【作者】 胡宏宇；

【导师】 王殿海；

【作者基本信息】 吉林大学 ， 交通信息工程及控制， 2010， 博士

【摘要】 基于视频处理的交通事件识别是智能交通领域中较有前途的应用技术之一,可为城市交通管理与控制提供全面、实时的交通状态信息。目前基于视频处理的智能化交通事件识别系统仍处于发展阶段,在某些关键技术环节,特别是在运动目标交通行为分析以及交通事件识别模型构建等方面尚需进一步研究。本文结合图像处理、模式识别、机器学习等理论与方法,围绕交通事件视频识别过程中的若干关键技术问题进行系统性的研究。首先针对背景模型的背景初始化、背景表达与更新进行深入探讨,实现了复杂场景下的自适应前景运动区域检测;同时给出了运动阴影去除的解决方案,提高了运动目标的检测精度。结合运动目标的形态特性与运动特性设计了基于多类支持向量机的分类算法,实现了混合交通运动目标的类别判断。提出了基于卡尔曼滤波的多特征匹配跟踪算法与基于历史运动信息补偿的遮挡处理方法,保证了复杂遮挡情况下的运动状态准确估计。综合考虑轨迹的空间特性、方向特性、类别特性,提出了运动模式的多层次学习方法,并由此构建了基于Bayes空间模式匹配以及基于起讫点方向模式匹配的异常行为检测方法。将运动目标的状态属性与交通场景中的上下文相关信息相结合,定义了简单事件、复杂事件等具体概念,为事件识别提供了通用的表达形式;在此基础上,构建了基于Bayes分类器与逻辑约束相结合的基本事件识别方法及基于隐马尔可夫模型的复杂事件识别方法。通过对实际场景中交通事件的识别验证了本文方法的有效性。本文的研究成果深化了交通事件信息视频检测理论与方法,为运动目标的交通行为特性研究、智能化交通事件信息采集提供了有效的技术手段,具有重要的理论意义与实用价值。更多还原

【Abstract】 In recent years, video detection technology is applied more and more in the field of intelligent transportation. Traffic event recognition based on video is one of the most promising applications in it. Traffic event automatic recognition technology could provide basic information for urban intelligent transportation management and control, and play an important role in easing traffic jams, reducing traffic accidents and ensuring travel safety. Currently, research of intelligent traffic event recognition system based on video processing is still at the exploratory stage that many key technical issues remain to be resolved. In this paper, key techniques of the process of video-based traffic event recognition are studied systematically, including moving object detection, object recognition and tracking, traffic behavior analysis and traffic event recognition.Moving object detection is the basic component of video detection and surveillance, which provides essential resource for behavior analysis and event recognition. Background model is crucial to obtain the background image for detecting moving targets effectively. Generally, background model is consisting of three parts: background initialization, background representation and background update. Among them, background initialization is the premise of background representation and update, while background representation and updating is the basis to maintain and update background image in complicated scenes for a long time. A background initialization algorithm based on the stable interval sequence searching is used in the paper. All stable non-overlapping intervals in the temporal sequence of each pixel are detected to obtain probably backgrounds first, and then background set is constructed with the pixel value variable-constrained to realize background initialization. From experiments it can be seen that the algorithm proposed could overcome "pseudo-background" generated from the slow motion of large-scale targets, as well as the accurate initial background could extracted when foreground coverage is more than 50% in the training sequence. Based on background initialization, realtime background is represented by Gaussian Mixture Model whose parameters are solved by EM algorithm. In addition, for realizing long-term background maintenance, the object-level background update algorithm is presented with the consideration of foreground motion regions information detected. This algorithm can solve the problem that suspended foreground objects become one part of the background during background update process. Foreground motion regions detected are often including moving shadow which has a significant impact on feature representation, classification and tracking of moving targets. So object-level moving shadow detection model based on RGB color variable degree is utilized to eliminate moving shadow of multi-objects effectively. Tests through several video sequences obtained from realistic scenarios show that proposed methods have good robustness and self-adaptability. This part of research enriches theories and methods of video detection and surveillance, as well as lays the foundation for traffic event recognition based on video processing.Moving objects’categories and spatio-temporal motion information could be obtained from object classification and tracking. In the aspect of the moving object recognition and classification, two basic issues are focused on: feature selection and representation, as well as classification model choice and learning. In order to provide information of target types of mixed traffic, a simple effective feature representation algorithm based on‘centro-bias’moment is proposed. Centro-bias moments feature has the invariability of rotation, translation and scale, which can overcome the influence of the moving status and dynamic environment. Meantime, object velocity is extracted as the motion feature. Combination of these two representation feature, multi-class support vector machine (SVM) is used to construct optimal hyperplanes for classifying moving object into vehicles, bikes and pedestrians. Experimental results show that the classification accuracy rate can reach 89.4%. In the aspect of motion tracking of multi-class objects in mixed traffic, multi-feature matching method based on Kalman filter is developed, which is combined with motion feature, shape feature and color feature. Additionly, to solve the problem of temporary trajectory missing caused by occlusion in the process of tracking, occlusion handling method based on historical motion information compensation is proposed. The method ensures an accurate estimate of the state of motion in the complex environment. The proposed methods are validated under different traffic scenes. Results show that the proposed tracking method is robust and adaptive, and has a good real-time property that the processing speed is below 0.02 seconds / frame. This part of research provides effective technical means to obtain object feature information for behavior analysis and event recognition.Traffic objects motion pattern and abnormal behavior can be obtained by trajectory distribution learning. In order to solve limitations that most exist methods of behavior pattern recognition rely on spatial characteristics, multi-level trajectory pattern learning algorithm is presented considering spatial characteristic, orientation characteristic and type characteristic. First, improved Hausdorff distance measure approach is utilized to construct spatial similarity matrix of a collection of traces and spectral clustering is used to realizing spatial pattern learning. Secondly, distribution of trace begin-points and end-points are fitted by GMM model. Orientation pattern can be learn From this. Thirdly, trace type pattern is obtained by hierarchical clustering algorithm with object categories. Multi-level trajectory pattern learning algorithm proposed is tested with realistic scenarios videos and good performances are showed in the experimental results. Abnormal behavior detection algorithms based on spatial pattern matching and orientation pattern matching are proposed respectively. Traffic behaviors of lane-changing and reverse-driving are detected effectively through these algorithms. This part of research provides technical support to study behavior characteristics of traffic targets.Traffic event recognition not only depends on reasoning and analysis of object behavior, but also is closely related with context information. If there is a lack of context information, the meaning and content of event can’t be described accurately. Hence, traffic event representation and recognition method based on context is developed and the concept of context information in traffic event is defined. Context is divided into spatial context, temporal context, object context and special parameter context according to the event content. And then, event unit is constructed with object’s property and one context information. Based on this, a common semantics representation form is realized involving basic event and complex event. For traffic event recognition, basic event recognition method based on Bayes classifier combined with logical restriction and complex event recognition method based on HMM model are developed. Events of pedestrian illegal crossing and temporary parking are recognized effectively through our approaches. This part of research provides a theoretical guidance and reference for traffic event recognition system.In summary, the achievement of our research results deepens the theories and approaches of traffic event information video detection, and provides some guidance and reference meaning for the follow-up studies. On the other hand, the research has an important application value that it provides effective technical support for traffic behavioral characteristics research and intelligent traffic event information collection.更多还原