Froehlich gives lecture on data-collection methods
Despite comprehensive civil rights legislation that has improved the quality of life in the U.S., there is still a lot that needs to be done. For 30.6 million people with physical disabilities, ambulatory activities present many hardships. About half of those people have reported using assistive aid, including wheelchairs, canes, crutches. or walkers. Consequently, many city streets, sidewalks, and businesses are inaccessible for individuals that rely on assistive aid. For example, a pole positioned in the middle of a sidewalk panel would be considered an inaccessible place for a person traveling on a wheelchair.
However, quickly reconstructing such areas is not feasible for many cities, especially ones that are burdened by economic constraints. Yet, assistant professor of computer science at the University of Maryland Jon Froehlich believes that the solution to the problem lies in developing mechanisms to identify accessible places and making that data available for individuals who need it.
On Wednesday, Froehlich shared his vision and research with Carnegie Mellon as part of a lecture series hosted by the university’s Human Computer Interaction Institute (HCII). Titled “Characterizing Physical World Accessibility at Scale Using Crowdsourcing, Computer Vision, and Machine Learning,” the lecture focused on novel scalable data-collection methods for obtaining information about how accessible manmade places are.
In collaboration with graduate students from his research group Makeability Lab, along with another Maryland computer science professor David Jacobs, Froehlich used a combination of crowdsourcing, computer vision, and online map imagery (i.e. Google Street View) within a series of projects.
Crowdsourcing, in this case, involves asking human subjects to identify and label inaccessible areas when given photos on the computer. Similarly, computer vision (CV) uses computer algorithms to identify such areas. By separately applying crowdsourcing and CV to images taken from Google Street View (GSV) — a search engine that showcases panoramic views from positions along many streets in the world — the members of the Makeability Lab were able to experiment with methods in which accessibility information could be collected and visualized.
However, Froehlich also presented the limitations in his methods. With crowdsourcing comes subjectivity and significant time consumption. In one study, crowd workers were found to identify areas that did not actually present inaccessibility issues. In contrast, the CV technique worked more quickly. But given the automated nature of computer algorithms, CV was statistically found to show less accurate results. Additionally, the age of many GSV images overlooked recent reconstruction of inaccessible areas.
Keeping these problems in mind, Froehlich and his group focused on combining crowdsourcing and CV. From their work, they developed the first “smart” system, Tohme, which incorporates two work flows: human labeling and CV labeling with human verification. The work flows are scheduled dynamically based on predicted performance. In their research, they used 1,086 GSV images of street intersections in four North American cities, as well as data from 403 crowd workers.
From their results, they saw that the system was able to perform similarly to crowdsourcing alone, but with a 13 percent reduction in time costs. While the study only focused on curb ramps, the group believes that the approaches used in developing Tohme can be applied to other locational obstacles. Pointing out room for improvement in Tohme, Froehlich also emphasized the need for more advanced computer and machine algorithms. Nevertheless, he believes that Tohme can be made more efficient.
During the question-and-answer portion of the seminar, many attendees expressed their fascination with Froehlich’s research, which may potentially improve the lives of many people struggling with physical disabilities.
In addition, Froehlich’s lecture presented many implications of the increasing role of data collection in everyday lives.