Research On Image Feature Point Detection Method Based On GPU

Videogrammertric measurement(VM)is a method of measuring the coordinates and shape of the measured objects by sticking artificial feature points on the measured object(generally using retro-reflective marker points)and then collecting its images.VM only needs simple equipment such as cameras,marking points,and light sources to realize noncontact measurements of three-dimensional coordinates.Typical commercial VM systems,such as the US "V-STARS" and the domestic "XJTUDP measurement system",need to collect a large number of multi-view/large overlapping feature point images,but it is very time-consuming to calculate 3D coordinates based on a large number of feature point images.This has become a technical bottleneck restricting the real-time measurement of VM.Given that GPU has a large number of stream processors,it has much stronger data-parallel computing capability and more convenient programming environment than FPGA.GPU has become the mainstream platform for hardware acceleration of images real-time detection.Therefore,this paper focuses on the GPU-based image-feature point detection method to carry out research:1)A hardware-parallel architecture and timing control method for image feature point detection is proposed.Using the time interval of image acquisition by high-speed cameras,the functions of image data transmission,image feature point detection,and deformation measurement are decomposed,time-analyzed,and optimally designed,which enables each decomposed or combined functional module to complete the corresponding work within the collection interval.From this,the timing control constraint formula is derived.The timing control simulation example shows the relationship between the number of images transmitted from the acquisition card to the memory at one time,the number of images packaged to the GPU for processing at one time,and the delay time.2)Aiming at the problems of repeated searches and excessive time consumption in the search for contours in the feature point search domain,a fast detection method for imagefeature point search domain segmentation is proposed.Firstly,the search domain is divided into four equally divided blocks,and then the contours are searched for the four blocks in the four directions of up,down,left,and right at the same time,and finally the pixels are spliced into a complete outline of feature points.Accordingly,a temporal prediction model for image feature point detection based on GPU is created.The test results show that this method is 3.8times higher than the existing detection method,and the obtained coordinate value is the same as the data accuracy of the original method.3)A feature point real-time detection algorithm for image direct transmission to GPU is proposed,and the parallel architecture and timing control method of image direct transmission to GPU is given.In this method,the feature point image is distributed to each GPU through a load-balancing control algorithm to complete image feature point detection.Based on this,the timing control constraint formula of image-direct transmission to GPU is deduced.The timing control simulation example shows that: compared with the hardware architecture of the image first entering the memory and then to the GPU,the delay time of the real-time detection algorithm of the feature points of the image directly transmitted to the GPU can be shortened by 24.98% to 44.43%.4)Design and develop a real-time detection system based on GPU image feature points.The system has the functions of deformation calculation and data storage and performs an image-feature point detection algorithm on GPU.The relative error of the experimental results of wing vibration frequency measurement compared with commercial software Lammps is only 1.3%.The error between the experimentally measured acceleration and the theoretical value of the uniformly accelerated moving car is 5.97%.It shows that the featurepoint tracking detection method in this paper is correct.The experimental results of real-time detection of feature points show that: based on the hardware of AMD-7302,PCIE4.0,RTX3090,and EVT2100 cameras,the detection rate of feature points in this paper meets the acquisition rate of 10.58GB/s of the camera.