Research Of Human 3D Reconstruction Methods Based On Monocular Video

Human motion capture and 3D reconstruction based on vision is an important research area in the computer vision , and also is a challenging research area in the engineering and science . Generally speaking , there are two types of methods for the human 3D reconstruction : Generative and Discriminative . This paper do some research on the 3D reconstruction in the framework of discriminative methods .Firstly , this paper analysis several methods of human 3D reconstruction , and the research status at home and abroad , at last we decide to use discriminative reconstruction method based on database as the framework of this paper . Then we discuss the technology roadmap of human motion capture and introduce the preparation for the 3D reconstruction based on database , including the choice of human model , the pretreatment of standard motion capture data and the motion segmentation .Secondly , this paper tentatively uses human silhouette information as the image match input to retrieve the similar candidates in the database . So we compare the different human silhouette features, these feature expressions include Hu moment invariants , affine moment invariants , contour signature , Fourier descriptor and shape context . At last this paper choose the combined features based on Fourier descriptor and shape context to descript the human silhouette and search data in the database , the experimental results also show subsequently .Finally , the paper tentatively use human joints position information to retrieve the similar image in the database . This paper use the existing study in our lab to calibrate the human key joints . The problem of posture sequence optimization is also discussed in this paper , the resolve of this problem can keep the 3D human posture coherent in the human vision .As showed in the experimental results , the image march algorithm based on combined features and human joint position information proposed in this paper can retrieve similar candidates in the database , and the strategy of posture sequence optimization can keep the human 3D coherent in the space of 3D posture .