DEEP STAIRS DETECTION BASED ON CONVOLUTIONAL NEURAL NETWORKS FOR VISUALLY IMPAIRED PERSONS

In 2022, an alarming 2.2 billion people worldwide suffered from some form of vision impairment, with 237 million individuals experiencing moderate to severe visual impairment, putting them at a heightened risk of accidents and injuries. Clinical studies have shown that individuals with significant vision loss are at heightened risk of accidents, particularly during activities involving motion and orientation. Recent research on assistive devices has leveraged deep learning techniques to enhance safety, particularly by detecting stairs and preventing falls. This project aims to evaluate and compare the performance of well-known convolutional neural networks (CNNs) in detecting and classifying stairs, specifically MobileNetV2, GoogleNet, and AlexNet. A dataset of 3,000 RGB images, captured at a resolution of 2268 x 4032 pixels, was used, featuring images of stairs (upstairs, downstairs) and non-stair elements. The labeled dataset was augmented to match the input layer size of the pre-trained models and processed using MATLAB R2022b. Model performance was assessed by analyzing training and validation accuracy and loss through a training progress graph. Additionally, testing accuracy, precision, recall, and F1 score were evaluated using a confusion matrix. The results demonstrated that AlexNet outperformed GoogleNet and MobileNetV2, achieving an impressive 99% accuracy across all performance metrics.