Research On Depth Map Generation For Natural Three-dimensional Teleyision

Along with the development of the computer science, more and more technology needs the depth information to truly represent the real world, make observers feel right on the scene. Natural 3DTV uses "2D video plus depth plus 3D augmentation information" coding scheme, the accurate depth information is needed in advance. Therefore, depth extraction technology is a key factor that affects the properties of natural 3DTV. Research on the depth extraction technology is of far-reaching significance.Up to now, there are two main methods to acquire depth information. One is active stereo method that uses time-of-flight range sensor technique. The other is passive stereo method which obtains depth based on two or more 2D images. This paper compares their advantages and disadvantages, and focuses on the research of combining active and passive stereo methods to generate depth information. The main achievements are as follows:(1) A novel passive stereo algorithm Adaptive Window Aggregation and Dynamic Programming is proposed. Only based on the intensities of stereo image pairs, the algorithm chooses an appropriate support window adaptively for each pixel and optimizes the raw disparity map using DP algorithm. Compared with other stereo matching methods, this algorithm can produce accurate disparity maps more easily and efficiently.(2) A real-time depth generation system is constructed which contains the function of image acquisition, pre-processing and depth map obtaining. The whole system is accelerated by GPU parallel calculation technique. Based on the characteristics of CUDA platform, Adaptive Window Aggregation and Dynamic Programming algorithm is rewritten properly so that it can be suitable for parallel computing.(3) This paper proposes a simple and effective fusion method to generate high-quality depth map by utilizing the complementary characteristics of ToF camera and passive stereo methods. In order to integrate their respective advantages, we measure their reliabilities and construct a new aggregation cost volume. This algorithm improves the accuracy and robustness, the generated depth map is much better than that obtained from an individual method.