3D CNN hand pose estimation with end-to-end hierarchical model and physical constraints from depth images

• DOI: 10.14311/nnw.2023.33.003
• Publication Date: 2023-03-21T08:01:42Z
• Authors: Zhengze Xu; Wenjun Zhang
• Keywords: End-to-end principle

Previous studies are mainly focused on the works that depth image is treated as flat image, and then data tends to be mapped gray values during convolution processing features extraction. To address this issue, an approach of 3D CNN hand pose estimation with end-to-end hierarchical model physical constraints proposed. After reconstruction space structure from converted into voxel grid for further by CNN. The method makes improvements embedding algorithm networks, resulting train at fast convergence rate avoid unrealistic pose. According experimental results, it reaches 87.98% mean accuracy 8.82 mm absolute error (MAE) all 21 joints within 24 ms inference time, which consistently outperforms several well-known gesture recognition algorithms.