Barrier detection using sensor data from unimpaired pedestrians

Abstract

There are several barriers (e.g., steps, slopes) that hinder the free movement of impaired people, both indoors and outdoors. Existing approaches for detecting barriers have an accuracy/coverage trade-off problem. For example, approaches that use a wheelchair with an accelerometer cannot detect barriers in areas that wheelchair users have not gone through. However, approaches that try to detect barriers from street images on the Internet fail to increase the accuracy of barrier detection because of occlusions that obscure the surface of the road. To address this problem, we propose a barrier detection approach that uses a machine learning model trained with acceleration data acquired from smartphones of able-bodied pedestrians. This idea uses pedestrians as sensor nodes for detecting barriers. This approach enables us to collect barrier information for a large area with high accuracy without any special investigators or devices. The results of the evaluation using acceleration data of pedestrians show that our method could identify barriers accurately by using hand-crafted features. Furthermore, we also clarify that the identification accuracy improves when features auto-generated by deep learning (Denoising Autoencoders) are used.