Research On Hardware Acceleration Technology Of Optical Flow Method

In Video Measurement,the pixels displacement are often employed to extract the motion characteristics of objects in a scene.Optical flow is a vector field describing the relative displacement of pixels between two consecutive frames in a video sequence.Optical flow method is a method of estimating the vector field,which can be used to obtain the motion characteristics of an object.The optical flow method is currently used in the fields of target identification,obtaining moving target masks,etc.,and is also indispensable for global morphological resistance measurements on the surface of wind tunnel test models.Among many optical flow methods,the variational optical flow method represented by the Horn-Schunck method has attracted the attention of researchers for its ability to obtain dense optical flow field information.However,as a high-quality optical flow method,the variational optical flow method has the problems of complex computation and large amount of computation,for high-resolution images are difficult to meet the requirements of real-time processing,which limits its application in practical engineering.Various existing hardware acceleration schemes are difficult to achieve the effect of real-time optical flow calculation for high-resolution images.The research on the hardware acceleration technology of the optical flow method has great research significance in the engineering field that needs to obtain the target motion information quickly.To this end,the following studies is carried out around high-resolution image real-time optical flow computation:1、Starting from the theoretical model of the variational optical flow method,this paper analyzes in detail the technical strategies used in the variational optical flow method.It is proposed that: 1)AnIncremental updating of the numerical differentiation method to reduce the time overhead of the gradient calculation phase;2)An image sequence processing strategy based on dynamic gray-scale thresholds to remove some pixels affected by light changes,and improving Algorithm efficiency.2、 By summarizing the general workflow of variational optical flow computation and analyzing its parallelism and time complexity,a GPU-based parallel optimization strategy is designed: 1)using shared memory to improve data access speed and speeding up the calculation process;2)designing images Data parallel storage and retrieval structure to reduce data transmission time and achieve the purpose of improving real-time performance.3、A study of hardware resource optimization schemes and time consumption prediction models.For a given hardware resource,predicting the time consuming optical flow computation based on image resolution and the number of iterations is used as a reference by the researcher to modify the image resolution,number of iterations,or camera frame rate,etc.The hardware resource optimization analysis provides a hardware resource allocation scheme for achieving better real-time processing results.4、Building a CPU-GPUbased real time flow computing system.First of all,we designed the overall system scheme,built the system hardware implementation platform and arranged the PCIE data transmission timing.Interactive software for real-time monitoring and dynamic display of the optical flow field was developed,which also configures other parameters such as the camera and stores the results of optical flow computation.The system has a real-time processing speed of 80 frames per second(fps)for images with a resolution of 1728 × 2352.