Enhancing point cloud regression for human pose estimation with hyperbolic embedding

Abstract

In the domain of human pose estimation using mmWave radar-generated point cloud data, the integration of advanced neural network models and novel embedding techniques is crucial for enhancing accuracy and efficiency. This dissertation introduces a method, HyperPose, which builds upon the robust feature extraction capabilities of the PointMLP and DGCNN models, integrating hyperbolic embedding as a novel enhancement for exceptional feature representation. Hyperbolic space, known for its superiority in modeling hierarchical data compared to Euclidean spaces, is utilized to address the compositional complexity of human poses. By embedding the extracted features into hyperbolic space, HyperPose effectively captures the complex hierarchical relationships inherent in human poses, thereby significantly improving the accuracy of pose estimation. Our comprehensive experiments conducted on the mm-fi dataset demonstrate HyperPose’s competitive performance in terms of precision and computational efficiency. Furthermore, experiment confirm the critical role of hyperbolic embedding in enhancing the model’s ability to discern complex human poses from mmWave radar-generated point clouds. The results not only underline the effectiveness of HyperPose but also open new avenues for research into point cloud-based human pose estimation and the application of hyperbolic spaces in deep learning architectures