节点文献
混合交通视频检测算法研究
Research on Video Detection Algorithms in Mixed Traffic
【作者】 曲昭伟;
【导师】 王殿海;
【作者基本信息】 吉林大学 , 交通信息工程及控制, 2009, 博士
【摘要】 混合交通是我国交通的主要特征。交通信息采集是有效进行交通管理与控制的前提。视频检测技术可自动获取全面的交通场景信息,广泛应用于道路交通数据采集、交叉口控制、交通事故处理、远程视频监控等交通领域,是未来交通信息采集的主要手段。针对混合交通视频检测算法在背景初始化、背景模型更新、阴影检测和物体分类识别等方面存在的主要问题,本文进行了深入系统的研究,提出了一种基于聚类识别的背景初始化算法,能够获取具有运动物体的初始背景,可克服缓慢运动大型物体造成的影响,实现遮挡率大于50%的背景初始化;提出了基于对象级的混合高斯背景模型更新方法,克服了像素级混合高斯模型中长时间停车和交通拥挤等现象可能造成的背景模型不能有效更新、使运动物体成为背景的一部分的问题;提出了基于RGB颜色变化度的自适应阴影检测算法,克服了颜色特性阴影检测方法中固定阈值的缺陷,可根据当前目标的特点自适应地进行阴影检测;提出了基于矩向量的混合交通物体特征提取与表达方法,利用SVM多分类学习机制,建立了基于SVM的识别分类算法,满足了混合交通物体识别的需要。本文取得的上述成果,为进一步研究混合交通视频检测算法奠定了基础。
【Abstract】 Mixed traffic is the main characteristic of our country traffic. Traffic information collection is the precondition of the effective traffic management and controlling. Video detection technology can automatically obtain comprehensive information mixed traffic scenario, Can be widely used in road traffic data collection, intersection control, traffic processing, remote video surveillance and other traffic areas, and it will be the main means of collecting information in the future.At present, it has made significant research results at the traffic video detection technology in China and abroad, but there are still some problems to constrain the traffic video detection technology, mostly of them are: 1) Background initialization and update, currently background model were assumed initial background , it does not include the prospect of sport, and the background update process does not consider the prospect of exercise, it limits the use of background model conditions and application scope, to make background model update can not overcome the traffic congestion, signal control, the formation of long-stay parking and traffic objects slow movement on the impact of the background updates, which will make movement object update as the background, causing movement objects false; 2) Shadow detection algorithm, the objects with the movement shadow has the same moving peculiarity of objects, and the shadow region and move objects closely linked, it probably caused the mistaken classification of objects, however the shadow detection algorithm used a fixed pre-determined threshold for background detection at present, that is difficult to satisfy the needs of practical applications; 3) Camera calibration parameters, the camera parameter calibration can achieve the projection mapping of world coordinates and image coordinates, which is the basis for carrying out traffic movement objects analysis and determination, their algorithm directly affects the precision and accuracy of the traffic analysis; 4) Mixed traffic object recognition, the mixed traffic is the main characteristics of our country transport, mixed traffic classification is the basic and premise of mixed traffic information obtained, At present, domestic and international traffic video detection algorithm is mainly identify vehicle classification, it lacks methods to recognize mixed traffic objects, restricting the scope of application, and also difficult to satisfy the actual needs of mixed traffic in our country.For the main problems of mixed traffic video detection, this article has been made a systematic research on the background model, the shadow detection, camera calibration, mixed-traffic objects classification and recognition, the main study results are as follows: In order to obtain the effective movement initial background to make the background model has selection updated in accordance to object moving information, this paper presents the background of identification based on cluster initialization algorithms and object-level based on a mixture of Gaussian background model update method. Background initialization phase, which is using the variable sliding window detect each pixel smoothing of all non-overlapping sub-sequences, to obtain probably background; Then select each smooth middle sample point of sub-sequences to build a classification sequence set; Finally under the unknown category of unsupervised clustering to identify thought, access to background subset implementation background initialization. Background update stage, according to motion segmentation, object recognition, kalman tracking sports and other high-level semantic expression, access to the prospect of targeted movement object information to determine whether the current point is the prospect objects or not. If it is the prospect objects, it does not be updated, otherwise, in accordance with the pixel-level Gaussian background model updating, in order to achieve the target level based on a mixture of Gaussian background update. All different traffic video sequences experimental results show that, this method has good adaptable, it could overcome the effect of the slow movement of large objects, the prospects for the impact of blocking the implementation rate of more than 50% of the background initialization. The method overcomes the pixel-level mixed-Gaussian model of long-stay parking and traffic situation on the impact of the background update.In light of the shadow detection algorithm does not consider the characteristics of sports objects and adaptive threshold, and adopting fixed, uniform threshold segmentation exercise shadow problem, according to the movement objectives characteristics of RGB color changing degree, adapting to the shadow detection algorithm is based on the RGB color changing degrees.The algorithm based on region segmentation, access to regional sports objects; using the optical properties of the shadow to access the candidate region; in accordance with the method of hypothesis testing, statistics for each candidate region of the RGB color changing degree of the shadow, making the shadow of the regional expression of the candidate for the mixed Gaussian distribution; Using EM algorithm implementation Gaussian Mixture Parameter Estimation, to obtain the shadow of the Gaussian distribution region; In accordance with the principles of 2.5σadaptive define the shadow detection threshold, making the shadow detected implementation. The algorithm overcomes the color characteristics of the shadow detection method of defects fixed threshold, and in accordance with the characteristics of the current goal for adaptively Detected shadow. Under different traffic status Videos treatment effect shows that the method has good robustness and adaptability, it satisfies to the actual needs of Traffic Video Detection.For existing scene reconstruction algorithm, it has high computational complexity and robustness of the bad disadvantage, this paper uses a black-box based on the mapping transformation method has been applied to traffic video calibration tests quickly and easily achieve 2-D space measurement Express. The method firstly is based on the Harris operator sub-pixel level corner detection algorithm, to obtain the characteristics of angle reference points, and select Group 4 corner; Combined with the corresponding corner of the image coordinates and world coordinates, according to the linear mapping model of the image coordinates and world coordinates using the matrix equation to obtain the camera parameters. There is only one piece of template images, in calibration process, comparing with traditional methods, this this reduces the number of feature points with the template to simplify the calculated. Experimental results show that the method is stability, high accuracy, the characteristics of simple algorithms.As for the lack of mixed traffic object classification and recognition rapid and effective algorithms couldn’t satisfy to actual needs of mixed traffic in our country, this paper bring forward the method of mixed traffic object feature extraction and expression which is based on the moment vector, using SVM multi-classification learning mechanism, set up SVM-based identification classification algorithm. In the feature extraction phase, firstly to obtain objects focus, then determine the moving object contour points with the focus characteristics of the expression, to build sports moment vector, as the characteristics of movement objects; In studying classification phase, according to different objects Moment vector samples, using SVM learning mechanism to obtain samples of each type of optimal classification hyperplane; In the identification stage, using objects characteristics vector and samples of optimal classification hyperplane in study phase to obtain the movement objects classification and recognition. Different video sequences of objects of mixed traffic detection results show that the classification algorithm has good performance and mixed traffic can meet the needs of object recognition.In this paper, the mainly study results achieved on mapping traffic scenes, background initialization, background update model, shadow detection and identify and classification of mixed-traffic, which are the theoretical basis for further study mixed traffic video detection algorithms and achieve mixed traffic video detection technology.
【Key words】 Mixed traffic; Video detection; Background initialization; Background model; shadow Detection; Feature extraction and expression;