Research On Parallel Algorithm Of Optical3D Reconstruction Based On Cuda

The dynamic three-dimensional surface shape of objects can be recovered withone deformation fringe pattern in Fourier Transform Profilometry,Hilbert TransformProfilometry and Wavelet Transform Profilometry.However,it is difficult to achievethe real-time measurement of dynamic objects that are larger with CPU and DSPserial processing that is slow and low efficiency. So the key of improving the speed ofthe three-dimensional measurement is how to minimize the calculation time.Fortunately the phase computation based on Fourier Transform, HilbertTransform and Wavelet Transform is not dependence on the neighborhood pixels withhigh arithmetic intensity, which suitable that the same program is executed onmultiple threads in parallel. This allows the phase computation done on the GPU withthe help of the powerful unified architecture graphics processor parallel computationalcapability. This paper analyzes the characteristics of the phase computation, andimplements the phase computation based on Fourier Transform, Hilbert Transformand Wavelet Transform with the advantages of CUDA.So the computation with higharithmetic intensity and parallelizable was done on GPU to allow the CPU with GPUworking together, then optimizes the parallel speed with GPU shared memory andsynchronization technology with Wavelet Transform to solve the problem of the lowefficiency of the phase computation on CPU.Besides,the processing speed andaccuracy of the three algorithms with CPU and GPU were analyzedrespectively.Experiments show that it obtains the speedup of one to two orders ofmagnitude to CPU by GPU acceleration with the same quality of the phasecalculation.So it lays the solid foundation for the Fourier TransformProfilometry,Hilbert Transform Profilometry and Wavelet Transform Profilometryapplying to the real-time three-dimensional measurement.