The use of formal grammars in syntactic pattern recognition is reasonable if decomposition of the entity under consideration into a set of primitives that lend themselves to automatic recognition is possible. In this paper we presented a method of combining the probabilistic model of simple ``vocabulary'' gestures and the higher level structural knowledge for the tasks of recognition of temporal behaviors and activities. The method is used for semantic segmentation, disambiguation and labeling complex behaviors that feature the decompositional properties.
With all the advantages of the method there are still problems that were not addressed in our work, for instance, grammar inference is hard to do in the framework of stochastic parsing. We have not researched the opportunities of fully utilizing the production probabilities. In the experiments above we determined them heuristically using simple reasoning based on our understanding of the process. The rule probabilities, which reflect ``typicality'' of a string derivation, will play a significantly more important role in recognizing and interpretation of activities of the higher complexity than those presented in the paper. We plan exploring the added advantages of the learned probabilities in our further research.