Assistive Navigation for the Visually Impaired: Real-Time Crosswalk Detection Using Satellite Images and Machine Learning

Abstract

Visually impaired individuals face many challenges when navigating urban areas, particularly when they are crossing streets. Existing navigation tools lack real-time detection of crosswalks, and depend on immediate surroundings, posing challenges to safe and independent navigation. This study addresses this limitation by presenting a wearable that utilizes machine learning to detect crosswalks in real-time from satellite imagery. This approach with the use of Google Maps satellite imagery ensures accurate and timely crosswalk detection, unlike the proximity-based navigation systems that rely on local sensors. A dataset of 1,740 satellite images annotated for binary classification of crosswalks and non-crosswalks instances was trained using the YOLOv5s object detection model. To enhance detection accuracy, the model was fine-tuned using hyperparameter optimization, including batch size, learning rate, and the number of epochs. The precision of the trained model was 95.24%, recall 98.45%, and mean average precision (mAP@0.5) 99.71%. The trained model, along with a Raspberry Pi 4 and GPS receiver, gives audio notifications at 5 m from a crosswalk, enabling the user’s safety and independence. Simultaneously, the dependence on internet connectivity to obtain satellite images, and the image quality being hampered due to low resolutions can be named as the limitations of the study. The contribution of this study lies in bridging the gap between existing assistive technologies and the need for real-time, reliable crosswalk detection. This research offers a novel, energy-efficient, and cost-effective solution for visually impaired individuals navigating urban environments by integrating machine learning with satellite imagery. Future directions will expand the dataset by including additional images, optimizing model efficiency, and enhancing adaptability across diverse environmental conditions. Unlike traditional navigation aids, which lack real-time crosswalk detection, this system provides an innovative way to ensure safety and independence, especially in urban areas.