Action Graph: A Versatile Data Structure for Action Recognition

This work presents a novel and generic data-driven method for recognizing human full body ac-tions from live motion data originating from various sources. The method queries an annotated motion capture database for similar motion segments, capable to handle temporal deviations from the original motion. The approach is online-capable, works in realtime, requires virtually no pre-processing and is shown to work with a variety of feature sets extracted from input data including positional data, sparse accelerometer signals, skeletons extracted from depth sensors and even video data. Evaluation is done by comparing against a frame-based Support Vector Machine ap-proach on a freely available motion capture database as well as a database containing Judo referee signal motions and concludes by demonstrating the applicability of the method in a vision-based scenario using video data.