Research On Methods Of Human Action Recognition Based On Manifold Learning

With the development of computer technology and network technology, people get information from images and videos become easier. Among these vast amounts of information, people always pay more attention to the information about human action. Human action recognition is one of the popular research direction of computer vision and machine learning. It is to study how to make computer capture and recognize human action in the videos or images automatically. The research of human action recognition has a wide range of applications such as: intelligent video surveillance, content-based video retrieval, motion analysis, human-computer interaction and so on. However, due to the complexity and diversity of human actions, there are still some big challenges in improving the accuracy and speed of human action recognition.This thesis focuses on the application of manifold learning in human action recognition, especially for the human action recognition in videos and images. The main research content of this thesis is about the human action features extraction and classifier design. In the part of human action feature extraction, we discuss several manifold learning methods that can be used to dimensionality reduction of action features, and propose a framework for action recognition based on manifold learning. Then by considering of the manifold learning algorithms principle and experiment results, this thesis analyze the limitations of some traditional manifold learning methods when they are used for the dimensionality reduction of action features. Then some improved methods and algorithms are proposed. In the part of classifier design, this thesis chose DTW classifier which is widely used in the field of speech recognition. And then we make some improvement on it. The main work and achievements of thesis can be concluded as follows:(1) In the process of action features dimensionality reduction, traditional manifold learning methods have a difficulty in dimensionality reduction of data on disconnected manifold. This thesis solve it by combining the manifold learning algorithms with the spectral clustering algorithm. In this thesis there are two kinds of traditional algorithms Isometric Feature Mapping(ISOMAP) and Laplacian Eigenmaps(LE) improved under this idea. Finally, experiments on the artificial dataset and human action database show these two improved algorithms can both get good results.(2) In this thesis, Nystr?m approximation is used to improve the incremental learning ability of classical Laplacian Eigenmaps. Then this method also be used to improve the disconnected manifold learning algorithm based on LE aboved. Finally, some experiments are used to check the effectiveness of improved algorithms.(3) The classical DTW classifier cannot keep good performance when it be used to recognize the robustness test action dataset. An improved classifier combine DTW with KNN algorithm is used to solve this problem. The experiments on Weizmann robustness action database show the effectiveness of the improved classifier.