Representation and Symbolization of Motion Captured Human Action by Locality Preserving Projections

Human motion analysis and assessment are important in determining Parkinsons disease and stroke, or in measuring skill quality in basic motions. Reduced space is useful in representing motion segments and finding basic behavioral patterns for humanoid robot control using the modularized approach. In the current paper, we represent motion-captured data of human action in a reduced space of nonlinear degrees of freedom in which the original motion is characterized. First, we represent high-dimensional data, such as motion sequence of the position of joints in Cartesian space, in a reduced space using the locality preserving projection (LPP) method. Second, we find a similarity measure between the actions. Finally, we assess human motions using a similarity measure to find the most similar one. The LPP is a linear dimensionality reduction algorithm that builds a graph for neighborhood information and maps data points to a reduced space. The reason for using LPP in our study is that it is defined globally, and any new data element can be mapped in the reduced space. Our method includes the generation of symbolic code sequence corresponding to complex, high-dimensional motion. Interdisciplinary synergy combined with information technology and wearable sensor systems can broaden the possible future applications in rehabilitation engineering.