Mind Your Crossings: Mining GIS Imagery for Crosswalk Localization
Dragan Ahmetovic, Roberto Manduchi, James M. Coughlan, Sergio Mascetti · 2017 · ACM Transactions on Accessible Computing (TACCESS) · doi:10.1145/3046790
Summary
This paper presents a computer vision system for automatically detecting and localizing zebra crosswalks in urban environments by mining existing geospatial image databases, specifically Google satellite imagery and Google Street View. The motivation is that blind pedestrians face significant safety challenges at street crossings, and knowing the precise location of marked crosswalks is valuable for both trip planning and real-time navigation — yet crosswalks are not systematically catalogued in existing GIS databases like OpenStreetMap or Google Maps. The system uses a cascaded detection approach: first, a line segment detector identifies candidate zebra stripe patterns in satellite images, filtering by geometric properties such as parallelism, spacing, and orientation; then, candidates are validated against spatially registered Street View panoramic images using a second detection stage that looks for the characteristic vertical stripe pattern of zebra crossings from ground level. This two-stage approach leverages the complementary strengths of each image source — satellite images provide broad coverage for initial detection while Street View images offer higher-resolution confirmation from a pedestrian perspective. The authors also conducted a survey of nine Orientation and Mobility (O&M) experts from the United States and Italy to understand the role crosswalks play in blind pedestrian travel. The survey revealed stark cultural differences: U.S. experts emphasized standard O&M techniques at all legal crossings regardless of markings, while Italian experts strongly recommended crossing at marked crosswalks whenever possible, reflecting differences in traffic regulations between the two countries.
Key findings
The algorithm was evaluated on four urban regions — two in San Francisco, one in New York City, and one in Milan, Italy — covering a total area of 7.5 square kilometres with manually labelled ground truth. The cascaded satellite-plus-Street-View approach achieved precision ranging from 0.902 to 0.971 and recall from 0.772 to 0.950 across the four regions. A key finding was that the two detection stages are complementary: satellite detection alone produced many false positives from stairways and similar striped patterns, while Street View validation effectively filtered these out. To further improve accuracy, the authors developed the Pedestrian Crossing Human Validation (PCHV) web service for crowdsourced verification. In a preliminary study with 10 human validators on the Milan region, crowdsourcing boosted recall from 0.772 to as high as 0.97 and precision from 0.948 to 0.989, with each validator spending an average of 27 minutes. The O&M expert survey confirmed that crosswalk location information would be valuable for blind travellers, particularly for trip planning, and highlighted that the type and amount of information needed varies by the traveller's experience level and familiarity with an area.
Relevance
This work addresses a fundamental gap in accessible pedestrian infrastructure data: despite the critical importance of crosswalks for safe street crossing by blind pedestrians, no existing mapping service systematically catalogues their locations. The automated detection approach offers a scalable solution that could populate navigation databases used by blind travellers, enabling both safer route planning and real-time guidance. The cross-cultural O&M survey findings are particularly valuable for practitioners, revealing that accessibility solutions must account for local traffic regulations and cultural practices — a crosswalk navigation aid designed for U.S. users may need different assumptions than one for European users. The combination of automated computer vision with crowdsourced human validation demonstrates a practical model for building large-scale accessibility databases cost-effectively. For transit accessibility advocates, this research strengthens the case that spatial data infrastructure should include pedestrian crossing features as a standard layer, not an afterthought.
Tags: visual impairment · orientation and mobility · computer vision · navigation · crowdsourcing · GIS · pedestrian safety · urban accessibility