Efficient Motion-Searching Using the Joint Decomposition Approach
Kosuke Sugano, Taishi Yamamoto, Kenta Oku, Kyoji Kawagoe · 2017
A large amount of human action data can be captured in databases, with the rapid performance improvement and popularization of sensor devices. Although many methods of human motions search based on multidimensional time series from sensors have already been proposed, both the accuracy should be improved and the processing cost should be reduced. In this paper, the authors verify that the noise generated in a specific joint can be reduced by performing similarity calculation for each joint. The authors propose the Joint Decomposition approach and apply it to A-LTK (Approximation using Local features at Thinned-out Keypoints) to enable similarity calculation and weighting for each joint. The Joint Decomposition approach focuses on parameter of A-LTK and sets for each joint. Parameter represents the approximation level. The authors evaluate the Joint Decomposition approach along with existing methods. The evaluation result shows that the Joint Decomposition approach is better than other methods. From this result, it is found that noise generated in a specific joint can be reduced.