Convolutional Neural Network-based Object Detection for Enhanced Navigation Assistance to the Visually Impaired

Abstract

The gap in the study is expected to be addressed by using this experiment and hence assisting the visually impaired individuals. To attain this a machine learning solution is proposed involving a walking stick with a camera to capture real time image of the environment. These images go through a convolutional neural network (CNN) to determine the presence of obstacles along the route. The research proposes three single-layered CNN architecture models (CNN1) and two and three-layered models (CNN2 and CNN3), using two learning rates (0.0005 and 0.005) on the basis of the accuracy, precision, recall, F1 score, and ROC-AUC metrics. CNN3 had the highest accuracy of 88 percent and the learning rate of 0.0005 with equal F1 scores of 0.88 in the obstacle and the structure category. Close behind was CNN2 with 86% accuracy and close second was generalisation of CNN1 with 79% accuracy with moderate class balance. These results were also confirmed by the application of ROC wherein CNN3 and CNN2 showed an AUC value of approximately 0.91 and 0.92 respectively which is an indicator of a good classification. The training was analysed and the learning rate of 0.0005 was found to be more appropriate in terms of generalisation and less overfitting. CNN3 was identified as the most appropriate model; this model not only has high discriminative power, but also it possesses good learning behaviour that has rendered it a reliable solution in the identification of barriers of real life in assistive navigation systems.