Video Analysis Using Semi-supervised Manifold Learning

Video analysis is an active research topic in computer vision and pattern recognition, but the dimensionality of the video data is very high.High dimensional data is hard to understand for human being and will bring high computational complexity . So learning the low dimensional representations for video data is necessary.Linear dimensionality reduction,for example PCA,is an effective technique for data compression.However, due to the nonlinear variation in video data,traditional linear dimensionality reduction methods are not applicable.So we need to find some nonlinear dimensionality reduction techniques to extract features from the video data.Manifold Learning is a nonlinear dimensionality reduction technique developed recently.It can well preserve the nonlinearities of the manifold.As one of the main manifolds learning algorithms, Isomap is based on MDS and it aims to preserve the inner geometric properties of data points via preserving geodesic distance between paris of points, in order to reveal the essential structure of the data with high dimension. When the input data points are drawn from multiple manifolds, many manifold learning approaches suffer. In the case where the multiple manifolds are separated by a gap, Isomap may discover the different connected components in the local neighborhood graph.This paper presents a research of semi-supervised manifold learning in applications of video analysis to resolve the multi-manifolds learning problem. Firstly, an analysis on node-weighted MDS (nw-MDS) which can be used to solve the prior problem is presented, and then, based on nw-MDS, leading into the idea of semi-supervised learning--giving the part labels of submanifolds, an algorithm named semi-supervised Isomap based on nw-MDS is proposed. This algorithm can partition an input dataset into clusters where each cluster contains data points from a single, simple low dimensional manifold. Experiments demonstrated the usefulness and advantages of the algorithm in video analysis, it outperforms the original MDS and Isomap for seeking the intrinsic feature distribution and human action classification and recognition.