Research On Visual Enhancement Of Subtle Motions And Hardware Acceleration Technology

Intelligent video monitoring is an important direction of computer vision.It mainly focuses on target recognition,tracking and behavior description with computer technol-ogy.The targets are usually the object or movement that we can see.The spatial size,movement amplitude and frequency of an object are limited by our eyes,so there are some subtle variations that cannot be seen with our naked eyes.For example,the color of the human face is changing slightly while the blood is circulating in body.The ab-domen has a slight fluctuation while breathing.In medical care,these two kinds of subtle motions can help nurses to monitor the physical condition of patients,as well as the quality of sleep.Therefore,we need to enhance these subtle motions in monitor videos,so make them visible with our naked eyes.When enhance subtle motions,we first need to identify the target,then extract the target area and amplify the subtle motion,finally re-render the video.Because the target recognition and area extraction are mature algorithms in video monitoring,so the main topics of this dissertation are subtle motion amplification algorithm and the application of video monitoring in medical care.The First of all,the dissertation describes the subtle motion enhancement technology.Then,combined with the characteristics of video monitoring,we optimize the subtle motion enhancement algorithm,and obtain good visualization effect of subtle motion.Finally,we complete the real-time processing of subtle motion enhancement.The specific work of the dissertation is as follows.(1)Although the linear Eulerian motion magnification algorithm can amplify the subtle color and motion changes of the target at the same time,there are serious noise pollution and artifacts in the output video.Phased-based Eulerian motion magnification algorithm can solve the problems of noise and artifacts,but the computational complex-ity is high and will produce a large amount of intermediate data.In order to solve the problems of the computational complexity,noise and artifacts,we first decompose the input video sequence into a complete pyramid.Then we choose different band-pass filter to extract the subtle motions according to the different types of motion.As the extracted image will be doped with the noise of same frequency,so we smooth the im-age to remove the noise.Finally,the dissertation will amplify the denoised image and re-render the video.(2)As above algorithm is computationally intensive and the video monitoring is real time.Hence,an FPGA-based hardware accelerator is presented in the dissertation.In the hardware implementation,hardware designs of the color space conversion mod-ule.pyramid decomposition module,denoising module and filter module of the subtle motion amplification algorithm are firstly presented.Then the pipeline architecture is designed based on the processing flow of these modules.In order to achieve continu-ous and fast access to DDR,the dissertation designs a ping-pong data buffer structure between DDR and FPGA.(3)The dissertation first validates the subtle motion amplification algorithm on CPU.Then subtle motion visualization in video monitoring is implemented on KC705 FPGA.The analyzation on time slice,brightness change,PSNR of the output video is presented,and the proposed algorithm has better visual effect of subtle motions,and shows better effect on noise reduction.Finally,we design a monitoring system of subtle motions,and complete the real-time monitoring.