Human Motion Recognition Based On NPE And LDCRF

Since the birth of computer, people have dreamed of understanding and finding the world better by computers. In recent years, with the rapid developments of computer technology and other related subjects, visual analysis of human motion as an important direction of computer vision gets more and more people's attention. Human motion analysis mainly investigates four aspects which are detecting, tracking, recognizing people and understanding human behaviors from the image containing humans, it involves artificial intelligence, image processing, pattern recognition, statistics and other complex subjects. Currently, moving human detection and tracking technology have been much riper and have very good applications; however, human motion recognition as a senior processing part of human motion analysis has been a hotspot for experts and scholars. The purpose of human motion recognition is detecting the moving people from continuous image sequences, extracting human body segments and then getting motion feature, at last identifying the motion as a particular type through some classification mechanism. Human motion recognition has a very broad application prospects and practical value, such as in national defense, military, industrial, and even our daily life. Currently, human motion recognition has been widely used in human-computer interaction, intelligent control, motion analysis and virtual reality, and it looks forward to having some further improvements.In this paper, we deeply investigate several commonly used dimension reduction methods and human recognition models. On the basis of reading and studying a large number of excellent papers, we present a human motion recognition method based on a manifold learning, that is Neighborhood Preserving Embedding(NPE) and Latent-Dynamic Conditional Random Fields(LDCRF). and the experiment results proved that this method has a high recognition rate.Firstly, we analyze several commonly used methods of moving objects detection in details, which are temporal difference, background subtraction and optical flow, and explain their advantages, disadvantages and their application occasions by discussing their principles. To the characteristics of the data we used, considering the real-time, segment results and operational complexity, we use background subtraction method to get binary images containing people, and then get the better foreground images for the further work by dilating and eroding.Secondly, human activities are non-rigid, how can we represent the specific human motion information as abstract parameters and preserve them well is a problem of human motion recognition, also human motion feature as the input of recognition model is very important to the recognition result. Therefore, this paper introduces the extraction and representation of human motion feature and mainly introduces model based and non-model based methods, and some other classification method. After contrasting and analyzing these methods, we present our method, that is, on the basis of extracting human contours, we re-sample on the contour, distribute the sampling points uniformly, and then calculate the distance to the centroid clockwise and get N-dimension feature vectors as the model input.Thirdly, the real world data are usually high-dimension, how can we find the data internal structure and correlation, and reduce the redundancy and noise is a problem. That is given rise to the conception of dimension reduction. Dimension reduction techniques are generally divided into linear and nonlinear. The common methods are Principal Component Analysis(PCA), Isometric Feature Mapping(Isomap), Locally Linear Embedding(LLE). and recent year it proposed a Locality Preserving Projections(LPP) and Neighborhood Preserving Embedding(NPE)methods. In this paper, we introduce these five common methods, and compare them by based on global or local features, linear or nonlinear, convex function optimization, parametric or nonparametric, for artificial data or real data, whether generates mapping to test data. PCA is based on linear hypothesis, and getting the optimal subspace through covariance matrix, aiming at preserving the global Euclidean distance. Isomap and LLE are nonlinear, although they can get good results in some synthetic data, the computation is very complex. And they can only generate mapping in training set, the mapping for the test set is unknown. On the contrary, LPP and NPE preserve the local neighborhood structure of the data manifold, and NPE is not sensitive to outlier as PCA and the computational time is shorter than methods based on global features. It can generate mapping in both training set and test set. overcoming the lack of Isomap and LLE. In this paper, we use NPE to reduce the feature vector. Since choice of the parameter k in the method is not clear, we do some experiments on it. And we also carry out some experiments on the low-dimensional embedding dimension. So we have plenty of data for the further work.Finally, the current methods for human motion recognition are usually based on Hidden Markov Model(HMMs) and Support Vector Machine(SVM). A new statistical machine learning method-Conditional Random Fields(CRFs) has been proposed in the sequence labeling task(especially natural language processing) and has a great success. And it has been employed to the field of human motion recognition. We use the LDCRF in this paper, which is an improvement of CRF. It avoids the independence assumption of the probability model, and it can use the context feature more flexible, and combine the internal structure and extrinsic dynamics by adding a layer of hidden state variables between observations and labels, which makes the classification more accurate. We discuss the principle of LDCRF and contrast it with HMMs and CRF. And we also test different window sizes of this method and contrast the recognition results with CRF. The experimental results show that the LDCRF is more stable and has a high recognition rate.In summary, this paper proposes a human motion recognition method based on NPE and LDCRF. Firstly, we introduce several methods for moving objects detection and contrast them, and use background subtraction to get binary images. After dilating and eroding, we get better foreground images. Then, we describe the representation of human motion feature. After getting the human contours, we calculate the human centroid and distribute some sampling points on the contours, calculate the distance to the centroid. getting the N-dimension feature vector. After that, we use NPE for dimension reduction. At last, we use a novel model LDCRF for recognition and contract it with CRF by different window sizes. The experimental results show that this method has a high recognition rate and a good stability. With the development of science and technology and the improvement of computer performance, the study of human motion recognition will be much riper and will be gradually applied into the real life. It is hoped that the work in this paper can make some contribution for the further investigation on this topic.