Obstacle Segmentation with Encoder-Decoder Architectures in Low Structured Environments for the Navigation of...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Assistive Technology

Background:

• Visually impaired individuals often rely on mobility aids like white canes for navigation.
• Current assistance systems capture environmental data but require efficient processing for real-time use.

Purpose of the Study:

• To develop an efficient deep learning approach for segmenting obstacles in outdoor environments for visually impaired people.
• To evaluate different encoder-decoder architectures using depth images for navigation assistance.

Main Methods:

• Utilized encoder-decoder deep learning architectures with depth images for obstacle segmentation.
• Developed an efficient training data generation method using the v-disparity technique.
• Evaluated depth-based and RGB-D fused architectures on a mobile computing unit.

Main Results:

• Demonstrated the applicability of the approach for mapping surrounding obstacles to support navigation.
• Quantitative evaluation showed some limitations, while qualitative assessment confirmed practical use.
• Achieved efficient obstacle segmentation in low-structured outdoor environments.

Conclusions:

• The developed approach effectively combines classical image processing, environmental knowledge, and deep learning for navigation support.
• This technology can significantly enhance the independence and quality of life for visually impaired individuals.
• The system aids in mapping obstacles, facilitating safer and more self-sufficient navigation.