Deep Learning-Enabled Obstacle Detection and Navigation Aid for Visually Impaired Using YOLOv4-Tiny

Visual impairment presents a significant barrier to independent mobility and personal safety for millions worldwide. This work introduces a deep learning-based assistive system built on the YOLOv4-tiny architecture, designed to support visually impaired users through real-time object detection, distance measurement, and navigation assistance. Unlike earlier YOLOv3-or YOLOv5-based aids that often trade off speed for accuracy or lack efficient feedback mechanisms, the proposed system strikes a balance between high-speed processing and high-precision detection. It uniquely integrates auditory and haptic feedback mechanisms to promptly alert users to nearby obstacles (within a 30-inch threshold) and potholes, addressing challenges in dynamic environments. Experimental evaluation demonstrates strong system performance, achieving a mean Average Precision (mAP) of 0.93, an Intersection over Union (IoU) of 0.89, and a precision score of 0.97 on the KITTI dataset, all while maintaining real-time processing speeds of 45 frames per second. Supplementary dynamic visualizations—such as continuous distance-time plots and periodic object class histograms—provide operational transparency and aid in system calibration. Notably, the system outperforms YOLOv3-based implementations by 20% in detection accuracy and operates with a latency of under 50 milliseconds, ensuring a seamless user experience on low-power edge devices. These results affirm the solution’s potential as a scalable, cost-effective mobility aid for visually impaired individuals in real-world, diverse environments.