Combining Crowdsourcing and Machine Learning to Collect Sidewalk Accessibility Data at Scale

Abstract

We are developing new data collection approaches that use a combination of remote crowdsourcing, machine learning, and online map imagery. Our newest effort, called Project Sidewalk, enables online crowdworkers to remotely label pedestrian-related accessibility problems by virtually walking through city streets in Google Street View. In 2019, we completed an 18-month deployment in Washington, D.C.: 1,150+ users provided over 200,000 geo-located sidewalk accessibility labels and audited 3,000 miles of D.C. streets. With simple quality control mechanisms, we found that minimally trained remote crowd workers could find and label 92 percent of accessibility problems in street view scenes, including missing curb ramps, obstacles in the path, surface problems, and missing sidewalks. For our PacTrans project, we proposed three threads of additional work. (1) First, we are deploying Project Sidewalk into three more cities, including two in the Pacific Northwest: Seattle, Washington, and Newberg, Oregon, to enable us to study and compare sidewalk accessibility factors across cities. (2) Second, to further scale our approach, we proposed new methods to automatically identify and classify sidewalk problems using deep learning techniques, which would be uniquely enabled by our large dataset. (3) Finally, we proposed new sidewalk accessibility models and interactive visualization tools to give stakeholders—from citizens to transit authorities—new understanding of their city’s accessibility.