A Comparative Study of Neural Network Structures for Detection of Accessibility Problems

抄録

Identifying accessibility problems (e.g., steps, steep road) is beneficial for enabling the smooth movement of impaired/elderly people. To construct accessibility maps that satisfy both the accuracy and coverage, we have proposed a crowdsourcing platform that requires people to acquire inertial sensor data during walking; accessibility problems are detected by a neural network that analyzes the sensor data. However, appropriate network structures for detection of accessibility problems have not been discussed. Accordingly, in this paper, we compare neural network structures for detection of accessibility problems. The preliminary study results showed that Type-wise structure network that concatenates data according to data type (i.e., acceleration data or rotation rate data) yielded the highest performance in detecting accessibility problems.