Research on Optimization of SURF Algorithm Based on Embedded CUDA Platform

As the key technology of robot vision positioning, binocular stereo matching algorithm has become a hot area in the field of vision. In practical application, the existing binocular stereo matching algorithm has the disadvantage of poor portability, low real time and insufficient precision. In this paper, the SURF (Speeded up Robust Features) of image matching algorithm is optimized and implemented on embedded CUDA platform. The parallel computing power of GPU is used to accelerate the stereo matching algorithm. By analyzing the multi-scale feature points of SURF algorithm, we choose the epipolar constraint condition and the difference constraint condition as the conditional judgment to reduce the search scope to make it achieve the purpose of real-time. At the same time, to improve the speed of the algorithm without affecting the matching accuracy, we do Gauss filtering on the input images, and the SURF algorithm is improved by adding K-means and RANSAC algorithm that eliminates the mismatch points. The experimental results show that the improved matching algorithm on the embedded CUDA platform has better recognition accuracy, cross-platform portability and real-time, to meet the needs of the robot vision positioning.