Evaluating Road Surface Condition by using Wheelchair Driving Data and Positional Information based Weakly Supervision

Abstract

Providing accessibility information on sidewalks for mobility impaired people is an important social issue. Until now, the authors have evaluated the accessibility of sidewalks by estimating the road surface condition by supervised learning on the accelerometer data mounted on wheelchairs. Video recording and data labeling to accelerometer data based on the video for teacher data require enormous costs and become problematic. This paper proposed and evaluated a new weakly supervised road surface condition evaluation system of using positional information automatically acquired at driving as a label. The evaluation result showed that weakly supervised learning method using locational label captured detailed features of road surfaces, and classified moving on slopes, curb climbing, moving on tactile indicators, and others with a mean F-score of 0.57 and accuracy of 0.71 close to those of supervised learning method.