【作者】 赵旭东；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 人工智能与信息处理， 2012， 博士

【摘要】 复杂多变的天气环境影响人类生产生活的方方面面。将带有复杂天气环境的视觉表现作为输入，会使基于计算机视觉的诸多领域应用面临前所未有的机遇和挑战。一方面，户外成像设备所获图像极易受到复杂天气环境的影响，导致其视觉效果、数据质量乃至应用价值下降。另一方面，如何从户外场景中有效地提取出所含天气环境的视觉表现，将为虚拟现实和基于计算机视觉的辅助气象预报提供相当大的技术支持。仅就户外视频分析而言，移动目标检测作为绝大部分实际应用的预处理模块，其检测结果不可避免地会受到复杂天气环境的干扰。为克服上述干扰，本文立足所创建的一种天气场景建模的通用研究框架，面向移动目标检测提出了鲁棒的天气场景建模方法，具体内容如下：(1)提出了一种基于多示例的场景动态区域分割方法。动态区域是指场景中存在明显像素变化的动态点集。作为带有复杂天气环境场景分类应用的预处理模块，简单而有效的动态区域分割方法有助于在场景中选出代表天气环境的关键位置用于分类。本文将场景动态区域分割抽象成一个多示例分类问题，通过构建示例特征的提取和排序、包距离度量公式、改进的多示例K均值聚类算法等一系列步骤，成功地完成有关移动目标检测的场景动态区域分割。(2)创建了一种带有复杂天气环境的场景客观分类方法。首先，根据对比不同天气环境的视觉表现所得出的三条假设构建了一个天气环境层次分类框架。在此基础上，从场景的动态区域中提取了一组时间、空间和彩色特征用于对带有复杂天气环境视觉表现的视频进行分类。此后，设计了一个将基于分类和回归树（Classification and Regression Tree, CART）与支持向量机（C-SVM）紧密结合的非度量分类器，用于完成对复杂天气环境的客观有效分类。(3)提出了一种基于变化时间窗的动态天气环境去除方法，作为动态天气条件下目标检测的预处理模块。现有的研究工作主要围绕如何从视频中有效地检测雨雪展开。本文则着重关注如何有效地去除检测到的动态天气环境。为此，本文改进了现存的动态天气环境检测算法，亦即构建了一种将K均值聚类的离线学习和基于高斯分布的在线学习相结合的学习策略。在此基础上，设计一个像素变化时间窗，用于去除视频中的动态天气环境。(4)创建了一种适应户外光照变化的自回归——纹理模型，用于在带有户外光照变化的视频中进行移动目标检测。根据复杂天气环境所含的不同视觉表现，考虑为不同的天气环境建立不同的场景模型。为此，本文提出了一种适用于缓慢光照变化和快速光照变化的自回归——纹理模型。该模型分别由适用于平稳图像序列的自回归（Auto Regression, AR）模型和对光照变化不敏感的纹理（Texture）模型构成。其一，针对不同的户外光照变化分别建立了基于帧间像素亮度差统计直方图的像素亮度扰动阈值，并将其引入到AR模型的快速背景估计中。其二，设计了一种递推的最小二乘法对AR模型进行实时地参数估计。其三，创建了一种精确的纹理度量，用于完善针对移动目标的估计。其四，基于亮度扰动阈值和自回归——纹理模型（TAR）构建了像素亮度置信区间和纹理置信区间，用于在带有户外光照变化的场景中完成移动目标检测。(5)提出了一种基于选择注意隐喻的分段记忆模型，用于复杂背景环境的移动目标检测。如何克服记忆容量的限制，对出现频率较低的确定背景状态建立模型，这是所有基于背景剪除的移动目标检测算法所面临的问题。解决上述问题的关键在于：采用何种实时检测和快速适应复杂环境变化（如不同的背景光照变化）的背景识别方法，以及如何建立带有记忆功能的背景模型。为此，本文提出了一种基于高斯混合模型（Gaussian Mixture Model, GMM）的分段记忆框架，用以解决有关移动目标检测的研究所面临的除语义反馈之外的诸如背景光照变化、背景周期性运动、背景的稳定性等一系列问题。上述有关天气场景建模的研究工作，为户外复杂天气条件下的移动目标检测提供了鲁棒的方法。依托所提的带有复杂天气环境的场景建模的通用研究框架，可以很好地克服甚至去除天气环境对场景中的移动目标所带来的干扰。此外，所提的分段记忆模型，为带有长时记忆背景模型的建立提供了一个新的思路。更多还原

【Abstract】 The complex and uncertain weather conditions afect every aspects of our human’sdaily life. The visual analysis on the complex manifestations of weather conditions chal-lenge many relevant methodologies and applications in computer vision. On the one hand,the manifestations of weather conditions will lead to a poor visual efect, low data qualityand reduced importance of application. On the other hand, an efective visualization ofweather conditions extracted from outdoor scenes will provide a first-step support on thevirtual reality and vision-aided weather forecast.Narrowing down to outdoor video analysis, object detection, which is regarded as apre-processing module of many practical applications, will inevitably obtain a poor detec-tion resulting from the disturbance of weather conditions. To this end, a robust weatherscene modelling method under severe weather conditions is proposed in this thesis basedon an established general weather modelling framework. More details are provided below.(1) A dynamic region segmentation approach is proposed based on multiple instancelearning (MIL). The dynamic region refers to the set of locations afected by temporalpixel changes in video. As a pre-processing module of complex weather classification, asimple but efective region segmentation approach makes a contribution to the selection ofkey locations representing the weather condition for the classification of diferent weatherconditions. In this research, the dynamic region segmentation is converted to the problemof the MIL. Through the phases of bag description, instance definition, instance sorting,distance measure and MI-based K-means clustering, we accomplish the dynamic regionsegmentation on videos.(2) A quantitative classification method on the visual efects of diferent weatherconditions is proposed. Due to the complex manifestations of weather conditions, threehypotheses are made. Moreover, a two-stage classification scheme is provided. Then,features derived from the spatio-temporal and chromatic space are extracted. Using theserepresentative features, we develop a quantitative classifier based on an experiential deci-sion binary tree associated with C-SVM.(3) A varying temporal window approach is proposed to remove dynamic weatherefects from a video. Particularly, we focus more on the removal of rain and snow ratherthan the detection method in this research. Dynamic weather conditions are detected by integrating an of-line K-means clustering with an on-line parameter maintenance ofGaussian distribution. Moreover, a variable time window containing adaptive backgroundedges is presented for removal of rain and snow.(4) An autoregressive texture-based background model combining short-term andlong-term analysis is established for accurate foreground detection from videos with vary-ing outdoor illumination. Diferent weather scenes can be modelled in terms of theirdiferent visual properties. Firstly, we discuss autocorrelation-based features for the iden-tification of foreground and outdoor illumination variations in short-term sequences, andpropose an adaptive threshold learning approach insensitive to inner-pixel fast illumina-tion variation based on the obtained histograms of intensity diferences between succes-sive frames. Then, an iterative orthogonal least squares (OLS) algorithm is designed toestimate the parameters for the auto regression (AR) model against gradual illuminationchange for background estimation in long-term sequence. Finally, we devise a texturemeasure to eliminate the regional efect of fast illumination variation.(5) A piecewise memorizing framework is proposed, which is capable of subtractinglong period video background under the restriction of memory capacity. Three major con-tributions can be claimed. Firstly, hypotheses of background subtraction indicating whatto recognize and memorize are proposed, taking the metaphors of psychological selectiveattention theory into consideration. Secondly, a prior perception-concerned recognition ofrapid illumination change is presented based on a segmented stationarity test. Thirdly, amemorizing framework based on GMM is put forward for the storage of long period back-ground. This framework is capable of identifying long period background appearances,as well as circumventing numerous typical problems except for semantic feed-back.The above research work on weather scene modelling methods results in robustforeground detection under complex weather conditions. Based on the proposed generalframework of scene modelling for the complex manifestations of weather conditions, thedisturbance of weather conditions can be overcome, and even removed. Besides, the pre-sented piecewise memorizing framework provides a new idea for background modellingwith long period memory.更多还原