【作者】 张睿；

【导师】 龚卫国；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2014， 博士

【摘要】 在智能视频监控技术中背景建模是一项位于底层的关键技术，其性能将直接决定上层各种智能视频分析功能的可实现性及鲁棒性。对背景建模技术的研究近十年来一直是视频分析与安防监控领域的研究热点与难点，因此开展与背景建模相关的研究具有重要理论意义和实际工程意义。目前，大多数背景建模方法在实用化程度上仍存在不足，具体表现为无法应对现实监控场景的复杂多样性，其核心问题在于：已构建的背景模型无法快速有效地学习场景在时空维度上的各种随机性变化。于是，对背景模型自适应更新问题的研究成为背景建模技术实用化的关键一步。现有的主流背景模型自适应更新方法存在以下不足：需人工设置背景模型的初始学习率，自适应性有待提升；背景模型学习率的调控策略依赖于具体的背景模型，通用性不高；逐点式地计算背景模型学习率，运算效率低。为克服传统方法的上述不足，本论文提出了一种新颖的背景模型自适应更新方法。论文的主要研究工作如下：①受物理学中原子能级跃迁模型启发，论文提出将视频中像素灰度变化理解为像素点样本在不同灰阶（即光强能级）间发生了迁移，进而提出了以视频灰阶为对象提取视频变化统计信息的视频低层数据挖掘新范式——灰阶迁移统计法。相比于传统视频低层数据挖掘三大范式（即像素点分析范式、区域分析范式和子空间分析范式），灰阶迁移统计法能够从监控视频中挖掘出传统范式所无法获得的独特统计信息，该统计信息被证明可有效地用于控制背景模型的自适应更新过程。②针对传统背景模型自适应更新方法的不足，提出了一种基于灰阶迁移统计法的全局化背景模型自适应更新方法。该方法对视频中全局场景进行灰阶迁移统计，生成一种被称为全局灰阶迁移概率图的二维离散概率分布函数，然后将全局灰阶迁移概率图作为在线学习率查询表，以查表方式快速获取背景模型更新所需的学习率。该方法有以下优点：1）无需人工设置初始学习率，自适应程度高；2）学习率的产生不依赖具体背景模型，通用性好；3）学习率的产生由快速查表方式实现，运算效率高。实验表明，该方法可有效提高背景模型的自适应性与鲁棒性。③对于某些具有复杂局部动态性的监控场景，由②中方法计算出的全局灰阶迁移概率图可能出现误差。为此，通过对②中的全局化背景模型自适应更新方法进行改进，论文提出了一种基于灰阶迁移统计法的区域化背景模型自适应更新方法。该方法包含以下关键步骤：1）自适应的场景动态性估计；2）基于场景动态性的自适应场景区域分割；3）对不同的场景区域分别进行灰阶迁移统计，生成对应的区域灰阶迁移概率图；4）将区域灰阶迁移概率图作为对应区域内背景模型学习率的查询表。实验表明，区域化的方法能够有效地克服全局化方法存在的不足。④当场景中出现某些特殊事件（例如出现遗留物），在③中提出的区域化背景模型自适应更新方法将可能在特殊事件区域内失效。为此，论文提出了一种基于灰阶迁移统计法的特殊事件区域背景模型自适应更新方法，其由两部分组成：1）基于灰阶迁移概率图的非参数化特殊事件区域检测与分割；2）基于人类进行拼图游戏时的视觉感知机制对特殊事件区域内的背景模型进行自适应更新。最后，上述特殊事件区域背景模型自适应更新方法被整合到③中提出的区域化背景模型自适应更新方法中，从而有效地改进了区域化背景模型自适应更新方法的鲁棒性。通过在背景建模领域较权威的Changedetection标准测试数据集上的一系列实验表明：灰阶迁移统计法这种视频低层数据挖掘范式在应用上具有多样性，能有效挖掘出监控视频中隐藏的多种独特且有价值的统计信息，而基于灰阶迁移统计法的背景模型自适应更新方法明显优于传统的背景模型自适应更新方法。更多还原

【Abstract】 Background modeling (BGM) is a key technology in intelligent video surveillance,and its performance will determine the realization and robustness of various high-levelintelligent video analyses. Over the past decade, the study of BGM has been a popular,but challenging topic in the fields of video analysis and security monitoring. Therefore,it has both theoretical and engineering significance to carry out studies related to BGM.So far most BGM methods have insufficient practicability, due to the complexityand diversity of real-world scenes. The core problem is that a built background modelcannot rapidly and effectively learn all kinds of random changes in the temporal andspatial dimensions of the scenes. Hence, the study of adaptive background modelupdating is a critical step in BGM’s practical application. The existing popular methodsof adaptive background model updating usually have the following drawbacks: Initiallearning rates of the background models must be set manually, which leads toinsufficient adaptability; The learning rate control schemes usually depend on specificbackground models, which leads to poor generality; The pixel-wise calculation oflearning rates is needed, which leads to low efficiency. To overcome the abovedrawbacks of the traditional methods, a novel method of adaptive background modelupdating is proposed in this thesis. The main work of the thesis is as follows:①Inspired by the model of atomic energy level transition in physics, the thesisproposes that the pixels’ intensity changes in videos can be interpreted as the migrationsof pixel samples between different intensity levels. On this basis, a new paradigm oflow-level video data mining for surveillance videos, called intensity-level migrationstatistics (IMS), is proposed. Compared to three traditional paradigms of low-levelvideo data mining (i.e., pixel-based, regional-based, and subspace-based paradigms),IMS can mine unique statistical information that the traditional paradigms cannot obtainfrom surveillance videos. It is proved that the statistical information mined by IMS canbe effectively applied to control the adaptive background model updating.②To resolve the drawbacks of traditional adaptive background model updatingmethods, an IMS-based global method of adaptive background model updating isproposed. By calculating the statistics of the intensity-level migrations of pixels withinthe global surveillance scene, the method can generate a two-dimensional discreteprobability function called global intensity-level migration probability map (IMPM). On this basis, the global IMPM is utilized as an online learning rate lookup table, which isemployed to rapidly retrieve the suitable adaptive learning rates for background modelupdating. This method has the following advantages:1) It has good generality since thelearning rate generation is independent of background models;2) It has goodadaptability since there is no need to manually set any initial learning rate;3) It hasgood computational efficiency since all pixels’ learning rates can be rapidly retrievedfrom a lookup table. Experimental results show that the proposed method caneffectively enhance background models’ adaptability and robustness.③For certain surveillance videos with complex regional scene dynamics, eorrsmight occur in the above global IMPM. To improve the global adaptive backgroundmodel updating method proposed in②, an IMS-based regional method of adaptivebackground model updating is proposed. The method consists of the following steps:1)Adaptive scene dynamics estimation;2) Scene-dynamics based adaptive scenesegmentation;3) The generation of regional IMPMs by calculating the statistics of theintensity-level migrations of pixels within different scene regions;4) To utilize theregional IMPMs as the learning rate lookup tables for the corresponding regions.Experimental results show that the regional adaptive updating method can effectivelyovercome the defect of the global adaptive background model updating method.④When certain particular incidents (e.g., abandoned objects) occur in surveillancescenes, the regional adaptive background model updating method proposed in③maybe ineffective. Hence, an IMS-based adaptive background model updating method forthe particular incident region (PIR) is proposed. The method comprises two parts:1)IMPM-based nonparametric PIR detection and segmentation;2) Adaptive backgroundmodel updating for the PIR based on the human visual perception for jigsaw puzzles.Finally, the adaptive updating method for PIR is integrated into the regional adaptiveupdating method in③, therefore whose robustness is effectively improved.Through a series of experiments carried on the Changedetection benchmark dataset,it shows that the IMS could have a variety of possible applications and can mine uniqueand valuable statistical information from surveillance videos. Meanwhile, theIMS-based adaptive background model updating method can significantly outperformthe traditional adaptive background model updating methods.更多还原