Auto-Regressive Model Based Depth Recovery Using CUDA

Acquiring scene depth is one of the basic challenge in computer vision. Its application covers robot navigation, model reconstruction and human-computer interaction and so on. The quality of depth image acquired by TOF camera or Kinect camera is unsatisfactory,which needs excellent depth recovery algorithm. Auto-regressive model based depth recovery algorithm recovers depth image acquired by 3d-TOF camera preeminently. It difficult to work on real-time system because it needs mass of calculation for parameters of auto-regressive model.Graphics processor unit(GPU) are born with the powerful computing ability. CUDA makes the development of parallel computation easy, therefore Programmers can develop parallel programs easily by using GPU. GPU has been widely used in the field of non-graphics and general purpose computing is researched in-depth based on GPU. The rapid development of GPU promotes development of application field such as image processing, virtual reality, computer simulation; and provides a good platform for general purpose computing. Image processing base on CUDA and general purpose computing has become the focus of research in graphics and high performance computing.This paper researches on parallel computation design and CUDA, and implements a general computational framework for depth image recovery algorithm. Input of auto-regressive model of depth recovery algorithm are depth image of low quality and color image of high quality. To compensate for the depth information using the texture information of color image, recover the depth image of high quality. The auto-regressive model contains both depth and color information, therefore the calculation is very time-consuming because of a great deal of matrix operation. Study on operation mode of CUDA hardware,and software configuration in this paper, implemented a parallel version of the auto-regressive model of depth recovery algorithm. This algorithm in CUDA is divided into six steps, each step has been optimized. Finally, through an extensive experimental comparison, the parallel version of the algorithm has excellent and stable performance.