Human Motion Estimation Based On The Probability Model Of The Latent Variable

With the development of the social modernization, image processing, machine learning and intelligent computing have become more important gradually, which are the branches of the technology of information and computer. In recent years, human motion estimation is focused by many domestic and foreign scholars, which is applied to many fields, such as the production of character animation, the production of 3D human film, medical diagnosis of human motion and so on.The research of human motion estimation originates from the processing of the object tracking, detection and occlusion of monitoring in the video. However, it is found that using the simple labels to track the object can not obtain the needed parameters of the object motion, and the human motion estimation just belongs to this research of the problem above. Human motion estimation and object tracking have the similar feature, it can be widely applied to benefit various aspects of our life, thus the research of the human motion estimation is needed to be analyzed more precisely in order to satify the higher requirement for the information of the human motion.In this dissertation, the research of the human motion estimation mainly starts with the probability model of the latent variable, meanwhile it is combined with the theory of the image processing and intelligent computing. There are two key problems in the research. Firstly, the probability model of the latent variable is used to learn the known high dimensional data samples of human motion in 3D. Therefore, the unknown human motion in 3D is estimated. Secondly, after processing the multi-view image of human motion to extract the silhouettes, the silhouettes is used to estimate the corresponding human motion in 3D. The work of this dissertation is summarized as follow:1. The data samples of the incomplete human motion cycle can not be learnt by GPDM to estimate the human motion. Thus a algorithm called the spatial constraints-based probability estimation can be used to solve this problem, in which the latent variable data of missing frames from incomplete human motion cycle in the latent space is calculated, so that the latent variable data can be used to estimate the corresponding missing poses(high dimensional data sample). Then, the latent variable data of incomplete human motion cycle and the corresponding poses(high dimensional data sample) can be complemented, which can be used to train GDPM again to achieve the human motion estimation.2. The learning algorithm based on the feature similarity optimization of the latent variable data is proposed to estimate the transition human motion of the two different human motions. This algorithm originates from the improvement of the learning algorithm of B-GPDM, it uses the distance and length of projection from the corresponding latent variables data to build the objective function, and the latent variables data which are initialized randomly can be optimized during the learning, which belonged to the transition motion. The optimization in the above learning is called feature similarity optimization(FSO). When the learning is finished, the transition human motion of the two different human motions can be estimated.3. The manifold latent pobabilistic optimization based on orthogonal subspace searching(MLPO-OSS) is proposed to achieve the better fitting of the high dimensional data from two different human motions. The high dimensional data samples contain the samples of imcomplete human motion cycle. MLPO-OSS can be used to improve the fitting of these high dimensional data samples, so that the human motion can be estimated better.4. An algorithm called dual latent variable spaces local particle search(DLVSLPS) is proposed to estimate the human motion in 3D from the silhouettes of the multi-view images more accurately, in which GPDM is used to build the dual latent variable spaces and the mapping from latent variable data to high dimensional data. In addition, the low dimensional particle is used to generate the better high dimensional particle through the neighbor weight prior condition search(NWPCS) in the dual latent spaces, so that the corresponding frame of the human motion in 3D can be estimated. DLVSLPS is more effective than the traditional particle filters.5. When a small amount of the high dimensional data samples are known, low dimensional space incremental learning(LDSIL) is proposed to estimate the human motion in 3D from the silhouettes of the multi-view images more accurately, in which stochastic extremum memory adaptive searching(SEMAS) and incremental probabilistic dimension reduction model(IPDRM) are used to estimate the human motion in 3D, and the corresponding new high dimensional data samples are collected. Then, IPDRM can be used to obtain the low dimensional data from the new collected high dimensional data samples through incremental dimension reduction, the new collected high dimensional data samples can be selected by comparing the distances among the corresponding low dimensional data, they can update the mapping from low dimensional space to high dimensional space to complete the incremental learning, so that the corresponding human motion in 3D can be estimated. LDSIL has better performance of the estimation than the previous algorithms.6. Gaussion incremental dimension reduction and manifold Boltzmann optimization(GIDRMBO) is proposed to enhance the accuracy of the estimation of human motion in 3D from the silhouettes of the multi-view images, in which the high dimensional data describing the human motion model is divided into two subvectors according to the infomation of spatial position and pose respectively, then the Gaussion incremental dimension reduction model(GIDRM) is used to obtain the low dimensional data of the samples of two subvectors through the dimension reduction and build the corresponding low dimensional space(latent variable space) and mapping repectively. Manifold Boltzmann optimization is also used to find the best subvector matching with the silhouettes of the multi-view images in the corresponding low dimensional space of the subvector, so that the final high dimensional data for the human motion model consists of these two subvectors for the estimation. GIDRMBO also has the better performance than the previous algorithms.It is shown from the results of our researches that using the probability model of the latent variable properly can be used to estimate the human motion in 3D better. With this model, the high dimensional data samples which describe the human motion in 3D can be learnt, and these data samples can be simplified, so that the features of the human motion can be extracted easily. More effectiveness is achieved by the intelligent computing based on this model.