Several GPU Accelerated 3D Face Analysis And Recognition Algorithms

With the maturities of 3D scan technology and improvement of the computability of CPU, there have been great progresses in 3D face analysis and recognition.3D face recognition not only has the advantage of friendly, non-intrusive and hard to forge but also almost unaffected by the variation of gesture and illumination which will terribly affect the performance of 2D face recognition algorithmes.But the high demand for computing power of 3D face analysis and recognition is one of its current limiting in its development and extensive application. Contemporary GPU with its powerful floating-point calculation ability, in the computer system not only plays the role of graphics processors but also begun to bear the burden of general-purpose computing. In many fields such as astronomy calculation, fluid mechanics simulation, molecular dynamics simulation, applications with the use of GPU acceleration won several times, dozens of times, even hundreds performance improvements. Explore the application of GPU general-purpose computing in 3D face analysis and recognition is the starting point of my work. Based on the introduction of the 3D face data, the research status of the 3D face analysis and recogniotn and the existing problems, reviews of the development of GPU general-purpose computing and the open computing standard OpenCL, the paper mainly completed the following jobs:(1)Analyzed the 3D face denoising algorithms spike noise removing, holes filling, Gaussian smoothing and bilateral mesh denoising. Then gave the GPU accelerated realization of these algorithms.(2)A 3D face nosetip detection algoritm based on double models matching which is robust to arbitrary posture of the 3D face is proposed and also the algorithm's GPU accelerated realizations(3)Analyzed the extensively used algorithm ICP in the field of 3D face analysis and recognition and its quickly but approximately realization. Then a GPU accelerated realization scheme was given. Finally A realization method of the k-d tree which is used for the strictly nearest neighbout search of spatial point sets for the GPU and a GPU parallel summation algorithm were proposed and therefore realized the GPU accelerated ICP algoritm.