面向智能视频监控的高动态场景建模和修复

发布时间：2018-12-07 21:12

【摘要】：在智能视频监控中,场景建模和修复是目标检测和场景理解的核心内容。计算机视觉科技的进步使得背景建模的算法日臻成熟,对于普通场景中的前景物体变化都能做到较为准确的识别。但在视频监控中常常存在着一些复杂高动态的场景,其中前景在时域和空域中的比例都要大于背景,目前流行的背景建模算法难以做到实时、准确地建模。针对此问题,本文提出了一种基于“像素-模型”(Pixel to Model,P2M)距离的无参数背景模型。本论文在学术上的主要贡献如下。1.本文提出了一种全新的像素表达框架:基于压缩感知的理论,利用图像中各点上下文信息,通过提取局部特征的方式来对中心像素进行建模。本文分析了之前多种基于色彩值的背景建模方式,针对其局限性提出了引入空域信息对像素点进行建模的方法,并使用压缩感知的方法解决其面临的巨大计算量的问题,将广义的Haar-like特征引入背景建模,完成本文背景建模中像素点的特征表达。通过引入基于纹理和边缘的图像上下文信息,本文的背景模型在前景检测上比单像素色彩模型更为准确,在监控行业常见的高动态场景中的鲁棒性也有所提高。2.针对先前的“像素-像素”背景建模方法的不足,本文提出了“像素-模型”距离的概念,以量化像素点和背景模型之间的相似度。该量化方式是本文前景分割和模型更新的基础。同时,本文使用了最小和最大“像素-模型”距离来对像素点以及其邻域像素点的背景模型进行更新,并对其中一些参数的自适应性做了推导和说明。在复杂高动态环境的前景分割实验中,该模型优于主流的四种背景建模方法;在后续的应用开发上,利用该模型能进行有效的智能视频监控。3.受高斯混合建模求取期望进行背景生成以及通过历史帧的线性组合进行背景像素点计算方法的启发,本文提出了基于贝叶斯方法,使用最小“像素-模型”距离进行权重计算,并通过类似期望求取的办法对场景图像进行修复的方法,提高了视频监控中场景修复的准确性。
[Abstract]:In intelligent video surveillance, scene modeling and restoration are the core contents of target detection and scene understanding. With the development of computer vision technology, the background modeling algorithm is becoming more and more mature, which can recognize the change of foreground objects in common scene accurately. However, there are some complex and highly dynamic scenes in video surveillance, in which the foreground is larger than the background in the time domain and the spatial domain. The current popular background modeling algorithms are difficult to achieve real-time and accurate modeling. To solve this problem, a nonparametric background model based on "Pixel-Model" (Pixel to Model,P2M distance is proposed in this paper. The main contributions of this thesis are as follows. 1. In this paper, a new framework of pixel representation is proposed: based on the theory of compression perception, the central pixel is modeled by extracting local features by using the context information of each point in the image. In this paper, several methods of background modeling based on color value are analyzed, aiming at its limitation, a method of introducing spatial information to model pixels is put forward, and the problem of huge computation is solved by using compressed sensing method. The generalized Haar-like feature is introduced into the background modeling, and the feature representation of pixels in the background modeling is completed. By introducing context information based on texture and edge, the background model in this paper is more accurate than the single pixel color model in foreground detection, and the robustness in high dynamic scene is improved. 2. In this paper, the concept of "Pixel-Model" distance is proposed to quantify the similarity between pixel points and background models in view of the shortcomings of previous "Pixel-Pixel" background modeling methods. This quantization method is the basis of foreground segmentation and model updating in this paper. At the same time, the minimum and maximum "Pixel-Model" distance is used to update the background model of pixels and their neighboring pixels, and the self-adaptability of some of the parameters is deduced and explained. In the foreground segmentation experiment of complex high dynamic environment, the model is superior to the four main background modeling methods, and in the following application development, the model can be used to carry out effective intelligent video surveillance. 3. Inspired by Gao Si's mixed modeling and expected background generation, and by linear combination of historical frames, this paper proposes a method based on Bayesian method, which uses the minimum "Pixel-Model" distance to calculate the weight. The accuracy of scene restoration in video surveillance is improved by the method of scene image restoration.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;TN948.6

【共引文献】