Automatic Extraction of Semantic Features for Action Recognition

Deeper understanding on human actions is required in many applications, e.g., action recognition (security), animation (sport, 3D cartoon movies and virtual world), etc. The task of automatic extraction of semantic action features within the data set is gaining in importance. We propose novel approaches to automatically extract the action features from 3D Motion Capture data for high accuracy in real-time action recognition. In particular, we contribute to two different areas dealing with variation at different features levels, i.e., 1) Extract of Discriminate Patterns from Skeleton Sequences approach is based on features that are close to the raw data. It provides a foundation in lower dimensional representation for the movement sequence analysis, retrieval, identification and synthesis; and 2) Automatic Extraction of Semantic Action Features approach focuses on solving the high-dimensional computational problems arising from the human motion sequences. It supports the follow-up stages of processing the human movement on a natural language level. As one common underlying concept, the propose approach contains a retrieval component for extracted action features.