Research On Optical Image's 3D Reconstruction Based On GPU Acceleration And Polynomial Mapping

The optical image's 3D reconstruction is widely used in the fields of industrial design,medical image construction,3D image construction and image recognition.It has important research value and social practical value.In the process of 3D reconstruction of optical images,the camera calibration technique is one of the key points.The result is reflected in the mapping from the pixel coordinates of the two-dimensional image to the three-dimensional space coordinate,which is,the solution of the camera parameter matrix and the distortion coefficient,and the optical fringe center sitting in the sub-pixel precision is the base of three-dimensional reconstruction.Based on this,the extraction of optical stripe center and the mapping relation from pixel coordinates to 3D spatial coordinates are deeply studied in this paper.(1)This thesis introduced the method of extracting light stripe center,especially the Steger method.The Steger algorithm has high accuracy,stability and robustness,but the computation is very large and the operation time is long because of at least 5 times of two-dimensional Gauss convolution on every pixel on the image.The collateral computing advantage of GPU multi-grids,blocks and threads is combined with the independence of the convolution operation of each pixel point of Steger algorithm.When extracting the center of the light stripe,the related data is first copied from CPU memory to GPU memory,and a block is set for each pixel point.This block sets the number of thread,uses thread to compute 5 times of Gauss convolution template,and block calculates the Hessian matrix value which preprocesses of the pixel points required for the 5 times of the pixel convolution,and so on.The work every block and every thread in blocks have done is repeated,and the result is displayed in GPU.It is copied to CPU memory to get the center value of light stripe.The calculation method of the center coordinates of the light stripe can greatly improve the operation efficiency.In the experimental environment of GTX1060-6G graphics memory graphics card,four-core i5-4590 CPU and four-core AMD A6-7310 CPU,the programming of center coordinates extraction time of light stripe based on CPU and GPU is calculated.The experimental results show that the extraction time of the light stripe under GPU is one sixteenth to one twelfth of that of CPU,which greatly improves the computation efficiency and reduces the calculation time.(2)The traditional 3D reconstruction method needs to solve 4 internal parameters in the internal parameter matrix and 6 external parameters in the external parameter matrix,and also takes the calculation of the distortion coefficient into consideration under the radial distortion,the tangent distortion and the thin prism distortion.This method not only needs to consider the solution of the parameter matrix and the distortion correction,but also needs to use the calibration plate to each demarcating point,which is analyzed and modeled.Any errors occurring in the two processes will bring great errors to the calibration results.In this paper,a polynomial method is used which solves the related problems in the optical image processing,and the polynomial method is used to solve the mapping relationship between the two coordinate systems to each other which is simple and visualized.The polynomial mapping relation from the pixel coordinates to the three-dimensional space coordinates is derived and established.Polynomial calculation,instead of parameter equation and distortion coefficient calculation,simplifies the calculation process,reduces computation and improves accuracy.Using the experimental results and the hardware and software environment,the calibrated object with the height of Zw is 38,48,54,68,82,93 respectively.The experimental verification is carried out in the development environment of Python Wing IDE.The corresponding polynomial coefficients are solved.The average error of the Yw and Zw is compared with the actual measured values of the Yw and Zw of the calibrated object at different Zw heights.About 0.5058%,the average error of Zw is about 1.176%,which meets the requirement of error accuracy and greatly decreases the error rate.