A Method for Identification of Door Type in an Image by Machine Learning

Abstract

The importance of nursing robots is increasing. There are various doors and doors in a home, but it is difficult for a robot to move through all the rooms. Therefore, the purpose of this paper is to identify doorknobs and handles in images using machine learning as a preliminary step to the development of an autonomous mobile robot that can open and close arbitrary doors. Since there are various types of doors and doorknobs in a home, we narrowed down to two types, sliding doors and hinged doors, and prepared three types of doorknobs for each door type. In this study, we used Random Forest as one of the machine learning. This model uses majority voting or averaging to reduce overfitting, which is a problem with many decision tree machine learning techniques. We have prepared 800 images for each of the six typical doorknob patterns. It was learned by machine learning. As a result, the test data could be detected with a high accuracy of 0.96 or higher. On the other hand, the accuracy of the actual door data differs depending on the type of doorknob, and some have a high accuracy of 0.9 or higher and some have a high accuracy of 0.75. In the future, I will try other methods of machine learning.