Three Dimensional Human Motion Analysis From Uncalibrated Monocular Image Sequences

Three dimensional human motion analysis studies how to recover relative three dimensional information of human from image sequences, such as locations, motions and gestures, which plays an important roles in the wide applications, such as pose recognition, semantic analysis, behavior understanding, virtual reality, intelligent sur-veillance, human-computer interaction and motion analysis.In recent years, visual analysis of human motion received increasing attention from academy and industry. There are two major categories of research methods on visual analysis of three dimensional human motion analyses: model-based approaches and learning based approaches. Nevertheless, there are many theoretical and technical problems remaining open in this research field. For example, seldom researches pay attention to the analytical relationship between the projection of human model and its feature of image. Furthermore, the initial joint points are still set manually in the first frame during the process of estimation and tracking. Besides, although many advanced filtering methods have been proposed for human tracking, the transition model of states is absent giving rise to possible miss-tracking. In machine learning, suitable algorithms for human analysis are badly demanded.The research work of this thesis involves reconstruction of human model, feature subtraction, filtering algorithm and machine learning. The main contributions of this thesis are follows:1. This thesis proved two theorems of convolution surface projection. With this, a novel method for reconstruction of human model and pose estimation is presented. Aiming at the absence of analytical relationship between the projection of human model and the image feature, convolution surface is introduced to reconstruction human model by integration along the articulated human skeleton. This thesis has proven two theorems of projection relationship between the convolution surface and convolution curve. Therefore, the silhouette of human in image is modeled by convolution curve and the parameters of human motion are estimated by nonline-arly fitting the human contour.2. An automatic method finding the location of joint points is proposed. Aiming at overcoming the problem of joint locations initialization, this thesis proposes a automatic method to automatically initialize joint points and prediction of unde-tected points. These joint points lie on the skeleton line of a thinned region and pa-rameters are estimated by minimizing the distance between the detected joint points and the projection of three dimensional model.3. A two-layered filtering framework of human tracking is presented. Aiming at the absence of state transition model, the thesis gives a two-layered framework for tracking. In the first layer, joint points are tracked by Kalman filter to avoid unde-tected points. In the second layer, the interactive multiple models are introduced, in which the multiple models are trained by ridge regression from real motion capture data.4. Shared latent variable dynamical model for human tracking is proposed in this the-sis. In order that the state can be tracked in a low dimensional space, the thesis proposed shared latent variable dynamical model to get a shared low dimensional latent variables. During the offline training, the state equation, observation equa-tion and reconstruction equation are acquired. During the online tracking, the low dimensional variables are tracked using Condensation algorithm. At the same time, the real states of human motion are calculated by reconstruction equation.5. A closed form solution of shared latent structure is derived. We have proven that the Principal Component Analysis is equivalent to the shared latent model. Aiming at the lack of a closed -form solution for the shared latent model, this thesis gives the proof that the Principal Component Analysis is equivalent to the shared latent model. Furthermore, ridge regression is introduced into the framework of calculat-ing shared latent model.