Stochastic regular grammar-based learning for basic dance motion recognition

This paper presents a simple and computationally efficient framework for 3D dance basic motion recognition based on syntactic pattern recognition. In this research, a class of basic dance motions is modeled by a stochastic regular grammar (SRG), inferred from training dataset, in which key body poses that are learned from training dataset are selected as gesture primitives. To represent a dance motion, body pose of a dancer is represented by angular coordinate of 15 skeleton joints. This feature is later compacted into one-dimensional string of labels for grammar inference which makes the recognition process is considerably fast compared to statistical pattern classifier such as k-nearest neighbor (kNN). A single test using the learned grammar in average takes only about 5 ms compared to around 20s using kNN whilst the overhead time to build all grammars takes only about 3.4s. This compacting process, however, leads to information loss which is observed in slightly degraded recognition performance for low articulated motions but quite large degradation for high articulated dance motions. To overcome this, we investigate several reliable feature selection methods such as Sequential Feature Selection (SFS), Principal Component Analysis (PCA), and Heuristic Sequential Feature Selection (HSFS) compared to the use of whole features. Based on our experiment, the HSFS is the most suitable feature selection to overcome this problem.