Self-Supervised Soft Obstacle Detection for Safe Navigation of Visually Impaired People

the context of computer-assisted navigation of visually impaired people (VIP), time-efficient and robust obstacle detection methods are of major importance. Most commonly, obstacle detection algorithms exploit depth information, acquired from specialized sensors, to characterize objects as obstacles. These algorithms usually incorporate computationally expensive sub-processes, such as ground detection and removal, saliency estimation etc. In this work, we propose a self-supervised system based on a Convolutional Neural Network (CNN) that learns such an obstacle detection algorithm, and simulates it, with significantly lower computational requirements for safe navigation of VIP. The input of the CNN is an RGB image, and its output is a saliency map, softly approximating the image regions that correspond to possibly high-risk obstacles. The resemblance of the saliency maps simulated by the proposed system and the original algorithm is 71.46%, assessed in terms of the Judd implementation of the area under receiver operating characteristic (AUC-J), with a sufficiently comparable obstacle detection accuracy.