REAL TIME OBJECT DETECTION AND DISTANCE APPROXIMATION

Abstract

Ordinary diurnal tasks can be very strenuous for people with visual defects. Over the years many contemporary technologies have been developed to help visually impaired persons. However, these technologies are only able to narrate the contents on a mobile screen, and does not help in describing the real-world objects around a visually impaired person. The purpose of this research work is therefore, to provide a system that can detect an object and predict its distance and direction from an individual in real time. The proposed system is a combination of an object detection model and a novel algorithm to approximate the distance and direction of objects called distance approximation algorithm. The detection and localization of objects is carried out by MobileNet and Single Shot Detector which are deep neural networks and are pretrained on the COCO dataset. The detection model highlights the identified objects by means of labelled bounding boxes. The coordinates of these bounding boxes are then used by the distance approximation algorithm to predict an object’s distance and direction. The system is tested using different images and live video feed from a camera, however in order to determine the efficiency of the system images of a single object taken from various distances is used. Findings indicate that the system achieves an average accuracy of 96% in predicting the distance and thus, would be able to be effective in aiding visually impaired or blind persons.