Extraction of Semantic Dynamic Content from Videos with Probabilistic Motion Models