【作者】 张波；

【导师】 陈金玉；

【作者基本信息】 重庆大学 ， 控制科学与工程， 2013， 硕士

【摘要】 违章停车事件是视频交通监控系统监控的一类重要异常事件，尤其在具有车流量大、行车速度快等特点的高速公路运营管理中，一旦发生停车事件后，若不及时处理，极易引发交通拥堵甚至二次事故，严重影响高速公路的正常运营，而复杂场景下由于视频序列图像中干扰信息较多，这就可能使得检测算法的检测准确率大大降低，甚至可能完全失效。从当前基于视频的停车检测算法的研究现状来看，算法对场景环境的自适应性差，而实际的应用场景比较复杂，如光照变化、车辆阴影以及隧道中车辆灯光、摄像机拍摄范围偏差等，因此研究复杂场景下的道路停车检测具有重要的理论和现实意义。本文针对上述问题，结合隧道及露天典型道路场景的特点，重点从运动目标检测和停驶车辆识别两个方面，展开对复杂场景下基于视频的停车检测关键技术的研究，包括：背景建模、背景更新、停驶车辆识别等。在背景建模以及背景更新方面，结合统计直方图法和统计中值法，提出了一种统计中值邻域最大值法。该方法不仅能有效过滤掉车辆灯光对路面背景的影响，而且使用面积中值法能够获得更接近理论中值的实际中值点，在提高场景背景建立效率的同时，建立更可靠的背景图像。此外，考虑实际的交通场景通常比较复杂，为了能够使当前背景能够始终真实代表当前场景，结合分块处理以及形态学滤波法，建立一种基于分块处理的背景重建与更新算法，在场景突变或者缓慢变化时，该算法能够及时更新背景。在停驶车辆识别方面,为了有效消除运动前景对背景像素点的影响，提出了一种基于像素时间序列特征的稳态分析法，分析背景像素稳态变化，准确检测静止目标；在此基础上，为了过滤掉实际交通场景中与停车状态特征相类似的非机动车对象（如抛洒物），通过获取目标对象在检测区域的大小以及轮廓规则度等特征，引入占空比和圆度来准确识别出车辆。最后，利用典型的交通场景——隧道场景和高速公路露天场景验证本文提出的停车检测算法。实验结果表明，提出的算法在隧道场景中正检率可达到95%以上，而在高速公路露天场景中正检率达到90%以上，与其他停车检测算法相比较，在减少计算量、提高系统实时性的同时，具有较好的正检率。更多还原

【Abstract】 Illegal parking is an important class of abnormal traffic incidents in highway videomonitoring system. Once the illegal parking event happens, it will easily lead to trafficjams and even secondary accident without timely processing due to the heavy trafficand high-speed cars in the motorway. And that would seriously affect the normaloperation of the speedway. Furthermore, there is more interferential information in thevideo screen of complex scenes, which is possible to decrease the accuracy rate ofdetection, and even to be completely ineffective. The existing illegal parking detectionalgorithms based on monitoring system are just for some unique scenes and they cannotbe well adapted to the environment. In fact, the actual application scenarios are muchmore complex, including bad weathers, interference of car lights and other light sources,camera shakes, barriers between the vehicles and so on. Therefore, to research the roadparking event detection in complex scenes has important theoretical and practicalsignificance.To solve these problems, combined with the characteristics of the typical tunneland open road scenes, this paper focuses on two aspects, which are moving targetdetection and stopped vehicle identification, to research the key technologies of thedetection, background modeling, background updating, stopped vehicle identificationand so on.On the aspect of background modeling and updating, this thesis brings up onemethod of statistical neighborhood maximum value, which combines with the methodof statistical histogram and the method of statistical median. The method not only canfilter out the impact of the vehicle lighting on the background, but also can establish areliable background through obtaining the median pixel, what’s more, it can improveefficiency more effectively by this method. In addition, it is necessary to update thebackground in real time since the complex environment always change in reality. Someexisting background update algorithms even make the vehicles which have stopped forlong time updated to the background. To solve these problems, this thesis will utilize abackground update method based on block images, which is also need themorphological filtering as an auxiliary.On the side of identifying stopped vehicles, to reduce the effect of the moving foreground to the pixels of background, this thesis comes up with a method to analyzethe changes of background pixels which is one steady-state analysis based on timeseries features of pixels. Furthermore, it can detect the stationary targets much moreeffectively. On this basis, use the special characteristics of vehicles, like the contourfeatures and duty ratio to identification vehicles, which can tell different spilled materialapart.In the thesis, two typical traffic scenes, tunnel and open road scenes of highway areused to verify the parking detection algorithm. The results demonstrate that the positivedetection rate in the tunnel scene can reach more than95%, which also can do morethan90%in the open road scene of the motorway. Compared with the other algorithms,it not only reduces computations and improves the system real-time, but also with betteraccurate detection rate.更多还原