Model-based gesture recognition /
We describe work towards a new gesture recognition technique that integrates both physical and control models of gestural actions. Gesture recognition is the process of automating the recognition of distinct patterns of human motion that are intended to communicate information, intent or command. O...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
2000.
|
| Subjects: | |
| Online Access: | http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=727853311&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | We describe work towards a new gesture recognition technique that integrates both physical and control models of gestural actions. Gesture recognition is the process of automating the recognition of distinct patterns of human motion that are intended to communicate information, intent or command. Our technique incorporates an underlying model of the physical motion involved with performing a specific gesture. Typically gesture recognition techniques have fallen into categories based on their analytical approach towards recognizing a pattern. Model-based recognition starts with the underlying dynamics of the human performing the gesture. The work is analagous to the study of robot manipulators which involve setting up the dynamics and control of motion of a manipulator arm. We detail the development of a prototype arm model, dynamics, incorporation of state estimation into the matching portion of our recognizer, a control and blending system, and "proof of concept" validation experiments. Our technique integrates a set of underlying control models and blending functions with a dynamic model, with one controller and blending function for each of the gestures in the recognition set. These models augment a set of Kalman-filter-based recognizer modules that filter the input presuming that one of the gestures is being performed. The blending functions supplement the control system by mixing the filtered data with the underlying gesture model. The augmented filter outputs are compared with the outputs of an unaugmented Kalman filter to determine the recognized gesture. Experiments validate the generalizability of our algorithm across a user population, and the reliability and extendability to different sets of gestures. The results show reliability across a user population that exceeds 98% accuracy. |
|---|---|
| Item Description: | Vita. "Major Subject: Computer Science". |
| Physical Description: | xii, 107 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilm Inc. |
| Bibliography: | Includes bibliographical references (leaves 97-106). |