Prediction of Human Body Orientation based on Voxel Using 3D Convolutional Neural Network

Abstract

Robot interaction with humans requires intelligent robots that can understand human activities. The development of advanced 3D LiDAR sensors has greatly contributed to this capability. In this study, we specifically focus on the use of 3D LiDAR sensors to predict the orientation of the human body using 3D Convolutional Neural Networks (CNNs) based on voxelized datasets. The dataset used in this study was created using a 3D LiDAR sensor with 32-channel specifications. We divided the dataset into four categories representing different walking orientations. The goal was to explore the performance of four different 3D CNN architectures using independently generated datasets. Based on the experimental results and performance analysis, it was found that VGG16 outperformed the other three architectures in predicting body orientation. VGG16 achieved an accuracy of 0.95, which was higher than DenseNet121 with approximately 0.91, ResNet50V2 with 0.80, and ResNet50 with 0.73. In the future, this method will be developed with additional orientation and results of architectural testing so that it can be modified to be better for further research on understanding human activity by robots.