Indexing and retrieval of human motion data based on a growing self-organizing map

As low-cost depth cameras are released recently, motion data containing 3D coordinates of skeleton joints during a time period can be directly captured. Nevertheless, analyzing the motion data is usually a challenging problem and requires huge computation costs because of the high-dimensionality. Among several alternatives, the self-organizing map (SOM) is verified to be an effective technique to handle such motion data. Specifically, a captured motion sequence can be easily and precisely mapped to form an indexed motion string through the use of a trained SOM. However, the training process of the SOM is high computation complexity and is thus typically tedious. In view of this, we propose in this work to incorporate a hierarchical structure into the SOM technique. Generally speaking, our approach named as GQSOM (growing quadtree self-organizing map) helps to significantly reduce the required computation complexity while preserving the effectiveness of the SOM technique. Empirical studies using the WorkSU-10 exercise dataset show that our approach is both efficient and effective to perform indexing and retrieval tasks of motion data.