With the popularity of Motion Capture Systems since 1990s,more and more 3D motion capture databases have been created and widely used in the research field of gesture recognition. How to effectively analyze motion capture data, is important to the organization and the re-use of large-scale motion capture database, as well as the follow-up application of 3D information extracted from videos. Recently, research on motion capture data has focused on how to analyze the semantic meaning of motion data. The work of this thesis is to explore new approaches to human arm gesture analysis based on the motion capture database, and present the following algorithms including extraction of gesture features and gesture classification. The thesis is organized as follows: 1. A gesture database is collected by using Motion Capture. The database, which is composed of 450 sequences of gestures made by 30 subjects, is mainly about communication gestures in our daily life, so that it can be a good test platform for communication gesture recognition. 2. Build a simple skeleton model, which can be used for the repair and display of motion capture data. For the follow-up application of motion data, .trc data format is transformed into quaternion to represent the gesture. 3. Some novel features that explicitly contain communicative information and have outstanding descriptive ability are extracted. Based on the analysis of human body parts and the spatial-temporal characteristic of gestures, collaboration between two arms, key posture, as well as the periodicity of gestures is extracted. These features function as a good way of building bridges between bottom-layer features and high-layer semantics. 4. A three-layer semantic classification tree is constructed by incorporating the semantic features on different layers by prior knowledge. Based on different features applied to different layers, classifiers like GentleBoost and Nearest Neighbor are built on different layers
修改评论