Zebra Crossing Spotter: Automatic Population of Spatial Databases for Increased Safety of Blind Travelers
Dragan Ahmetovic, Roberto Manduchi, James M. Coughlan, Sergio Mascetti · 2015 · ASSETS '15: Proceedings of the 17th International ACM SIGACCESS Conference on Computers & Accessibility · doi:10.1145/2700648.2809847
Summary
This paper presents a computer vision system that automatically detects zebra crosswalks by mining satellite imagery and Google Street View, with the goal of augmenting geographic information systems (GIS) to help blind travelers plan safer routes. Knowing crosswalk locations is critical for blind pedestrians because marked crosswalks grant right-of-way to pedestrians and are more visible to drivers—crossing outside marked areas requires yielding to vehicles, which is extremely difficult without sight. The system uses a cascade classifier approach. First, it analyzes satellite images to identify candidate zebra crossings based on their distinctive striped pattern, using an adapted version of the ZebraLocalizer algorithm. Satellite images are efficient because they cover large areas with minimal data download. The algorithm only processes images containing roads (determined via Google Maps API), reducing unnecessary downloads by about one-third. Second, each candidate is validated against nearby Google Street View panoramas, which provide ground-level confirmation. This two-stage approach achieves both high accuracy and computational efficiency—processing a 1.6km² area requires only 39MB of satellite imagery plus 16MB of Street View images, versus 637MB if starting with Street View alone. The technique was evaluated on a dense urban area in San Francisco containing 141 manually-labeled zebra crossings and 152 transverse crossings.
Key findings
The cascade classifier achieved excellent accuracy: 97% precision (few false positives) and 95% recall (few missed crossings). Satellite detection alone achieved 97% recall but only 69% precision, with false positives often corresponding to rooftop patterns or building features. Street View validation filtered out 58 of 62 false positives while losing only 3 true positives. The system is computationally efficient. Processing 791 satellite images covering the test area took 142 seconds; Street View validation required only 406 images (averaging 1.8 panoramas per candidate crossing) and 18.5 seconds. Total CPU time was 161 seconds for 1.6km². The algorithm is also parallelizable for faster large-scale processing. False negatives occurred when crosswalk paint was discolored/faded or when crossings were obscured by trees, shadows, or vehicles. The few remaining false positives were patterns highly similar to zebra crossings, such as outdoor staircases. The authors note that crowdsourcing could provide a final validation stage—it is far easier for crowdworkers to confirm or reject a candidate than to exhaustively search for missed crossings.
Relevance
This research addresses a significant gap in accessible navigation infrastructure. While GPS-based navigation apps exist for blind travelers, they are only as useful as their underlying maps—and most GIS databases lack pedestrian-relevant features like crosswalk locations, curb ramps, and accessible pedestrian signals. By automatically extracting crosswalk data from existing imagery, this technique could rapidly populate accessibility-enhanced maps at scale. The practical applications are substantial. Blind travelers could use crosswalk databases for route planning (preferring routes with marked crossings), real-time navigation guidance, and integration with smartphone-based crosswalk detection apps that help with alignment and approach. The cascaded computer vision approach demonstrates how different imagery sources can be combined for efficient, accurate feature extraction. For organizations building accessible navigation systems, this work provides a model for augmenting GIS with accessibility-relevant features without requiring costly manual surveys. The technique could be extended to detect other pedestrian infrastructure such as curb ramps, accessible pedestrian signals, and tactile paving.
Tags: visual impairment · navigation · computer vision · pedestrian crossings · wayfinding · GIS · crowdsourcing · orientation and mobility