Study On 3d Reconstruction Of Dynamic Non-rigid Object Based On Kinect

The 3D digitalization technology of non-rigid object,such as clothing,has a broad application prospect in the field of virtual fitting system,animation production and game development.This technology has been a hot topic in computer vision and computer graphics.How to achieve high efficiency,high accuracy and cheap equipments has always been a focus in 3D digitalization technology.The traditional 3D digitalization technologies can well reconstruct the rigid object.But it is difficult to find a 3D digitalization method with high accuracy,rich texture information for non-rigid object.In recent years,the depth camera with range-ranging capability,which is represented by Kinect,can directly capture the depth information of the photo area,but the general depth camera measurement accuracy is low,too much noise,and the measurement results must be de-noised and optimized for 3D digitalization.This paper aims at optimizing the modeling results of depth camera,and studies the 3D digitalization technology of non-rigid object based on Kinect,which can be summarized as follows:1)3D digitalization of non-rigid object by fusing photometric stereo method with Kinect.Based on the color photometric stereo method and RGB-D camera,the high-precision 3D digitalization method of dynamic single reflectance/multireflectance objects is discussed.We use the red,green and blue light source to illuminate the object at the same time.The depth and color graph of the target is collected by the Kinect,and the construction result is solved by solving the cost function.The reconstruction result is more accurate than the traditional photometric method because of the depth dimension information.2)Full view 3D reconstruction with the rotating platform based on Kinect.After the single view digitization,we put the object on the rotating platform to realize the full view 3D digitization.The object is rotated on the rotating platform,and a Kinect camera is used to photograph it,taking its rotating weekly graph sequence.We use SURF operator to calculate the feature points between two frames,the coarse registration between adjacent frames is completed based on the feature point.After that the ICP registration technique was used to perform the accurate registration and error distribution method is used to modify the drift error.Then the Poisson reconstruction technique is used to fuse the point cloud and the color difference is corrected.Finally,a full view model with high precision can be reconstructed.3)Human body dynamic reconstruction based on S-SCAPE.Because of the non rigid bodies such as human body not only contain the whole displacement,but also have some local deformation,the non rigid point cloud registration has been a difficult problem in 3D reconstruction.This paper uses the human semantic model S-SCAPE to solve the registration problem.S-SCAPE uses 27 body parameters and 32 pose parameters to represent the human body model,which greatly reduces the solution space that the optimization algorithm needs to search.The acquisition system does not need to be calibrated in advance.The reconstruction method only need three frames of RGB-D datas to perform a complete human model,and has broad application prospects.The experimental results showed that the proposed method can realize the 3D digitalization of non-rigid object and can meet the specific needs of some applications.The research of this paper has laid a good foundation for realize nonrigid dynamic object reconstruction.