Assistive Object Recognition and Obstacle Detection System for the Visually Impaired Using YOLO

Human vision provides us with information that is irreplaceable and indispensable. Accessible visual data for humans is of utmost importance for their safety, navigation, identification, and many other activities. Therefore, to aid the visually hindered, the need of the hour is to develop a smart, automated system to gain characteristic information about the area around them and help them navigate an unknown and unmapped environment. With massive developments in machine learning and computer vision, various devices and applications have been developed to assist the visually impaired. The limitations of these techniques are their usability in a real-world scenario and affordability in third-world and developing countries. Thus, this paper comprehensively compares various state-of-the-art (SOTA) algorithms for detection and classification and outlines an application to assist the visually impaired by utilizing a handheld gadget. It incorporates different strategies to assemble a model that will classify and recognize the objects and provide an estimated distance from the nearest obstacle through an auditory medium. The model takes advantage of Raspberry Pi’s portability, computational power, and modularity, creating a cost-effective handheld. It uses the YOLOv3 algorithm to perform object recognition in near real-time scenarios, which provides the best overall ratio of mAP to Time Taken and utilizes the robust dataset of MS COCO. As a result, the proposed device provides the best performance in a real-time environment, computing results within 50ms with mAP comparable to other SOTA algorithms. It aims to provide an alternative to the walking stick and other external help required by the visually challenged.