A STUDY OF SKELETAL BASED IMAGE PROCESSING TECHNIQUE FOR CNN BASED IMAGE CLASSIFICATION

Abstract

In the twenty-first century, significant advancements in the field of computer vision facilitated a surge in the application of image classification in different industries. This work proposes an image classification technique that utilizes a Convolutional Neural Network (CNN) to simplify training by transforming raw images into reduced representations. This proposed technique is used in developing two CNN models. The first model is applied in a human-robot interactive game of Simon Says. In contrast, the second is applied in a fall detection system classifying human subjects’ actions as sitting, falling, or on-feet. An accuracy of 92.55% was achieved for the human-robot interactive game, while the fall detection algorithm yielded an accuracy of 90.79%. We hope this work will be a great addition to the research community as it can further be expanded to incorporate different areas of computer vision, such as human gesture recognition for autonomous vehicles.