Research And Implementation Of GPU Based Fast 3D Human Face Imaging

For decades,3D real-face reconstruction is one of the hotspots in the field of computer vision.With the development of digital 3D reconstruction technology and the corresponding industry,3D human face model has been used as an important data in different community.At the meantime,the increasing amount of data to be processed and the parallelism requirements of modern computer vision algorithms have greatly burdened traditional serial processors,e.g.CPU,and revealed their inadequate.Graphic Process Unit(GPU)is different from CPU,it is designed for processing large-scale parallel computing efficiently.Nowadays,for well-designed parallel algorithms,the computing speed of GPU is faster more than 20 times compared with CPU.This thesis proposed a GPU based fast 3D real-face reconstruction system.The system consists of hardware and software.The hardware component includes face capturing platform(FCP),as slave computer,and algorithm implementing platform(AIP),as master computer.The FCP consists of main structure,capturing unit,light unit and synchronization unit,and is used for acquiring high-quality human face image.The AIP employed “CPU+GPU” framework which meets the need of both logical controlling and accelerating parallel computation of algorithms.In order to fastly reconstructing 3D point cloud data of human face,this thesis proposed a GPU based local stereo matching method.This method employed Normalized Cross-Correlation(NCC)as a criterion for measuring similarities between feature points,and three constraints were used for checking the validation of calculated disparity.A four-layer image pyramid was built for each stereo image pair respectively,and a layer-wise iterative refinement method was adopted for improving the generated disparity map of each image layer.This strategy overcomes the weakness of local stereo matching that generates coarse disparity map,and instead generates a dense one,meanwhile accelerates the matching speed.The whole method was implemented in GPU,and the experiments showed that the method achieved an acceleration ratio of at least 50 on the GPU,compared with the 4-threads CPU implementation.This thesis also studied and analyzed the underlying impacts that the capturing unit might be suffered during its life cycle;and proposed an online calibration method for 3D human face reconstruction.The method employed the result of offline calibration as a prior,and updated calibration parameters according a criterion that keeping the optimized stereo configuration near the offline computed version.The author studied the characteristic of human facial skin,and proposed an efficient and robust method for extracting valid correspondences on facial skin.The experiment showed that the proposed online calibration method was robust under slight lens drifting condition.For those conditions that might greatly change the stereo configuration,e.g.the translation and the rotation between stereo cameras,we still have to re-calibrate the stereo system using offline calibration method.