Motion recognition based on manifold learning spectral clustering

With the emergence of numerous 3D human motion capture databases, the effective analysis and handling of human motion data have become a major challenge so that the use of motion capture databases can be maximized. To reduce the high-dimensional complexity of data, a type of geometrical feature based on 2D geometrical space law is first extracted from human motion for the application of motion data into a low-dimensional subspace. With the aim of achieving a low-dimensional feature, identification and classification in different motions are then conducted through spectral clustering based on manifold learning to realize the automatic identification and retrieval of 3D human motion. © 2014 SERSC.