【作者】 周宁；

【导师】 周曼丽；

【作者基本信息】 华中科技大学 ， 通信与信息系统， 2006， 博士

【摘要】 随着视频信息处理技术和网络技术的飞速发展,开发、利用监视视觉信息成为了一种不可避免的趋势。智能化监视视频的传输、分析、存储与检索、以及与其他控制系统无缝集成等,均依赖于对场景语义视频对象的处理。本文以语义视频对象为中心,研究公共安全监视中语义视频对象检测、跟踪以及监视系统与其他子系统在对象层次上的互操作等问题。本文分析了视频监视系统智能化的技术要求;回顾了运动对象的分割、跟踪,以及运动阴影检测的发展状况;阐述了BACnet协议作为智能控制系统通信平台,缺少基于监视视觉信息的系统互操作手段。本文根据监视应用的不同语义抽象,建立了三种不同的监视语义抽象:运动对象、运动阴影和运动blob。每个语义视频对象的属性包括图像的像素特征和语义特征。用这些属性构成语义描述符,以一种简洁的形式表达监视视觉数据。针对复杂场景中背景的不完整性、随机噪声以及目标运动快慢不一等影响背景估计的因素,本文提出了一种基于时空相似度量的复杂场景背景估计方法。首先度量同一位置不同时刻子块的相似性矩阵,然后度量候选背景子块的空间相似性,从而判别最可能的背景子块。该方法对噪声、运动目标速度有较强的适应性,计算代价较低。本文提出了一种基于反射率相似子区域分析的运动阴影抑制方法。算法通过分析反射率相似子区域中的环境光照特征和边缘能量信息,从而区分运动对象及其跟随的运动阴影。该方法对于室内投射阴影检测较为有效。本文提出了一种基于语义交互的运动对象跟踪算法。算法将运动人体初始化为头部、躯干和下肢等运动blob,表达为相应的blob描述符。通过投影blob描述符,更新、验证运动对象,实现对运动对象的跟踪。算法利用改进的快速高斯变换计算各个运动blob,并选择参与估计的目标数据和源数据样本,以降低计算代价。该跟踪方法处理简单、计算代价较低,能较好地处理不同运动对象之间的部分遮挡问题。本文首次提出了一种BACnet视频对象模型及视频点操作服务,实现了监视系统与其它控制系统之间在对象层次的互操作,并在此基础上搭建了基于场景事件的楼宇视频监视应用方案。为了使用户能够在复杂的楼宇分布控制系统中合理地部署系统智能,本文采用FIPA平台构建了一种多主体多服务器结构模型,充分考虑所集成子系统内部和子系统之间的请求信息交互,给出了主体的核心结构及主体间服务请求的控制管理方案。更多还原

【Abstract】 With the rapid development of video processing and network technology, exploiting visual surveillance information has become an inevitable trend. The surveillance system should be capable of transmitting video data safely, analyzing visual scene, storing and indexing video data by scene information, and integrating with other control systems. All those capabilities rely on processing of semantic video objects in scene. This thesis focuses on semantic video objects segmentation and tracking in surveillance scene, and interoperation based on objects among different systems.Technical requirements for intelligent surveillance system are discussed at first. Then the state-of-the-art of moving object segmentation, tracking and moving shadow detection techniques is reviewed. As communication platform for building control system, BACnet protocol needs interoperation based on surveillance information.Semantic abstract in scene are different for different surveillance applications. Three semantic video objects are defined in this thesis, which are moving object, moving shadow, and moving blob. Every semantic video object’s properties including pixel and semantic characteristics constitute semantic descriptors, which represent visual data semantically.Problems like unavailable background pixels, noise and moving objects’velocity in the scene make background estimation more difficult in video surveillance. A similarity-measurement method is provided to reconstruct background. By comparing with temporal blocks and spatial blocks, block similarity measurement helps to valid candidate background blocks. The method deals well with noise and moving object’s velocity automatically, and has lower computation cost.This thesis proposed a multi-feature moving shadow detection approach based on albedo ratio similarity region. After analyzing ambient illumination feature and edge information in those similarity regions, moving shadow can be detected from moving video object. The approach is suitable for indoor shadow detection.Semantic interaction based moving object tracking is put forword to deal with occlusion when two objects interact. The approach is based on modeling major color regions of moving human body such as head, torso, and lower limbs blob. These blobs are represented as moving blob descriptions. After projecting those descriptions, moving objects should be refined and validated. Improved fast gauss transform (IFGT) is exploited for semantic video object blob. By choosing target and source number, which used for IFGT, computation cost becomes lower. The tracking approach is simple and available for multi-object tracking.For system interoperation, a new BACnet video object model and video point operation service are proposed at first time. According to this model, a surveillance application scheme based on scene events is built in intelligent building system.In order to put intelligent video surveillance system into building control systems, FIPA-based multi-Agent multi-service integration architecture is put forward. The architecture discussed core framework of agent and request control among agents. It’s convenience for consumers to deploy system intelligent.更多还原