PointVoxel-Net: A Residual Learning Approach to Human Pose-Oriented 3D Classification Using PointGrid Features

The continuous evolution of 3D sensor technology has increased the demand for robust perception systems that can handle real-world complexities. Although deep learning has advanced 3D object recognition and orientation estimation, point cloud data remain sparse, noisy, and irregular, posing significant challenges. To overcome these issues, we propose PointVoxelNet, a hybrid architecture that integrates voxel grid representations with point-level features enhanced by residual learning. Instead of binary voxelization, sampled point features are embedded within voxels to enrich the geometric details and improve generalization. Extensive experiments confirmed the model’s effectiveness, with accuracies of 90.67 %, 92.29% on ModelNet10 and 99.64% on HumanPose (LiDAR-based). In occlusion scenarios, PointVoxelNet remained robust, achieving 86.67% accuracy at a 0.8 ratio, and outperformed VoxNet, PointNet, PointGrid, and VR5. Overall, the results demonstrate that combining residual learning with point-in-voxel representations enhances the accuracy, robustness under occlusion, and computational efficiency. Although it requires more resources than lightweight models, PointVoxelNet offers a favorable trade-off, establishing its strong potential for advancing 3D orientation recognition and object classification in both synthetic and real-world applications.