Modified CNN VoxNet Based Depthwise Separable Convolution for Voxel-Driven Body Orientation Classification

摘要

In this research, LiDAR sensor technology introduces a new representation of point cloud data for tasks in 3D object recognition. Point clouds provide rich information that can be utilized to predict the orientation pose of the human body. The orientation of the human body is crucial for various needs such as autonomous vehicles, robotics, and human activity monitoring. However, the inherently unstructured nature and permutation invariance of point cloud data pose significant challenges in accurately predicting human body orientation using 3D data. In this paper, we propose a solution by introducing a volumetric method based on CNN to effectively recognize human body ori-entation. A novel approach is proposed by embedding depthwise sparse convolution into 2D CNN to process voxel data of human body orientation. We evaluate the performance using KITTI data divided into four orientation classes: east, west, north, and south. Additionally, we compare it with other architectures using the ModelNet10 and ModelNet40 datasets. The comparison with other architectures shows superior performance with accuracies of 84.31 % for the KITTI dataset, 86.67 % for ModelNet10, and 75.20 % for ModelNet40. This evidence strongly supports the effectiveness of our proposed method, highlighting its potential in enhancing 3D object recognition systems significantly.