细微运动的视觉增强及硬件加速技术研究

发布时间：2017-12-31 13:27

本文关键词：细微运动的视觉增强及硬件加速技术研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：智能化的视频监控作为计算机视觉研究的一个重要方向,它主要是利用计算机技术,对监控视野内的目标进行识别、追踪、行为描述等处理,其关注的目标通常是我们裸眼所能分辨的物体或者运动;然而我们的眼睛在观察目标时,对物体的空间尺度大小、运动幅度和频率都有一定的要求;所以在监控视频中还存在着一些我们裸眼难以分辨的细微运动,比如脸部颜色随血液的流动会发生微弱变化,腹部在呼吸时会有微小起伏。在医疗看护中,上述两种细微运动可以辅助护理人员检测病人的身体状况及其睡眠质量。因此我们需要对视频监控中的细微运动进行增强,使其可视化。在增强视频监控中细微运动时,我们首先需要对目标进行识别,然后提取目标所在区域,对提取区域中的细微运动进行放大,最后重新渲染视频。由于目标识别和区域提取已经是视频监控中非常成熟的算法,所以本文结合视频监控在医疗看护中的应用,重点研究了细微运动增强算法。首先,本文阐述了细微运动增强算法的相关原理。随后,本文结合视频监控的特点,对细微运动增强算法进行了优化,获得了良好的可视化效果。最后,本文实现了细微运动的实时增强。本文的具体工作如下。(1)线性欧拉运动放大算法虽然能同时放大目标的颜色和运动变化,但是输出视频中会存在着严重的噪声污染和伪影。相位欧拉运动放大算法虽然能解决噪声和伪影问题,但是计算复杂度高并且会产生大量的中间数据。为了解决计算量、噪声和伪影难题,本文首先将输入的视频序列进行完全金字塔分解;然后根据不同的运动类型,选取不同的带通滤波器来提取我们感兴趣的细微运动。由于提取后的图像中会掺杂相同频率的噪声,因此本文对图像进行一次平滑去噪。最后本文将去噪后的图像进行放大并重新渲染视频。(2)由于上述算法属于计算密集型运算,同时还要满足视频监控中实时性的要求。所以本文提出了一种基于FPGA的硬件加速方案。在硬件实现过程中,本文首先对细微运动增强算法中的色彩空间转换模块、金字塔分解模块、去噪模块和滤波模块进行了硬件设计。然后根据各个模块的处理流程,设计了流水线架构。为了实现FPGA对DDR的连续快速访问,本文在DDR和FPGA间设计了ping-pong数据缓冲结构。(3)本文首先在CPU上对细微运动增强算法进行了验证,然后结合Xilinx KC705板卡实现了视频监控中细微运动的可视化。我们对输出视频的时间切片、亮度变化、PSNR等参数进行了分析,本文算法在具有良好可视化效果的同时,对噪声具有良好的抑制能力。最后我们设计了细微运动的监控系统,并且完成了细微运动的实时监控。
[Abstract]:As an important direction of computer vision research, intelligent video surveillance mainly uses computer technology to identify, track and describe the target in the field of surveillance. The object of attention is usually the object or motion that we can distinguish with the naked eye. However, when our eyes observe the target, there are certain requirements for the spatial scale, motion amplitude and frequency of the object. So there are some subtle movements that we can't tell from the naked eye in the surveillance video, such as the slight change in color of the face with the flow of blood, the slight fluctuation of the abdomen as it breathes. In medical care. Both of these subtle exercises can assist nursing staff in monitoring the patient's physical condition and sleep quality, so we need to enhance the fine movement in video surveillance. In order to enhance the fine motion in video surveillance, we first need to identify the target, then extract the region where the target is located, and enlarge the fine motion in the extracted region. Because target recognition and region extraction are very mature algorithms in video surveillance, this paper combines video surveillance in medical care applications. Focus on the fine motion enhancement algorithm. Firstly, this paper describes the principle of the fine motion enhancement algorithm. Then, combining the characteristics of video surveillance, this paper optimizes the fine motion enhancement algorithm. A good visualization effect is obtained. Finally. In this paper, the real-time enhancement of fine motion is realized. The specific work of this paper is as follows: 1) Linear Euler motion amplification algorithm can amplify both color and motion changes of target at the same time. However, there will be serious noise pollution and artifacts in the output video. Although the phase Euler motion amplification algorithm can solve the noise and artifact problems. In order to solve the problem of computational complexity, noise and artifact, the input video sequences are decomposed into complete pyramids. Then according to different motion types, different bandpass filters are selected to extract the fine motion we are interested in. Because the extracted image will be doped with the same frequency of noise. Therefore, the image is de-noised once. Finally, the de-noised image is amplified and the video is rerendered. (2) because the above algorithm is computationally intensive. At the same time to meet the real-time requirements of video surveillance. So this paper proposes a hardware acceleration scheme based on FPGA. In the process of hardware implementation. In this paper, the color space conversion module, pyramid decomposition module, denoising module and filtering module are designed firstly. Then according to the processing flow of each module. In order to realize the continuous and fast access to DDR by FPGA, the pipeline architecture is designed. In this paper, we design a ping-pong data buffer structure between DDR and FPGA.) in this paper, we first verify the fine motion enhancement algorithm on CPU. Then we realize the visualization of fine motion in video surveillance with Xilinx KC705 card. We analyze the time slice, brightness change and other parameters of video output. At the same time, the algorithm has a good ability to suppress noise. Finally, we design a monitoring system for fine motion, and complete the real-time monitoring of fine motion.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】