Abstract
Purpose :
Many blind people take public transportation, especially in metropolitan cities. Due to the lack of pin-point accuracy of conventional geo-location information provided by mobile devices, when navigating to the bus stops, they sometimes may experience the ‘last-10-meter’ problem, which refers to a gap between the bus stop and the destination announced by the mobile devices. The gap may sometimes be large enough for bus drivers to misunderstand the blind people’s intention and not stop the buses for them. This pilot study evaluates an AI solution to address the localization problem.
Methods :
We developed a mobile AI app, All Aboard, which detects bus stop signs in the users’ vicinity in the images captured by the smartphone camera in real-time. The deep neural network for bus stop sign detection was custom-trained using 5000 to 10000 images collected for a given city. The app guides the users to approach bus stop signs through auditory cues, with pitch coding the distance to the target. In a pilot test, we evaluated the accuracy of localizing bus stops using the All Aboard app versus the Google Map navigation for 20 bus stop locations in the Los Angeles area. 10 bus stops were in the downtown area near high-rises, and the other 10 stops were in suburban areas. The experimenter walked under the guidance of auditory cues given by All Aboard app and Google Maps, respectively, starting from about 25 meters away from each bus stop. The localization error was measured as the distance from the actual bus stop signs to the location where the apps indicated the user arrived at the destinations.
Results :
The successful navigation rate with Google Maps was 60% (failed to localize 7 stops in downtown and 1 in suburban), and with All Aboard app was 95% (failed to detect 1 downtown stop). The All Aboard app was able to detect bus stop signs starting from an average distance of 9.7±6.3 (SD) meters. Excluding navigation failures, the localization error with All Aboard app (average ± SD: 1.1±0.6 meters, max. = 2.2 meters) was significantly lower than Google Maps (7.4±3.8 meters, max. = 13 meters) (paired t-test, p < 0.001).
Conclusions :
The promising results suggest that the bus stop sign detection app can potentially address the “last-10-meter” problem for blind individuals taking bus transportation.
This is a 2021 ARVO Annual Meeting abstract.