Enhancing Voxel-Based Human Pose Classification Using CNN with Modified VGG16 Method

Human pose classification based on 3D point cloud is a challenging problem in pattern recognition and computer vision. The human pose prediction based on a 3D point cloud is a first step in human monitoring because the advantages are robustness to light and having an accurate location in 3D space. This research proposed a novel method for human pose classification based on a 3D point cloud to overcome that condition. However, with the proposed method of voxel-based feature extraction and Convolutional Neural Network (CNN) modified VGG16 achieved great success in classifying human poses with 3D point cloud inputs. This research proposes a CNN with modified VGG16 network to classify 3D point cloud human poses by present voxel-based feature extraction. This work uses our primary 3D point cloud data from LiDAR 32-channel. Before 3D point cloud learning, the first step is pre-processing data with normalization and extracting features with voxelization. Our experiment uses two types of classification cases, namely the classification for binary-class and multi-class 3D point cloud human poses. Experimental results show that our proposed method performs well and excellently, obtaining an accuracy value of 90% for the binary-class case and outperforming other existing methods. With our proposed method, it will be possible to recognize human poses better.