基于DSP的固定场景视频消抖快速算法研究与实现

发布时间：2018-01-12 09:05

本文关键词：基于DSP的固定场景视频消抖快速算法研究与实现　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】：随着视频监控产品市场的繁荣,户外视频监控产品种类越来越多,应用场合也越来越广泛。此类产品的摄像头容易受到外界环境因素的干扰,比如吹风或摄像头基座晃动,导致成像视频画面的抖动。抖动的视频会使观察者产生视觉疲劳,造成对视频内容信息的误判;同时,视频抖动会严重干扰后期应用的一些视频处理的效果。因此,消除监控视频的抖动已成为一个亟待解决的问题。视频消抖,也被称为稳像,是计算机视觉领域重要的研究方向之一,其实质是利用几何变换关系补偿图像序列间的非规则运动,以实现视频的稳定。针对固定场景的视频,即拍摄视频的摄像头没有大幅移动、监控固定区域,本文分析了该场景视频抖动的特点,研究了稳像的原理和系统结构,深入探讨了固定场景稳像各环节的关键技术,包括图像变换模型、全局运动估计、运动补偿和稳像质量评价方法。针对稳像的现实需求,本文提出了一种固定场景的视频消抖算法,完成了在监控产品DSP处理器上运行的稳像算法库的开发。具体的研究内容和成果如下:综合考虑时间开销和计算精度后,设计出以Harris角点检测算子和金字塔LK光流法为基础的全局运动估计算法,选择描述图像变换充分的单应性模型,同时通过特征点的布局优化、角点亚像素化、RANSAC剔除干扰因素和单纯形法优化运动参数这四个措施来提高运动参数估计的精度。在运动补偿环节,采用逆映射法和双线性插值提高重构图像的质量。多个方法的评价结果显示,该算法能快速、有效地解决固定场景视频抖动的问题。为了充分利用DSP的硬件资源,提高稳像算法库的运行速度,使用了 5种优化稳像代码的方法。这5种方法包括:(1)改善软件流水,提高并行化处理速度;(2)编译器优化,使编译器编译出高效的代码;(3)浮点运算转定点运算,体现DSP的计算优势;(4)使用内联函数和软件库,将部分函数的运行效率提高到汇编级;(5)设计DMA搬移数据,让CPU专心负责运算。对于720×480分辨率的视频,这些优化工作使得稳像程序的运行耗时从2300毫秒/帧降低到40毫秒/帧,从而能够满足实时性的要求。
[Abstract]:As the video surveillance products market prosperity, outdoor video surveillance products more and more types of applications, such products are more and more widely. The camera is susceptible to interference from external environmental factors, such as hair or camera base shaking, resulting in imaging video screen shake. Video jitter will make the observer to produce visual fatigue, resulting in the video at the same time, the information content of misjudgment; video jitter will seriously interfere with the late application of some video processing results. Therefore, the elimination of video jitter has become an urgent problem to be solved. Video debounce, also known as stabilization, is one of the important research directions in the field of computer vision, its essence is the use of geometric transformation the relationship between compensation between image sequences and irregular movement, in order to achieve video stability. For the fixed scene video, video camera is not a mobile monitoring The fixed area, this paper analyzes the characteristics of the scene video jitter, study the principle and system structure of image stabilization, in-depth study of the fixed scene image stabilizing the key techniques, including image transformation model, global motion estimation, motion compensation and image quality evaluation method. According to practical needs of stabilization, is proposed in this paper. Debounce algorithm for fixed scene video, completed the development of monitoring products run in the DSP processor on the image stabilization algorithm library. The specific research contents and results are as follows: considering the time cost and accuracy, design a global motion using Harris corner detection operator and the Pyramid LK optical flow method based estimation algorithm full description of the image transform, select the homography model, the layout optimization of feature points, sub-pixel corner of RANSAC, eliminating interference factors and simplex method optimization of motion parameters of the four. To improve the accuracy of motion parameter estimation. The motion compensation link, using the inverse mapping method and bilinear interpolation to improve the quality of the reconstructed image. To evaluate multiple methods. The results show that the algorithm can quickly and effectively solve the fixed scene video jitter problems. In order to make full use of DSP hardware resources, improve the image stabilization algorithm speed library, image stabilization code method uses 5 kinds of optimization. These 5 kinds of methods: (1) improve the software pipelining, improve the parallel processing speed; (2) the compiler optimization, the compiler to compile an efficient code; (3) floating-point operation to fixed-point arithmetic, reflect the computational advantages of DSP (4;) using inline functions and software libraries that will improve the efficiency of some functions to the assembly level; (5) the design of DMA transport data, let CPU attend to operation. For the 720 * 480 resolution video, these optimization work makes the stabilization program running time from 230 0 milliseconds / frames are reduced to 40 milliseconds / frames, which can meet the requirements of real time.

【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;TN948.6

【相似文献】