Human Posture Recognition Based On Manifold Learning

Recently, visual analysis of human motion is one of the hot topics in the field of computer vision. It mainly includes motion detection, moving object classification, human tracking and activity recognition and description. This paper mainly concerns human motion detection and the recognition of human posture, and the objective is to provide more information for human tracking, activity recognition, etc by analyzing human motion postures. The research results could be applied in smart surveillance, motion analysis, etc. This paper consists of image sequence preprocessing, dimensionality reduction of the silhouette sequences and human posture recognition.Image sequence preprocessing consists of two parts: (1) Silhouette sequences are obtained by extracting the human moving target from each original image sequence. (2) In order to effectively utilize the global feature of the silhouette and the relations among frames of each silhouette sequence, the manifold learning algorithm is employed and the images of each silhouette sequence are normalized into the same size.In some research fields, the dimensionality of the data is very high and the useful information is submerged by large quantities of data. How to reduce the dimensionality of the original data and extract the useful information is a critical problem. The dimensionality reduction algorithm can be divided into two classes: linear and nonlinear ones. Manifold learning belongs to the latter class. Seven influential dimensionality reduction algorithms are summarized and the principle's deductions, procedures and some applications of them are discussed in detail. The Locally Linear Embedding algorithm (LLE) is selected to process the silhouette sequences and the experimental results shows that relations among frames can be well preserved and the amount of computation in the following recognition part decreases sharply.The sample database is created by the silhouettes sequences. Based on the global feature of the human silhouette, the relations among frames of the silhouette sequences are mapped into the Euclidean distance by making use of manifold learning algorithm, recognition process is dealt with in the low dimensional Euclidean space. The test silhouette sequence is recognized based on Mean Hausdorff Distance, Gaussian Model and Hidden Markov Model respectively. Finally, the analysis and summary are given.