Research On Pose Estimation And Identity Recognition From Human Motion Sequence

With the rapid development of modern society,human motion analysis has attracted many researches' attention in many fields.Human motion analysis aims to automatically reconstruct human motion,perceive and understand of human behavior and identity on the semantic level.Compare to traditional appearance features,3D human skeleton features are more robust to external light and visual angle changes,and can accurately describe and explain human motion process from the aspects of biology,physics and human kinematics.Based on these reasons,this thesis focuses on two important issues of human motion analysis: 3D human motion estimation and individual identification.The main contributions of our studies are summarized as follows:Since it is difficult to obtain 3D skeleton directly,this paper proposes a 3D human pose estimation method based on Riemannian manifold.Based on the analysis of kinematics and geometric characteristics of motion trajectory,a second-order stochastic dynamic model is built according to tangent bundles of Riemannian manifold.We propose Extended Rauch Tung Striebel Smoother(RERTSS)method to estimate the3 D skeleton pose by generalizing the RTSS to Riemannian manifold.Moreover,further optimization of the reconstruction results can be done by the local simplex optimization method.The proposed method can accurately estimate the 3D human pose without training data,and also can alleviate the ambiguity problem in 3D reconstruction.Euclidean feature is unable to describe non-linear human motion accurately,we propose motion feature extraction method based on Riemannian manifold,which can accurately describe individual differences in non-linear space.In order to overcome the limitation of traditional metric learning methods which neglect the spatio-temporal structure of motion feature,we propose Spatio-Temporal Large Margin Nearest Neighbor(ST-LMNN),which combines bilinear model with classical metric learning framework to measure the similarity between individual motion features by the bilinear metric function.Considering the differences of spatio-temporal structure between individuals,and inspired by the nearest class mean classifier,we propose Spatio-Temporal MultiMetric Learning(STMM)method.In this method,samples belonging to the same class tend to be closer to its geometrical mean,while increasing the distance between the means with different labels.In addition,considering small scale and less covariant factors in most ID dataset,we build a motion database containing multiple external and emotional factors.Considering covariant factors influences on human motion process,we propose Spatio-Temporal Multi-Factor Discriminant Analysis(ST-MFDA)to alleviate the low recognition accuracy caused by multi-covariant factors.In ST-MFDA,a pair of spatiotemporal projection matrices are learned for each covariant factor to project motion features from different factors into a common subspace,and keep small intra-class divergence and large inter-class divergence in the common subspace via the generalized Fisher discriminant criterion.In summary,this article is developed in following three key aspects: 3D pose estimation,spatio-temporal metric learning and subspace learning,and proposes a series of modeling methods to solve the key sub-problems in human motion analysis.Both of the theoretical analysis and the experimental results show that the proposed pose estimation,spatio-temporal metric learning and spatio-temporal multi-factor discriminant methods are more effective than traditional methods.