Road Safety Conditions Pilot in California Using Computer Vision and Artificial Intelligence Reported 97 Percent Accuracy in Pothole Detection.
Pilot Project in San José Evaluated if Automated Object Detection Could Increase Efficiency in Identifying and Responding to Hazards on City Streets.
San Jose, California, United States
Road Safety Conditions Pilot AI Object Detection Initiative Status Report
Summary Information
San José launched its Road Safety Conditions Pilot in December 2023 to use Artificial Intelligence (AI) for automatically identifying and addressing roadway hazards. The pilot used cameras mounted on City vehicles that fed data into an AI model and tested whether AI and computer vision could improve street safety by automatically detecting hazards such as potholes, illegal dumping, and blocked bike lanes. Road Safety Conditions Pilot ran from December 2023 through July 2024, collecting footage primarily from streets within the City’s District 10. The City hoped to move beyond resident reporting through the 311 system and create a more proactive, automated process, thus reducing delays and improving maintenance planning.
METHODOLOGY
The City engaged four vendors to test AI detection capabilities and evaluated them across four categories: 1) technology availability, 2) performance, 3) compatibility with City processes, and 4) public input. City staff reviewed test footage and compared it to the AI detection results, confirming the accuracy above and beyond vendor-reported statistics. Testing focused on identifying potholes, obstructive trash, and 72-hour parking violations. Graffiti and encampment detection were removed from the pilot due to capacity limitations and privacy concerns. All personally identifiable information was blurred during collection.
FINDINGS
- Pothole detection achieved high accuracy (97 percent true positive rate). Most vendors already provide pothole detection, and public feedback shows strong support for using AI to identify potholes and accelerate repairs.
- Trash and debris detection had an 88 percent accuracy rate, though some concerns were raised about misidentification of encampments.
- 72-hour parking violation detection was unreliable, as AI model was unable to track vehicles that moved slightly.
- Recreational vehicles detection was 70 percent accurate, while lived-in vehicle detection was only 12.5 percent accurate.
