrian infrastructure assets.

Main Methods:

• Designed and optimized a Segment Anything Model (SAM) workflow for geospatial data.
• Utilized mobile LiDAR point-cloud data to generate street-view images.
• Integrated street-view images with satellite imagery for multi-modal input.
• Employed zero-shot segmentation to identify pedestrian infrastructure and street furniture.

Main Results:

• The SAM-based approach efficiently segments pedestrian infrastructure from combined LiDAR-derived street-view and satellite imagery.
• Street-view images from mobile LiDAR, when paired with satellite data, facilitate effective zero-shot segmentation.
• The method successfully expands the definition of pedestrian infrastructure to include accessibility features.

Conclusions:

• The proposed workflow offers a scalable and efficient method for creating comprehensive pedestrian-accessible infrastructure inventories.
• This approach provides immediate benefits for GIS professionals, city managers, and transportation planners.
• The findings particularly support improved accessibility for individuals with travel-limiting disabilities.