Study And Implementation Of 3D Digitization Based On Photometric Information

Three dimensional data of objects and scenes contain more abundant visual information,which is of great value and broad application prospects in production and life.The three-dimensional reconstruction technology of computer vision has become a research hotspot because of its low cost,high efficiency,freedom from volume and space,It has been gradually applied to various 3D perception scenarios.The three-dimensional reconstruction technology based on photometric information has great advantages in restoring the surface details of the object,but recovering the three-dimensional shape of object from the single photometric information of the image often has the problem of depth distortion.At the same time,the ideal reconstruction are performed in a darkroom with multiple calibrated light sources,which greatly reduces the utility and robustness of the technique.In addition,in recent years,with the popular of depth cameras such as Microsoft Kinect,it is faster and simpler to directly acquire the three-dimensional information of the object through the sensing device,however the depth data directly collected contains too much noise and the detail loss is serious.It is difficult to apply in some scenarios.Based on these problems,the main work of this paper is fusion reconstruction,which use the photometric information of RGB image and the depth information directly acquired by the depth camera,so that a three-dimensional model of object with relatively high precision and rich details is obtained.Firstly,in the ideal darkroom environment,the three-dimensional reconstruction of the single perspective of the object is completed by combining the photometric information of the RGB image and the depth information directly acquired by the Kinect.Then,based on the improved illumination model,the RGBD information is integrated to realize the three-dimensional reconstruction of object under the global illumination.Finally,reconstructing from multiple perspectives,extracting point cloud data,and registering,Thereby obtaining a complete 3D model.The specific work is summarized as follows:(1)Three-dimensional digitization of partial differential photometric stereo based on RGBD fusion.In an ideal darkroom environment,a parallel white light source is first used to illuminate the measured object at a plurality of different direction,and a set of RGB image information and depth information of the surface illumination change of the object is acquired using a Kinect camera,and the light source direction for each exposure is calculated from the calibration sphere in the images.Through partial differential photometric stereo method based on RGBD fusion,the reflectivity is eliminated by the photometric ratio between the three channels of the RGB image,and a linear partial differential equation with the gradient as unknown is constructed.That is,the photometric constraint of the cost function,then construct the depth constraint term using the depth information directly acquired by the Kinect,and construct the Laplacian smoothing term of the surface of the object with the Laplacian operator,and finally obtain a cost function with the depth as unknown which consist of a partial differential photometric terms,depth terms,and Laplacian smoothing terms.The over-determined linear equations are solved by the normal equation of least squares estimation,and a three-dimensional information of object with relatively high precision and rich details is obtained.(2)Three-dimensional digitization of RGBD fusion under global illumination.The partial differential photometric stereo based on RGBD fusion can obtain more accurate three-dimensional information of the object,but the reconstruction environment is the ideal darkroom,and the illumination condition is the calibrated parallel light source.Such constraints reduce the robustness and practicability of the algorithm.Therefore,based on the partial differential photometric stereo,the global illumination model is used to represent the mixed illumination in the natural environment with the first-order spherical harmonic function,and the partial differential photometric constraint with global illumination is constructed to remove the limit of reconstruction environment and the parallel source calibration.combining the prior depth information to construct the depth constraint term,and solve the object depth by the cost function consist of each component,so as to realize the three-dimensional digitization of the object under arbitrary illumination conditions.(3)Three-dimensional digitization of full view.After completing the 2.5 dimensional reconstruction,in order to obtain the full-view three-dimensional information of the object,the RGBD fusion algorithm under global illumination is used to reconstruct the object in multiple perspectives,and the point cloud data is extracted,then the method combining NDT(Normal Distribution Transform)coarse registration and ICP(Iterative Closet Point)fine registration for splicing and fusing the multi-frame point cloud data.Firstly,considering the problem that the transformation angle between the different frame point clouds is large,the number of points is large,the ICP is easy to fall into the local optimum,that cause registration failure,the fast coarse registration of the point cloud is performed by NDT,and then the fine registration is performed by ICP.Finally,the Poisson surface reconstruction of the complete point cloud is carried out and the chromatic aberration is further corrected,and a three-dimensional model of the full-view object is obtained.Through the simulation in the ideal environment and real reconstruction experiments,the validity of the proposed algorithm is verified.At the same time,the correlation algorithm is used as a reference.The error comparison between the reconstruction results and the standard model further demonstrates the advantage of accuracy of the algorithm.The idea of algorithm is information fusion and variable transformation,through the surface photometric information of the object and the depth information directly collected to constrain the reconstruction,and finally obtain a more precise three-dimensional information of the object with complete surface details in any illumination environment,which have a certain practical value in realistic three-dimensional perception scenes.