GuideDogNet: A Deep Learning Model for Guiding the Blind in Walking Environments

Abstract

A guide dog is a critical companion for the blind, which enables independent travel of the blind. However, due to the costly and time-consuming training process, only 1.7% of the blind who wish to adopt a guide dog can take it. In order to alleviate this social problem, previous studies have suggested several blind guiding systems heavily based on hardware devices, such as GPS(Global Positioning System), RFID(Radio-Frequency Identification), and ultrasonic devices. However, those techniques lack administrative feasibility to use in real-world environments. Moreover, those techniques are deficient in warning of obstacles, which makes the system non-user-friendly. To guide the blind universally and provide accurate information about the obstacles without cumbersome devices, we propose a novel deep learning-based blind guiding system, GuideDogNet. The proposed system consists of an object detection network, depth prediction network, and post-processing module. To provide user-friendly outputs for the blind, we propose a rule-based post-processing module that outputs the label, direction, and distance of the obstacles by combining the results of the object detection network and the depth prediction network. We achieved an mAP of 67.8 on the AI Hub Sidewalks dataset which is publicly available. To the best of our knowledge, this is the first attempt at a deep learning-based blind guiding system.

The code will be available on https://github.com/Yunseo-Hwang/AI-GuideDog