Privacy by design
Faces, license plates, and clothing are filtered out at the source, leaving only the shapes and motion the map actually needs.
Structures as Sensors Lab
Neighborhood Drive
Privacy-preserving mobile sensors on community vehicles map neighborhood conditions in San José, allowing cities and community organizations to better serve unhoused residents.
Project Introduction
Neighborhood-scale perception with privacy built in.
Serving unhoused residents well depends on current, ground-level knowledge of where people are. People living in RVs, encampments, and shelters move across neighborhoods over weeks and days in response to their own needs and external stressors, particularly displacement. Service providers and city agencies need to know where they move and how their needs have changed, including environmental conditions like air quality, heat, and sanitation.
Mobile, not fixed
Sensors ride on vehicles already serving the community, covering far more ground than fixed cameras at a fraction of the cost.
Community-engaged
The system is shaped by ongoing surveys with the people it affects. Their input informs our definition of "privacy-preserving" and determines what gets sensed and how.
Our Work
Privacy-preserving vehicle detection
A pipeline that turns RGB video from vehicle-mounted cameras into bounding boxes around vehicles, with appearance details removed at the source. Thermal data downstream identifies which detected vehicles are lived-in.
Environmental sensing
Gas, temperature, humidity, and noise sensors map ambient conditions block by block, capturing air quality, heat exposure, and acoustic environments along service routes.
Built-environment computer vision
Detection of supportive infrastructure, such as toilets and handwashing stations, and deterrent infrastructure, such as planters and concrete blocks, from street-level imagery.
Federated learning
Models are trained across multiple vehicles without centralizing raw data, and used to predict effective routes for service providers.
Community perspectives
Surveys with Loaves and Fishes clients on privacy and monitoring preferences, alongside public opinion polling on policies affecting unhoused residents. In partnership with the City of San Jose.
Project Team
Collaborators across sensing, cities, and community-centered AI.
PIs
Project Leader
Prof. Hae Young Noh
Department of Civil and Environmental Engineering, Stanford University
Co-PI
Prof. Sarah Billington
Department of Civil & Environmental Engineering, Stanford University
Co-PI
Prof. Carlee Joe-Wong
Department of Electrical & Computer Engineering, Carnegie Mellon University
Co-PI
Prof. Pei Zhang
Department of Electrical and Computer Engineering, University of Michigan
Students / Researchers
Postdoctoral Researcher
Dr. Amelie Bonde
Department of Civil & Environmental Engineering, Stanford University
PhD Student
Jatin Aggarwal
Department of Civil & Environmental Engineering, Stanford University
PhD Student
Zhili He
Department of Civil & Environmental Engineering
HKUST & Stanford University
MS Student
Shang Gao
Department of Civil & Environmental Engineering, Stanford University
PhD Student
Daria Fontani Herreros
Department of Civil & Environmental Engineering, Stanford University
PhD Student
Yi Hu
Department of Electrical & Computer Engineering, Carnegie Mellon University
PhD Student
Julia Gersey
Department of Electrical & Computer Engineering, University of Michigan
Previous Members
Outcomes
Publications and presentations from the NeighborDrive collaboration.
Conference Workshop Paper
NeighborDrive: Privacy-preserving Neighborhood Sensing through Outreach Vehicles
Best Paper Award, 2nd Place
Presented at the SocialSys workshop, ACM Sustainability Week Companion 2026, in Banff, AB, Canada (June 22-25, 2026).
Aggarwal, J., Gao, S., Bonde, A., He, Z., Gersey, J., Fernandez, T. S., Rejeev, R., Zhang, P., & Noh, H. (2026). NeighborDrive: Privacy-preserving Neighborhood Sensing through Outreach Vehicles. ACM BuildSys 4th International Workshop on Social Infrastructure Systems (SocialSys'26), June 22-25, 2026, Banff, AB, Canada.
Conference Paper
GLUE: Gradient-free Learning to Unify Experts
Presented at IEEE ICASSP 2026.
Park, J. I., Chaudhari, S., Pranav, S., Joe-Wong, C., & Moura, J. M. (2026, May). GLUE: Gradient-free Learning to Unify Experts. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 2681-2685). IEEE.
Conference Workshop Paper
Human vs. Machine: Comparing Urban Condition Classification Methods from Vehicular Vision
Presented at BuildSys 2025.
Julia Gersey, Troy Zhong, Jiale Zhang, Jesse Codling, Jackelyn Hwang, and Pei Zhang. 2025. Human vs. Machine: Comparing Urban Condition Classification Methods from Vehicular Vision. In Proceedings of the 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys '25). Association for Computing Machinery, New York, NY, USA, 427-430. https://doi.org/10.1145/3736425.3772319
Workspace Outputs
One image, multiple privacy-preserving representations.
Annotated
SAM3 mask overlays and detection boxes for prompt verification.
Privacy
Human regions replaced by solid black for identity protection.
Depth Mask
Only human regions become Depth-Anything-V2 depth; the rest remains RGB.
Estimated Depth
The complete scene rendered as a colorful Depth-Anything-V2 depth image.