Human Pose Classification Using a Large Kernel Attention-driven 3D Transformer Model (PointLARKA) with LiDAR Point Cloud Features

Human 3D point cloud data present significant challenges due to non-rigid articulation, pose variation, and complex surface geometries.To address these issues, we propose PointLARKA.This novel transformer-based framework integrates Geometry Disentangled Attention and Large Kernel Attention for effective 3D human point cloud classification.The proposed model learns complementary sharp and gentle geometric cues through GDA and refines them using Sharp-Gentle Complementary Attention.Instead of using a regular block, we utilize the residual mechanism in the large kernel attention block, then process these features to capture both fine-grained local geometry and long-range spatial dependencies efficiently.By combining graph-based structural reasoning with large receptive field modeling, PointLARKA achieves discriminative and context-aware feature learningExperimental findings show that the proposed method attains accuracies of 93.2% on ModelNet40, 96.59% on the 3D LiDAR Human Pose dataset, and 86.4% on ScanObjectNN, and an average LOSO accuracy of 88.6% ± 10.9% on our 3D LiDAR human-pose, surpassing the performance of current state-of-the-art approaches.These results highlight the effectiveness of the proposed GDA-LKA fusion in enhancing geometric understanding and classification performance for complex 3D human point clouds.Our implementation source code is available at: https://github.com/virgantara/PointLARKA.