|
||
|
||
Mathematics and Cognition Seminar Fall 2009 Tuesdays 12:15 Psych 161 Seminar Schedule: <http://math.la.asu.edu/~tom/cognition/math+cogsched.html> Cookies and Coffee Starting at 12:00 Note the New Location! Map ("X" marks the spot) |
On Tuesday, September 22, at 12:15 in Psych 161, |
 |
|
On the topic: |
||
"A Dynamic Bayesian Approach to Computational Laban Shape Quality Analysis " |
|
Abstract We are developing experiential media systems, a new generation of human-computer interfaces that are situated in the physical world and that respond meaningfully to natural human activity via immersive, multisensory feedback. Examples have already been developed for select applications in stroke patient rehabilitation, K-12 education, interactive dance performance, and the mediation of complex systems and data sets. Critical to the success of these systems has been the ability to develop a computational understanding of human movement at the level of meaning and intention. While as in language, many aspects of intention remain context/situation-specific, there do exist intermediate levels of representation, analogous to phonemes in speech, that serve as a precursor and that do generalize across contexts. One of the more complete and universally recognized frameworks across dance, kinesiology, and related disciplines for representing "body language" is Laban Movement Analysis (LMA), originated by Rudolf Laban in the 1920's. Despite the fact that humans can be trained to interpret and annotate all human movement in terms of LMA, it still remains difficult to recognize LMA qualities computationally from video or even motion capture data. The dimension of Shape (how the body forms itself directionally in space) is particularly difficult as there is no one prescribed way to perform any of the shape qualities (rising/sinking, advancing/retreating, or enclosing/spreading). One may "advance", for instance, by walking towards an object in a directed manner, or simply by craning one's neck forward in a subtle way. To this end, we have developed a robust and extensible Bayesian fusion approach for identifying LMA Shape qualities from motion capture data. Our approach uses a dynamic Bayesian network (DBN) to fuse novel movement features across the body and across time, and was developed in close collaboration with a certified Laban Movement Analyst (CLMA). The method operates in real-time and delivers excellent performance in preliminary studies comprising natural and improvisatory movements. Although currently relying on motion-capture data, the method can be readily adapted for low-cost, portable video camera arrays (eg. new Playstation cameras: $40/each) to be installed in everday public spaces and home environments. . Bio Harvey Thornburg joined the ASU faculty in 2005 with a joint appointment in Arts, Media and Engineering and Electrical Engineering. Current research activities involve audio sensing and content analysis, as well as multimodal data fusion. In a broader sense, his research addresses the representation of contextual knowledge emerging from flexible and uncertain structural forms (for instance: those arising from the syntax of music and dance) and the fusion of this knowledge with raw sensory information to improve detection and estimation capabilities. |