Intelligent Navigation System for the Visually Impaired - A Deep Learning Approach

摘要

Visually impaired individuals have been gradually claiming a significant stake in the population demographics. The proposed autonomous device aims to provide a holistic solution by engineering a smart navigation system that relentlessly scans the environment, detects and classifies neighboring objects using a 4 layered Convolutional Neural Network (CNN) that has been trained on a data set containing 2513 permutations of various images of household objects that an individual may encounter in daily life. The CNN follows the 80-20 rule for testing and training the self-learning model enabling it to learn recursively from the error rate. The proposed system then calculates distances of neighboring objects from the user and provides adaptive solutions in real time to manoeuvre the user to safety by providing auditory input in a simplistic manner which considers 10-24 frames per second while drafting the kinematic response for the user. The device has achieved an unprecedented success rate of serving within a response time of less than 50 ms. The accuracy of the CNN algorithm being at 94.6%, also sets a distinguished benchmark as an object detection algorithm thereby contributing to the success in simulations of the proposed device in a constrained environment.