Research On Improvement Of Center Extraction Algorithm Using CUDA

With the rapid development of computer and machine vision technology,the method of visual measurement has been widely used in industrial manufacturing,scientific research and medical testing.The visual measurement technology uses the sensor to get the images,and then makes the images as a mean or carrier for detecting and transmitting information,so as to obtain the precise mapping relationship between two-dimensional plane points and three-dimensional space points to achieve the measurement.The three-dimensional measurement technology for moving objects belongs to the category of visual measurement.In the process of dynamic three-dimensional measurement,it is usually necessary to paste some circular markers on the target to improve the accuracy and robustness of the measurement.The center extraction process of the markers on two-dimensional images has become the core of the three-dimensional measurement technology,which directly determines the accuracy and processing efficiency of the measurement.However,most of the existing center extraction algorithms are based on the CPU,and its computational efficiency is low,which can not meet the demand for real-time online measurement in engineering applications.In view of this,this paper proposes a CUDA-based center extraction algorithm.This algorithm makes full use of the characteristics of GPU's multithreading and external memory,and optimizes the traditional algorithm CPU's serial execution instructions in parallel,which greatly improves the computing efficiency and achieves a fast and accurate extraction process of the center coordinates to meet the requirements of real-time online measurement.Specific research work is as follows:Firstly,the principle of the traditional center extraction algorithm is analyzed.The algorithm mainly includes three key technologies: elliptic boundary annular neighborhood extraction based on image,binary image connected-component labeling and least squares solution.The binary image is obtained by extracting the elliptical region in the image,and then the information of the pixel on the ellipse is calculated by labeling the connected component of the binary image,and the coordinates of the circle are obtained by substituting into the elliptical least squares solution target equation.The whole process lays the foundation for subsequent parallel analysis.Secondly,a center extraction algorithm using CUDA is designed.Among them,the computing power of GPU and the programming module of CUDA have been analyzed,and the advantages and programming mode of GPU multi-threaded parallel computing have been summarized.The parallelizable analysis of each link in the process of center extraction is carried out.Combined with the characteristics of the programming module of CUDA,the improved algorithm of center extraction using CUDA has been realized by programming on JETSON TX2,embedded development board of NVIDIA.The experimental program is verified by experiments.The intermediate results of each stage of the algorithm are compared with the existing CPU serial algorithm.The correctness and execution efficiency of the algorithm are analyzed.The experimental results show that for the black-and-white image with a resolution of 2592×1944,the parallel algorithm designed in this paper guarantees the accuracy of the extraction of the center coordinates,and the acceleration ratio is nearly 7.4 times as high as that of the existing CPU algorithm.On the basis of the above,a set of monocular vision system was designed to be applied to the automatic docking process of large cylindrical structures.The principle of the application is analyzed,which embodies the important position of the center extraction process in the measurement process.The center extraction algorithm using CUDA implemented in this project is used to accelerate the overall measurement processing efficiency.By the test results,we can know that the algorithm studied in this paper owns good precision and high execution efficiency in practical applications,and meets the requirements of automatic docking,which fully reflects the important practical significance of the CUDA-based center extraction algorithm.