【作者】 侯杰；

【导师】 钱稼茹；

【作者基本信息】 清华大学 ， 土木工程， 2010， 博士

【摘要】 基于视频图像的探测方法具有探测范围广、响应时间短、成本低等优势,是解决高大空间建筑火灾探测问题的有效方法。基于视频图像的探测由三个环节组成:目标提取、目标跟踪和目标识别。由于高大空间中图像信息的复杂性,如光照变化、物件移动、目标数量多、运动模型多变、目标遮挡、镜头遮挡、小样本等,使得火灾探测的准确性、实时性和鲁棒性难以同时满足。本文改进和提出了关于目标提取、跟踪和识别的系列算法,并开发了针对高大空间建筑火灾探测的软件系统。本文的研究成果主要有以下几个方面:(1)目标提取:主要解决高大空间中由于光照变化、物件移动等而造成的不能准确、及时提取火灾目标(火焰、烟雾)问题。分析了当前主要目标提取方法的特点及实际效果;针对背景差分法中的背景更新问题,提出了一种基于跟踪和识别信息的时空自适应背景更新法:把目标跟踪、识别的结果信息进行反馈,用以指导背景更新,以考虑目标类型、位置和频率特征等差异。(2)目标跟踪:主要解决多目标跟踪的模型多变和遮挡问题。构建了用于多目标提取、跟踪和识别的数据存储结构;针对状态向量是测量向量扩展形式的平方根无迹卡尔曼滤波(square root unscented kalman filter, SR-UKF),提出了一种新的精简算法;针对高大空间中运动模型多变、目标遮挡和镜头遮挡等问题,把多模型、数据延迟和模糊自适应融入到多目标跟踪体系中,提出了一种新的多目标跟踪方法——基于多模型和数据延迟的模糊自适应多目标跟踪。(3)目标识别:针对高大空间条件下火灾识别的小样本问题,利用基于遗传算法的最小二乘支持向量机(genetic algorithm based least square support vector machine, GALSSVM)中GA样本训练的结果信息构建了模糊隶属度函数,形成了基于模糊隶属度和遗传算法的最小二乘支持向量机(fuzzy membership and genetic algorithm based least square support vector machine, FGALSSVM);针对单一、瞬态识别算法的准确性和鲁棒性不足问题,把瞬态融合和历史信息融合的概念引入到目标识别中,建立了基于算法融合的目标识别框架。除了改进算法本身,本文开发了各主要算法的源代码以及软件系统,并基于火灾实验的视频文件,对改进算法的实际效果进行了实验验证。更多还原

【Abstract】 Video-image fire detection performs well in detection range, response time and cost. So it is more efficient than traditional methods in fire detection for large space structures. In essence, video-image fire detection is composed of three parts: object extraction, tracking and recognition. Due to the complexity of the visual environment in the large space structures, such as change of illumination, moving of objects, large number of objects, variation of motion model, object shading, lens shading and the problem of the small sample size, it is hard to realize accurate, real-time and robust detection of fire at the same time. In this research, a series of algorithms about object extraction, tracking and recognition are improved and proposed. A prototype software system is developed to detect fire for large space structures. The main contents are as follows:(1) Object extraction. Due to the changing of lighting and moving of objects, it is difficult to achieve accurate and real-time object extraction. First, the feature and performance of main extraction algorithms are analyzed. Then a new space-time adaptive algorithm for background updating is proposed for the background subtraction. The advance is that the information about object tracking and recognition can be applied to guide the object extraction. So the type, position and frequency of the object can be gained and considered.(2) Object tracking. The key problem is the variation motion model and shading in multi-object tracking. First, a data structure is established for multi-object extraction, tracking and recognition. Then a new concised SR-UKF(square root unscented kalman filter) is proposed when the measuring vector is a subset of the state vector. At last, multiple motion model, data delay and fuzzy adaption are introduced into the multi-object tracking system to improve the tracking performance. So a new multi-object tracking algorithm is proposed. That is, fuzzy adaptive multi-object tracking algorithm based on multiple model and data delay.(3) Object recognition. In order to conquer the problem of the small sample size in fire detection for large space structures, functions of membership degree are constituted based on the final training result of the GA(genetic algorithm) in GALSSVM(genetic algorithm based least square support vector machine). So a new SVM is proposed with the name FGALSSVM(fuzzy membership and genetic algorithm based least square support vector machine). To improve the accuracy and robustness of single and transient recognition algorithms, the concept of transient data fusion and historical data fusion are introduced into the system of recognition. So a new type of recognition frame is established.Except for improving the algorithms, the source codes for main algorithms and the prototype detecting software are developed. A series of tests are conducted to verify the performance of the improved algorithms.更多还原