激光与光电子学进展, 2021, 58 (2): 0210020, 网络出版: 2021-01-11
一种适用于行星表面特征提取的实时SIFT算法 下载: 875次
A Real-Time SIFT Algorithm for Planetary Surface Feature Extraction
图像处理 尺度不变特征变换算法 快速高斯模糊 CUDA 实时性 image processing scale invariant feature transform algorithm fast Gaussian blur CUDA real time
摘要
在行星探测任务中,针对尺度不变特征变换(SIFT)算法计算量大,无法同时满足对导航算法准确性和实时性要求的问题,提出了一种基于快速高斯模糊的并行化SIFT算法,即FG-SIFT算法。首先,将算法中构建高斯金字塔的二维高斯核函数分离成两个一维高斯函数,降低算法的计算复杂度。然后,对于每一维高斯函数,使用两个无限脉冲响应滤波器串联进行逼近,进一步减少计算量。最后,利用并行化处理的优势,设计算法各部分的并行化计算方案。仿真结果表明,FG-SIFT算法的计算效率相较于原SIFT算法平均提高了15倍,相较于没有使用快速高斯模糊的SIFT算法,在图形处理器上的运行效率也有近2倍的提高,很大程度上减少了特征点提取的计算时长,提高了算法的实时性。
Abstract
In order to solve the problem that the scale invariant feature transform (SIFT) has a large amount of calculation and cannot meet the requirements of accuracy and real-time in the navigation algorithm, a parallel SIFT algorithm FG-SIFT based on fast Gaussian blur is proposed. First, the two-dimensional Gaussian kernel function, which constructs the Gaussian pyramid, is separated into two one-dimensional Gaussian functions to reduce the computational complexity. Then, two infinite impulse response filters are used in series to approximate each one-dimensional Gaussian kernel function to further reduce the computational complexity. Finally, using the advantage of parallel processing, the parallel computing scheme of each part of the algorithm is designed. Simulation results show that the computational efficiency of FG-SIFT algorithm is 15 times higher than that of the original SIFT algorithm, and the running efficiency of FG-SIFT algorithm on graphics processing unit is nearly 2 times higher than that of SIFT without fast Gaussian blur. This algorithm greatly reduces the calculation time of feature point extraction and improves the real-time performance.
单宝彦, 朱振才, 张永合, 邱成波. 一种适用于行星表面特征提取的实时SIFT算法[J]. 激光与光电子学进展, 2021, 58(2): 0210020. Baoyan Shan, Zhencai Zhu, Yonghe Zhang, Chengbo Qiu. A Real-Time SIFT Algorithm for Planetary Surface Feature Extraction[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0210020.