GPU-Accelerated Image Matching for Three-Dimensional Features

• Main Authors: Shih-Ying Hung, 洪世穎
• Other Authors: 蔡榮得
• Format: Others
• Language: zh-TW
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/86583465176067905511

碩士 === 國立中興大學 === 土木工程學系所 === 102 === The aerial images are getting larger than ever due to the emerging development in aerial sensor technology. It is time consuming in follow-up image processing using traditional central processing unit (CPU) approaches. Thanks to the recent development in graphics processing unit (GPU), which has much faster speed than CPU in computational capability and memory transfer mechanism, and related software development techniques, it is much easier than before to develop high-performance processing programs with the use of GPUs for aerial images. This research employs GPU technology to speed up SIFT approach in feature extraction from stereo pairs of aerial images. The solution of conjugate image points is also examined and evaluated by applying two-dimensional (2-D) affine transformation and RANSAC approaches for final computation of three-dimensional (3-D) coordinates in object space using space intersection. The performance in each step is compared for GPU and CPU approaches. The results of feature extraction from experimental images show that the speed-up performance is about 20 to 50 and more than 95% time saving for GPU approach versus CPU approach. The 2-D affine transformation is better than RANSAC approach for producing much expected conjugate image points with uniform distribution. However, it is time consuming when operator’s interference was introduced.