基于超像素分割的时空上下文模型视频追踪算法研究

发布时间：2018-03-31 04:27

本文选题：超像素分割　切入点：时空上下文　出处：《甘肃政法学院》2017年硕士论文

【摘要】：现今,随着天网全方位的建立,图像和视频已成为公安破案的重要依据,计算机视觉中的图像处理,视频的动态追踪也成为世界公安技术领域相关专家研究的热点问题,人工智能、机器学习的飞速发展,也使得计算机在针对图像、视频的分析获取得了快速发展。在道路监控视频中,往往由于一些树木,车辆或其他因素导致追踪目标被遮挡,计算机无法完成连续追踪,只能依靠人力去追踪,这就大大浪费了警力和时间,如何能够在遮挡或半遮挡的情况下,实现目标的快速动态追踪,是本文研究的重点,也是国内外的一个研究热点。同时,为了解决快速追踪的问题,本文采用了图像处理技术中的超像素分割技术,将像素级转化成超像素级问题,实现了快速追踪。由于图像分割的质量决定了后续问题的处理,所以在超像素分割算法的选择上也成为一个重要因素。本文利用融合了超像素分割技术的时空上下文先验模型来解决半遮挡问题。文章首先介绍现阶段国内外对于视频目标追踪技术的发展情况,提出研究的背景及意义,并提出利用基于熵率的超像素分割算法融合时空上下文先验模型进行视频目标的动态追踪,在实践中,通常采用低层特征处理图像视频,而很少采用中层视觉特征,本文则是利用了中层视觉特征,构建上下文模型,求解置信图,获得目标位置在新一帧的最大概率。首先建立时空上下文先验模型,利用贝叶斯框架来做目标模型构建,以及目标与背景邻近的子区域的关系,通过时空上下文模型确定目标位置的置信图,并在目标区域采用均匀采样点追踪器,来完成当前帧的目标位置确定。然后,使用超像素分割算法进行分割,利用像素位置距离超像素中心点的远近,算出所求目标在每个像素位置出现的概率,将置信图转化成像素级置信图,估计目标在所寻求区域的每个位置出现的概率问题,即概率最大的即为目标新一帧所出现的位置。最后,利用捷尚公司的测试视频和标准视频库中的视频,对本文的算法进行实验验证,验证了算法的正确性和有效性的同时,提出了算法分析及对运算结果的评价,最后对本文涉及的图像处理及视频跟踪技术的应用进行了总结展望。
[Abstract]:Nowadays, with the establishment of Skynet, images and videos have become an important basis for police to solve cases. Image processing and video dynamic tracking in computer vision have also become hot issues in the field of public security technology in the world. With the rapid development of artificial intelligence and machine learning, computer analysis of images and videos has also developed rapidly. In road surveillance videos, tracking targets are often blocked because of trees, vehicles or other factors. The computer can not complete continuous tracking, it can only rely on manpower to track, which greatly wasted the police force and time. How to achieve fast dynamic tracking of targets in occlusion or semi-occlusion, is the focus of this paper. At the same time, in order to solve the problem of fast tracking, this paper adopts the super-pixel segmentation technology in the image processing technology, which transforms the pixel level into the super-pixel level problem. Fast tracking is realized. Because the quality of image segmentation determines the processing of subsequent problems, Therefore, the selection of hyperpixel segmentation algorithm is also an important factor. In this paper, a spatio-temporal contextual priori model combining hyperpixel segmentation technology is used to solve the semi-occlusion problem. The development of video target tracking technology, The background and significance of the research are presented, and a hyper pixel segmentation algorithm based on entropy rate is proposed to fuse the temporal and spatial context prior model for dynamic tracking of video objects. In practice, low level features are usually used to process video images. However, the middle vision feature is seldom used. In this paper, the middle level visual feature is used to construct the context model, to solve the confidence chart, and to obtain the maximum probability of the target position in the new frame. Firstly, the temporal and spatial context priori model is established. The Bayesian framework is used to construct the target model and the relation between the target and the sub-region adjacent to the background. The confidence chart of the target location is determined by the spatio-temporal context model, and the uniform sampling point tracker is used in the target area. To determine the target position of the current frame. Then, using the hyperpixel segmentation algorithm, the probability of the target appearing at each pixel position is calculated by using the distance from the pixel position to the center of the super-pixel. The confidence chart is transformed into a pixel level confidence chart to estimate the probability problem of the target at each location in the region sought, that is, the position where the maximum probability is the occurrence of the new frame of the target. Finally, By using the test video and the video in the standard video library, the algorithm of this paper is verified by experiments, and the correctness and validity of the algorithm are verified. At the same time, the analysis of the algorithm and the evaluation of the operation result are presented. Finally, the application of image processing and video tracking technology in this paper is summarized and prospected.
【学位授予单位】：甘肃政法学院
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：D918

【相似文献】