SKELETON ACTION RECOGNITION BASED ON SINGULAR VALUE DECOMPOSITION

This work introduces new method using the singular value decomposition (SVD) to recognise human activities from skeleton motion sequences. The primary focus was on different activity durations, inaccurate placement of the joints and loss of information about position of the joints. For that we needed to develop a robust model. At first, the pose features are created for description of skeleton pose per frame, that is created by directional vectors to all joint pairwise combinations without repetition. The data timeline is divided in logical hierarchical parts using temporal pyramid decomposition. This helps to code the time information. On each temporal part, SVD is computed from each pose feature. From the final decomposition of SVD, the singular values representing variance and the right singular vectors representing the fitted planes in space-time are used to characterise the pose feature trajectory. The support vector machine classifier uses these features for recognition. Some of the results of the experiments are similar, and some outperform the results of the state-of-the-art methods. The next analysis verified the robustness of the proposed method to Gaussian noise and to the loss of coordinates data of joints. The primary contribution of proposed method is robustness against noise and information loss with preservation of the state-of-the-art results.