Obstacle Detection System for Navigation Assistance of Visually Impaired People Based on Deep Learning Techniques

Area of Science:

• Computer Vision
• Artificial Intelligence
• Assistive Technology

Background:

• Visually impaired individuals face mobility challenges impacting social integration and quality of life.
• Personal navigation systems are crucial for enhancing independence and confidence.
• Deep learning and Neural Architecture Search (NAS) offer potential for advanced navigation solutions.

Purpose of the Study:

• To develop an intelligent navigation assistance system for the visually impaired.
• To propose an efficient NAS technique for optimizing object detection models in computer vision.
• To address the computational limitations of NAS for real-world applications.

Main Methods:

• A novel, fast NAS approach was developed using a tailored reinforcement learning technique.
• The NAS method focused on optimizing the feature pyramid network and prediction stage of an anchor-free object detection model.
• The model was trained and evaluated on the Coco and Indoor Object Detection and Recognition (IODR) datasets.

Main Results:

• The proposed NAS method efficiently searched for an optimal object detection framework.
• The resulting model demonstrated a 2.6% improvement in average precision (AP) compared to the original model.
• The optimized model achieved this improvement with acceptable computational complexity.

Conclusions:

• The developed intelligent navigation system shows promise for improving the mobility and quality of life for visually impaired people.
• The proposed fast NAS technique is effective and efficient for custom object detection tasks in computer vision.
• This research contributes to advancing assistive technologies through efficient deep learning architecture optimization.