Research On Close-range Video Measurement And Its Hardware Acceleration Technology

Video Measurement(VM)can measure and estimate the structural parameters and motion parameters of the target from the image.It is a non-contact Measurement method and has the advantages of not damaging the target to be measured,not interfering with the experimental environment and high Measurement accuracy,which has been favored by scholars and researchers at home and abroad.Nowadays,video measurement has two development trends,which are higher precision video measurement and real-time video measurement.And with the development and progress of science,the application of video measurement is more and more extensive.This paper focuses on the relative orientation technology and the image mark point recognition technology in video measurement as follows:1)In order to solve the problem that the iterative solution may be far from the optimal solution when the traditional relatively directional nonlinear coplanar equation is optimized,which leads to the divergence of the iterative process,an optimization method of the coplanar equation is proposed by using the constraint interval to limit the iterative domain.Firstly,a method of solving the radius of the constraint interval with the initial value as the center is designed,and a coplanar equation minimization model of the interval constraint is established.Secondly,in order to facilitate the optimization solution,the penalty function is constructed to transform the original constraint minimization problem into an unconstrained minimization problem.In the end,the optimization process of a relatively directional coplanar equation is given,which limit the iteration process to complete within the constraint interval,so as to avoid iteration divergence.2)Aiming at the problem that image tag point recognition in video measurement takes a long time,and then affects the real-time video measurement,a parallel processing method for image tag point recognition is proposed.Firstly,an eight-step tracking method with direction rotation is proposed,which solves the problem of repetitive retrieval of contour search in the traditional eight-step tracking method,and saves the time of contour search of marked points.Secondly,a chunking image segmentation and parallel processing scheme for image tag point recognition is designed to improve the speed of image tag point recognition.Finally,a structure for quickly storing and retrieving image data is designed to hide the time-consuming data transmission and improve the overall recognition efficiency.3)Resource optimization and time prediction models are established.The prediction model establishes the relationship between the time cost of identification of marker points and the experimental conditions such as image resolution,the size of marker points,the number of marker points and hardware resources.It is convenient for the experimenter to verify the image resolution and image marker size in real-time processing.Resource optimization by analyzing the current resource situation,the actual optimal number of parallel paths can be determined to achieve the best real-time processing capability.4)Designed and developed a CPU-GPU heterogeneous real-time video measurement system,built a corresponding hardware experiment platform,and developed user interactionsoftware.System on the CPU to complete image monitoring,data transmission,calculation the values which are to be measured and data storage,and other functions,completes the real-time identification of image marker points on the GPU side and achieves the real-time processing speed of 2GB/s.