Dangerous Object Detection for Visually Impaired People using Computer Vision

Abstract

In this contemporary world, Artificial Intelligence and Machine Learning are one of the leading technologies creating an impact in the world by mimicking human behaviour to solve a particular problem. Hence, these technologies are widely used to aid different obstacles encountered by humans. One such problem widely faced by the mankind is visual impairment. According to World Health Organization, approximately 285 million people suffer with vision impairment. Therefore, applications of machine learning and computer vision can be applied to guide the people with such problems. This paper presents the idea of using object detection to aid the visually impaired people. In this paper, an experiment has been proposed which uses a custom-built image dataset of various dangerous objects. The objects have been categorized into 5 broad categories: Sharp objects, Danger signs, Broken glass, Manhole and Fires. A number of different algorithms have been trained on this custom image dataset containing the menacing objects and their performances have been evaluated. The evaluation indicators for the models are the validation error in terms mean Average Precision (mAP) and the processing time for each model. The models have also been tested in real world scenario by evaluating on a custom video to gauge their performance in terms of accuracy in detection of different objects as well as their ease in deployment by suggesting their frame rate handling capacity. The results are discussed and the most robust and balanced model is suggested at the end of the paper.