Fast 3D human body gesture recognition with multiple principal planes approximation

In applying object reconstruction techniques to the problem of 3D shape approximation, we develop two new and powerful improvements to increase the robustness and accuracy of 3D human body gesture recognition. The first, the momentpreserving principal, solves the problem of 3D shape approximation with multiple surfaces by minimizing the shape reconstruction error. The second, we represents a surface with an affine-invariant surface descriptor for representing a 3D shape with the bag-of-words (BoW) model. The approach also aims at generating a time-ordered pose codebook to speed up the keyposes detection and improve precision. Our experiments demonstrate that these contributions make the 3D human body gesture recognition not only tractable but also highly accurate for our example application.