3D Human Motion Retrieval Based on ISOMAP Dimension Reduction

In recent years, with the development and increasingly mature of motion capture technology, it has become one of the most widely used technologies to obtain realistic human motion in computer animation. With the increasing demands, motion dataset is becoming larger and larger. Due to motion feature data have the high-dimensional complexity, we first adopt nonlinear ISOMAP manifold learning algorithm to resolve the "curse of dimensionality" problem for motion feature data. In order to save the time of reducing dimension, we adopt the scarcity of neighboring-graph to improve ISOMAP algorithm for making it apply the massive human motion database. Then we build a motion string index for database, deploy Smith-Waterman algorithm to compare the retrieval samples' motion string with motion strings of candidate datasets, finally, we obtain the similar motion sequence. Experiment results show that the approach proposed in this paper is effective and efficient.