Behavior modeling and recognition on moving object is an important researching direction. As for behavior analysis, the extraction of feature is influenced by the conversion of viewpoint, the change of illumination, the occlusion, and the expression of behavior is unstable. How to extract the low level feature and express the behavior and how to describe the motion state and reduce calculation from the time and the dimension are both key problem of video understanding. As to the expression method of behavior modeling, how to classification and recognition of video action by machine recognition technology is a problem eager to solve. The extraction and expression of low level feature and the building of analysis model are very important for a behavior analysis system. we did a series of research job from feature aspect to analysis model. The main contributions of this thesis include following issues: 1.In accordance with the idea of the behavior to be analyzed from the simple to the complicated,we recalled the methods of behavior analysis in recent years from six aspects. There are six kinds of methods include in:non-parametric method, volumetric method, parametric method, graphical models method, syntactic method, knowledge based method. 2.An action analysis method based on invariant feature is proposed. Firstly, we project the image sequence to the Radon domain.Secondly, we do the R transform, then expand the curve to a two dimension image. Finally, the feature is extracted from this image by Fourier-mellin transform. In the experiment, we test and verify the invariance, the distinguishing and the anti-noising of our method.The experimental result show the effectiveness of our method. While before extracting this kind of feature, the silhouette of human should be extracted.So the effective foreground extraction method is very important for the application of this kind feature. 3.An action classification method based on prototype learning is proposed. The method of machine learning is used in action analysis more and more widely. The prototype learning method have the characteristic of low storage need and low computational complexity. So prototype learning method can be used in a human action classification system. The simple and easy-got feature is very important for the system too. Our method is verified by the experiment. 4.A behavior analysis method based on metric learning and prototype learning is proposed. Because distance metric is very import...
修改评论