| Summary: | 碩士 === 華梵大學 === 資訊管理學系碩士班 === 99 === In the literature for computer vision and image processing, there are lots of well-known algorithms which are useful for practical applications. However, these algorithms are usually designed for running in sequential modes. Some sequential algorithms may be time-consuming and not suitable for real-time applications. In order to speed up the computation time, we adopt the multiple threads. If the algorithm can be run in multiple threads, the computation time can be reduced. In the thesis, we select to deal with the problem of point correspondence that is time-consuming and implement in the CUDA platform using multithread.
For parallel processing, an image is divided into a few image blocks with the same sizes. In the thesis, we propose an algorithm for point correspondence: we store the coordinates of image feature points in an array, and then equally assign these feature points into each thread for image matching. Since the size of an image block assigned for a thread and the size of the search area are important for point correspondence, we design some experiments and analysis for time computation.
In our experimental results, the computation in the experimental setup, which the number of threads is 5120 and the block size is 64x64 to each thread, performs better. When the image size is 640x480, the time used for point correspondence is 127.279ms/frame.
|
|---|
