@incollection{endres_bayesian_2016,
	address = {Cham},
	title = {Bayesian {Approaches} for {Learning} of {Primitive}-{Based} {Compact} {Representations} of {Complex} {Human} {Activities}},
	isbn = {978-3-319-25739-6},
	url = {https://doi.org/10.1007/978-3-319-25739-6_6},
	doi = {10.1007/978-3-319-25739-6_6},
	abstract = {Human full-body activities, such as choreographed dances, are comprised of sequences of individual actions. Research in motor control shows that such individual actions can be approximated by superpositions of simplified elements, called movement primitives. Such primitives can be employed to model complex coordinated movements, as occurring in martial arts or dance. In this chapter, we will briefly outline several biologically-inspired definitions of movement primitives and will discuss a new algorithm that unifies many existing models and which identifies such primitives with higher accuracy than alternative unsupervised learning techniques. We combine this algorithm with methods from Bayesian inference to optimize the complexity of the learned models and to identify automatically the best generative model underlying the identification of such primitives. We also discuss efficient probabilistic methods for the automatic segmentation of action sequences. The developed unsupervised segmentation method is based on Bayesian binning, an algorithm that models a longer data stream by the concatenation of an optimal number of segments, at the same time estimating the optimal temporal boundaries between those segments. Applying this algorithm to motion capture data from a TaeKwonDo form, and comparing the automatically generated segmentation results with human psychophysical data, we found a good agreement between automatically generated segmentations and human performance. Furthermore, the segments agree with the minimum jerk hypothesis about human movement [32]. These results suggest that a similar approach might be useful for the decomposition of dances into primitive-like movement components, providing a new approach for the derivation of compressed descriptions of dances that is based on principles from biological motor control.},
	language = {en},
	urldate = {2026-02-21},
	booktitle = {Dance {Notations} and {Robot} {Motion}},
	publisher = {Springer International Publishing},
	author = {Endres, Dominik and Chiovetto, Enrico and Giese, Martin A.},
	editor = {Laumond, Jean-Paul and Abe, Naoko},
	year = {2016},
	keywords = {Bayesian Information Criterion, Independent Component Analysis, Segment Boundary, Source Function},
	pages = {117--137},
}