Independent pathfinding with collision avoidance for visually impaired individuals.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Assistive Technology

Background:

• Image segmentation is crucial for visually impaired navigation but requires detailed spatial information.
• Existing methods face challenges in providing robust and high-performance solutions for safe navigation.

Purpose of the Study:

• To propose a novel panoptic segmentation framework for an effective pathfinding system.
• To combine robust collision avoidance with high performance for visually impaired individuals.
• To support future integration with real-time feedback in visual prostheses.

Main Methods:

• Developed a single-stage instance segmentation method using a ResNet101-FPN encoder-decoder architecture.
• Created a customized panoptic labeled dataset tailored for the needs of visually impaired individuals.
• Evaluated the model qualitatively and quantitatively using the Panoptic Quality (PQ) metric.

Main Results:

• The proposed method achieved a Panoptic Quality (PQ) score 4.092 points higher than recent panoptic segmentation techniques.
• Demonstrated superior accuracy and efficiency compared to existing pathfinding systems.
• Showcased effectiveness under varied weather conditions.

Conclusions:

• The novel panoptic segmentation framework provides a strong foundation for advanced pathfinding systems.
• The customized dataset and ResNet101-FPN architecture contribute to improved performance for assistive technology.
• This work advances computer vision applications for enhancing the independence and safety of visually impaired individuals.