Improving 3D Human Pose Orientation Recognition Through Weight-Voxel Features And 3D CNNs

Abstract

Preprocessing is a widely used process in deep learning applications, and it has been applied in both 2D and 3D computer vision applications. In this research, we propose a preprocessing technique involving weighting to enhance classification performance, incorporated with a 3D CNN architecture. Unlike regular voxel preprocessing, which uses a zero-one (binary) approach, adding weighting incorporates stronger structural information into the voxels. This method is tested with 3D data represented in the form of voxels, followed by weighting preprocessing before entering the core 3D CNN architecture. We evaluate our approach using both public datasets, such as the KITTI dataset, and self-collected 3D human orientation data with four classes. Subsequently, we tested it with five 3D CNN architectures, including VGG16, ResNet50, ResNet50v2, DenseNet121, and VoxNet. Based on experiments conducted with this data, preprocessing with the 3D VGG16 architecture, among the five architectures tested, demonstrates an improvement in accuracy and a reduction in errors in 3D human orientation classification compared to using no preprocessing or other preprocessing methods on the 3D voxel data. The results show that the accuracy and loss in 3D object classification exhibit superior performance compared to specific preprocessing methods, such as binary processing within each voxel.