A Study to Improve the Accuracy of the Material Classification by Machine Learning Using Depth Data of an RGB-D Camera

Rescue robots have been developed to collect information and perform tasks in place of humans at disaster sites, and there is a demand for autonomous rescue robots. To make robots autonomous, map generation based on advanced environmental recognition is necessary. However, the reflectance characteristics of the laser or infrared light emitted by the distance sensor used for map generation are affected by the material of objects in the environment. So there was a problem that the accuracy of the map deteriorated depending on the materials of objects in the environment. Therefore, we have proposed a method to classify the materials of objects in the environment using data from a distance sensor. As the first step of this research, we verified whether wood, metal, and concrete could be classified using a 3D-LiDAR and an RGB-D camera in our previous studies. We compared three types of machine learning methods (decision tree, AdaBoost, and neural network), and obtained an accuracy rate of 97% for the 3D-LiDAR and 83% for the RGB-D camera using decision tree. In this study, we aim to increase the accuracy rate of the RGB-D camera to 95% or more. For the machine learning method, we use the decision tree, which was the best in previous studies. In order to improve accuracy, three types of five explanatory variables are added. In addition, we optimize the hyper parameters of the decision tree using the existing hyperparameter optimization tool Optuna. We considered the contribution of each explanatory variable to the classification with the SHAP. And the essential explanatory variables were identified. As a result, we were able to increase the accuracy rate of material classification using the RGB-D camera to 96.5% and the Macro-F1 to 96.6%. In addition, we discussed the reasons why these explanatory variables.