激光与光电子学进展, 2019, 56 (14): 141502, 网络出版: 2019-07-12
改进的尺度不变特征变换算法并行加速双目测距系统及其实现 下载: 1094次
Research and Implementation of Binocular Distance Measurement System Based on Improved Scale-Invariant Feature Transform Algorithm with Parallel Acceleration
机器视觉 双目测距 尺度不变特征变换算法 开放运算语言 并行加速 machine vision binocular distance measurement scale-invariant feature transform algorithm open computing language parallel acceleration
摘要
在数字图像处理领域中,尺度不变特征变换(SIFT)算法是特征点识别的代表性算法。以开放运算语言(OpenCL)并行计算为加速手段,建立了基于改进的SIFT算法的双目测距系统,深入研究了如何加快SIFT算法的运算速度。在加快SIFT算法方面,选取了积分均值模糊,并利用OpenCL对其进行并行加速,对算法进行并行优化后,使之能够在NVIDIA GPU硬件平台上进行实现。在获取精确视差方面,对原SIFT匹配方法进行了改进,极大地提高了匹配效率。此外,构建了双目测距异构计算实验平台,并进行实验。实验平台对采集的图像进行了实时处理,验证了基于SIFT算法的并行加速计算的可行性,同时可以直观地看到中间计算过程和测距的结果。实验结果表明,本文方法与已有的加速优化工作相比,计算时间消耗比原方案要少得多。
Abstract
The scale-invariant feature transform (SIFT) algorithm is the representative approach for key-point detection in the field of digital image processing. A binocular distance measurement system is established herein based on the improved SIFT algorithm by using open computing language (OpenCL) parallel computing as an acceleration method, and a profound study on how to speed up the operation of SIFT algorithm is performed. First, the integral mean blur is selected to speed up the SIFT algorithm operation. OpenCL parallel computing is then used to accelerate it. The parallel optimization of the algorithm is made to be implemented on NVIDIA GPU hardware platforms. The original SIFT matching method is improved to obtain an accurate parallax. Consequently, the matching efficiency has been greatly improved. Finally, a binocular distance measurement system heterogeneous computing experimental platform is constructed. The experimental platform performs a real-time processing on the acquired images. The feasibility of parallel acceleration based on the SIFT algorithm is verified. An intermediate calculation process and the distance measurement results can be directly obtained in the system. The experimental results show that compared with the previous accelerated optimization work to the SIFT, the computational time consumption of the proposed approach is much less than that in the original method.
张志强, 施文华. 改进的尺度不变特征变换算法并行加速双目测距系统及其实现[J]. 激光与光电子学进展, 2019, 56(14): 141502. Zhiqiang Zhang, Wenhua Shi. Research and Implementation of Binocular Distance Measurement System Based on Improved Scale-Invariant Feature Transform Algorithm with Parallel Acceleration[J]. Laser & Optoelectronics Progress, 2019, 56(14): 141502.