Owli-AI Logo Owli-AI
Menu

Owli-AI Research

An Electronic Guide Dog for the Blind based on Artificial Neural Networks

(2021) - Paper

S. Lopatin; F. v. Zabiensky; M. Kreutzer; K. Rinn; D. Bienhaus

Technische Hochschule Mittelhessen, University of Applied Sciences, Institute of Technology and Computer Science, Giessen, Germany

Notice: this page was machine-translated and is pending editorial review.

Visual

Functional scheme of the Electronic Guide Dog with camera input, CNN segmentation, and derived navigation information.
Fig. 1 Functional scheme of the Electronic Guide Dog.

Abstract

This paper presents a feasibility study for an electronic assistance system intended to support blind and visually impaired people in orientation and navigation in public space. The core approach is optical detection of walkable sidewalk areas using semantic segmentation with a neural network trained from scratch. In the practical implementation, an NVIDIA Jetson Nano is used as a mobile computing unit for on-device inference. Navigation cues are derived from detected sidewalk structures and output to the user as speech. The work therefore examines technical feasibility of a portable electronic guide dog based on computer vision and CNN.

Keywords

  • electronic travel aid
  • blind sidewalk detection
  • portable ETA system
  • electronic travel aid technology
  • computer vision
  • convolutional neural network

Download

Download PDF Back

Figures

7 visuals from the paper.

  1. Functional scheme of the Electronic Guide Dog with camera input, CNN segmentation, and derived navigation information.
    Fig. 1 Functional scheme of the Electronic Guide Dog.
  2. System architecture with ROS2 nodes for camera, CNN segmentation, environmental analysis, and audio output.
    Fig. 2 System architecture with data-flow based processing.
  3. Scatter plot comparing inference speed and mIoU of different segmentation architectures.
    Fig. 3 Comparison of speed and precision.
  4. Example segmentation results for sidewalk areas from the test dataset.
    Fig. 4 Comparison result of one test sample for DeepLabV3 and BiSeNetV2.
  5. Visualization of parameters used to calculate deviation from the optimal sidewalk path.
    Fig. 5 Parameters for deviation calculation from optimal path.
  6. Photos of the experimental prototype with chest camera, computing unit, and headphones.
    Fig. 6 Experimental prototype in field-test setup.
  7. Segmentation examples from field tests with marked walkable sidewalk area.
    Fig. 7 Segmentation examples from field test.