Improving Public Transit Accessibility for Blind Riders by Crowdsourcing Bus Stop Landmark Locations with Google Street View: An Extended Analysis
Kotaro Hara, Shiri Azenkot, Megan Campbell, Cynthia L. Bennett, Vicki Le, Sean Pannella, Robert Moore, Kelly Minckler, Rochelle H. Ng, Jon E. Froehlich · 2015 · ACM Transactions on Accessible Computing · doi:10.1145/2717513
Summary
This paper tackles a fundamental challenge in public transit accessibility: helping blind and low-vision riders locate bus stops, which often lack nonvisual markers or consistent landmark placement. The researchers developed Bus Stop CSI (Crowdsourcing Streetview Inspections), a web-based tool that enables crowd workers on Amazon Mechanical Turk to virtually navigate Google Street View (GSV) imagery and label physical landmarks near bus stops, such as shelters, benches, trash cans, mailboxes, and signage. The work is grounded in three complementary studies. Study 1 involved formative interviews with 18 people with visual impairments across urban, suburban, and small-town settings, exploring how they navigate to and identify bus stops. These interviews confirmed that physical landmarks like shelters, benches, and even nearby businesses play a critical role in nonvisual navigation, with participants reporting strategies ranging from listening for echoes off shelters to smelling nearby food vendors. Study 2 compared physical in-person bus stop audits with virtual audits conducted through GSV across four neighborhoods in Washington, DC and Seattle, surveying 179 bus stops along 42 linear kilometers. Using Krippendorff's Alpha for interrater reliability and Spearman rank correlation to compare datasets, the researchers established that GSV imagery closely mirrors on-the-ground conditions, validating it as a viable remote audit data source. Study 3 deployed Bus Stop CSI on MTurk, where 153 crowd workers labeled landmarks at 150 bus stop locations. The tool featured a four-stage interactive tutorial system and two interaction modes: Explorer Mode for navigating the GSV panorama, and Labeling Mode for placing landmark markers.
Key findings
Individual crowd workers achieved 82.5% accuracy in correctly identifying the presence or absence of bus stop landmarks across 150 locations. With a simple seven-turker majority vote for quality control, accuracy increased to 87.3%. Performance varied by landmark type: mailboxes and newspaper bins were labeled most accurately (88.8%), followed by bus stop shelters (88.6%) and benches (83.3%), while traffic signs and other poles had the lowest accuracy (66.2%) due to the category being open-ended and ambiguous. Bus stop locations rated as easy by researchers yielded 84.5% per-turker accuracy, compared to 74.3% for medium-to-hard locations. Scene difficulty was driven by five factors: occlusion from parked vehicles, distance from the GSV camera, ambiguous proximity of landmarks to bus stops, poor lighting and shadows, and inaccurate bus stop coordinates in the Google Maps API. The physical-versus-virtual audit comparison showed high Spearman rank correlations across all landmark types (all rho > 0.60, p < 0.001), with bus stop shelters and benches achieving the highest correlation (rho = 0.88 and 0.88). Analysis of the 21 poorly performing turkers revealed they tended to underlabel rather than overlabel, with higher false-negative rates (1.5 vs. 0.7 average) than the remaining 132 workers.
Relevance
This research demonstrates a scalable, low-cost method for collecting detailed bus stop accessibility information that transit agencies and navigation app developers could integrate into their services. The approach is particularly significant because it does not require on-the-ground volunteers or specialized equipment — anyone with internet access can contribute through the GSV-based crowdsourcing tool. For accessibility practitioners, the interview findings offer concrete evidence of what landmark information matters most to blind and low-vision travelers, which can inform how transit agencies describe their stops. The work also highlights a broader pattern relevant to accessibility: existing geographic datasets like Google Transit often lack the granularity needed by people with disabilities, but crowdsourcing combined with freely available street-level imagery can fill those gaps. Limitations include dependence on GSV image currency and coverage, the challenge of accurately representing bus stop coordinates, and the need for further research on what minimum accuracy threshold is useful for real-world navigation.
Tags: public transit · blind and low vision · crowdsourcing · wayfinding · landmarks · Google Street View · bus stops
Standards referenced: ADA