Project Objective & Overview

Phase I demonstrated a scalable, low-cost approach to collecting street-level data and linking built environment conditions to travel behavior, without third-party data.

• Designed, built, deployed, and operated a scalable, low-cost street-level data collection system to capture imagery at scale and enable nationwide scalability
• Fully re-architected our visual machine learning pipeline to extract infrastructure, context, and travel behavior features, transitioning to a zero-shot ensemble methodology
• Developed predictive models linking built environment features to travel behavior using data from a single deployment, enabling cost-effective, citywide travel behavior estimation

Infrastructure = physical street assets; Context = land use and urban form/design; Travel behavior = pedestrian, bicycle, and vehicle activity.

Key Results & Findings

• Designed and deployed a data collection system using cell phones, collecting 2M+ images across 8,000+ unique blocks in the Philadelphia/SEPTA service area, demonstrating the feasibility of independent, scalable street-level imagery collection and eliminating reliance on third-party data sources
• Expanded built environment features from 127 to 150 fully re-architected feature extraction pipeline using a new 298-model zero-shot ensemble architecture, enabling block-level extraction of infrastructure, context, and travel behavior features
• Achieved strong and consistent model performance (≈90% unbalanced, 88% balanced accuracy, average) and developed predictive models linking built environment features to travel behavior, achieving r² ≈ 0.5+ for pedestrian and vehicle activity
• Demonstrated that one-time built environment data collection, combined with limited observed travel behavior data, can support citywide travel behavior estimation, providing a cost-effective alternative to sensor- and mobile phone–based approaches for large-area analysis
• Lessons learned: Phase I highlighted the need to further improve hardware robustness, incorporate additional data modalities, and refine software workflows to support repeated deployments and operational use at scale