The Stereo Matching And De-interlacing Algorithm Based On CUDA

As an important passive depth measurement, Stereo vision is one of the research focuses in the field of computer vision. Stereo matching is a key step for stereo vision. The thesis investigates how to design a real-time stereo matching method. Among related literature concerning matching algorithms, local area based matching method has advantage of simple operations and hardware logic circuit implementation. The selection of matching window has a significant influence on the match result. Furthermore, proper acceleration method can be used to reduce computational time.In the typical surveillance scene, there are dynamic objects, like moving cars and pedestrians. The interlaced scan of surveillance cameras is the main factor for video quality deterioration. Thus De-interlacing algorithm is required to convert interlaced video into progressive scan format for future processing. Classical deinterlacing methods mainly includes motion adaptive deinterlacing and motion compensated deinterlacing. Both are analysized in the thesis. Motion adaptive method employs the same-parity 4-filed motion detection to extract motion information. Then the image is divided into three parts: static area, motive area and mixed area. The algorithm implements field merging in static area, enhances edge-dependent interpolation in motive area and performs motion vector based weighted average in mixed area. Motion-compensated deinterlacing uses bilateral motion estimation and then divides field blocks into three parts based on the motion vector. The algorithm implements enhanced edge-dependent interpolation for fast motion blocks and linear average along the motion trajectory for slow motion blocks. Other blocks are divided into smaller parts by using one of the two interpolation methods.CUDA is introduced by NVIDIA. It is a powerful general purpose parallel computation architecture with a software environment of GPU. It allows parallel processing with the help of the powerful GPU parallel computational capability. Based on CUDA, the thesis can accelerates processing time of stereo matching and deinterlacing algorithm effectively. It makes a comparison for the CUDA performance between different algorithm structures of stereo matching. Algorithm structure adjustment is carried out for deinterlacing algorithm. And a good result is obtained for motion detection within motion adaptive method and motion estimation within motion compensated method after CUDA acceleration.