Human articulated body parts bending motion classification based on Dictionary-Learning Sparse Representation

Abstract

In this paper a method is developed to estimate human articulated body parts bending motion based on Dictionary-Learning Sparse Representation (DLSR). The extracted features for training the dictionary are achieved by deformation gradient of proposed part, which is the non-translation portion of an affine transformation that determines the change between original shape and deformed shape. In order to train the dictionary for motion classification, we minimize the reconstruction error of the target shape. Then, all trained dictionaries from motion classes are combined to construct an over-complete dictionary for sparse representation and classification. We evaluate our approach to different topological structure of human arm and leg shape. The experimental results show the effectiveness of our approach for treating the bending motion classification in different images.