CAJ | 학술논문

针对平面单应性矩阵实时求解的需求，提出了RANSAC单应性算法在GPU上并行实现的方法。首先计算出理论上最优的迭代次数；然后根据迭代间数据的独立性将每次迭代看成一个独立的任务映射到CUDA线程上。实验表明，对于较高分辨率的图像，GT240M平台上平均加速比达到了8。
침대평면단응성구진실시구해적수구，제출료RANSAC단응성산법재GPU상병행실현적방법。수선계산출이론상최우적질대차수；연후근거질대간수거적독립성장매차질대간성일개독립적임무영사도CUDA선정상。실험표명，대우교고분변솔적도상，GT240M평태상평균가속비체도료8。
A massively parallel RANSAC algorithm implementation in GPU is proposed to meet the real-time requirements for homography estimation. At first, optimal iteration number is calculated. Then, each iteration task is assigned to a CUDA thread based on data independent between iterations. Experimental results on a GT 240M graphics card prove that the RANSAC homography algorithm could be used to deliver a speed-up of 8x with high resulution images on average.