交通监控视频中的车辆异常行为检测

发布时间：2018-11-03 15:20

【摘要】：近年来,我国经济快速迅速,机动车保有量迅速攀升。汽车给我们带来了许多便利,但随之频发的道路交通事故却对人民的生命财产安全造成了巨大的威胁。当前对交通监控视频的主要用途是在发生事故后进行事件回顾,并且很大程度上要依赖人工检索来定位,这样只能做到交通事故的事后查看,不能提前预防。为了进一步规范车辆行驶,缓解交通拥堵,减少交通事故,交通监控视频中的车辆异常行为检测成为了目前智能交通领域研究的重点与难点,这将为人们的日常生活、社会的安定与和谐带来重要保障。本文对交通监控视频中的车辆检测、车辆跟踪、车辆行驶轨迹提取及车辆行为的异常检测技术进行了研究,之后对其中的研究难点提出了若干算法改进与解决方案。主要工作如下:为了对车辆目标进行跟踪,并进一步提取车辆轨迹,分析行驶行为,首先要做到从视频监控数据中检测车辆目标。本文在分析、对比现有运动目标检测算法后,针对交通监控视频的特点提出了一种阈值自适应的Surendra背景差分改进算法,并将其与三帧差法结合,进行运动车辆的检测。最后通过实验结果表明,该改进算法能够结合背景差分法与帧差法的优点,抗环境干扰能力强,并且兼顾交通监控系统实时性与稳定性的要求,还原了车辆的真实目标区域,为车辆跟踪环节提供了车辆区域目标信息。现有CamShift算法能够实现视频中运动目标的跟踪,但存在需要手动选取跟踪区域与抗遮挡能力较差等问题。为了解决上述问题,实现跟踪效果的优化,本文将运动车辆检测结果输入CamShift算法初始步骤,并引入Kalman滤波器预测车辆的运动状态,提出基于Kalman滤波预测的CamShift车辆跟踪算法,缩小了下一帧中目标车辆的搜索范围,减小了CamShift算法的计算复杂度,并且分析了车辆因遮挡带来的跟踪失效问题,使用Kalman滤波器的预测值代替CamShift算法计算的目标位置,作为观测值去更新Kalman滤波器。实验证明,该改进算法能够有效抵抗由于目标遮挡而导致的跟踪失败。并且由于初始化搜索目标时使用了第三章车辆检测得出的结果,实现了运动车辆的自动跟踪。通过对车辆进行实时跟踪,可由目标外接矩形框获得车辆运动中心坐标。之后通过曲线拟合获取车辆运动轨迹。本文对轨迹数据进行了深入分析,提出了几种车辆运动行为的判别标准,包括车辆运动方向识别与车辆变道、调头、逆行等情况的判断。通过实验数据可以说明本章提出的判别方法能够广泛应用于车辆的违章识别,并且算法易于实现,稳定性较高。
[Abstract]:In recent years, China's economy is fast and the number of motor vehicles is rising rapidly. Automobile brings us a lot of convenience, but the frequent road traffic accidents pose a great threat to the safety of people's life and property. At present, the main purpose of traffic surveillance video is to review the incident after an accident, and to a large extent rely on manual search to locate, so that only after the traffic accident can be viewed, can not be prevented in advance. In order to further regulate vehicle driving, alleviate traffic congestion and reduce traffic accidents, the detection of abnormal behavior of vehicles in traffic surveillance video has become the focus and difficulty of the research in the field of intelligent transportation, which will be the daily life of people. Social stability and harmony bring important protection. In this paper, the techniques of vehicle detection, vehicle tracking, vehicle track extraction and abnormal detection of vehicle behavior in traffic surveillance video are studied. The main work is as follows: in order to track the vehicle target, further extract the vehicle trajectory and analyze the driving behavior, the first step is to detect the vehicle target from the video surveillance data. After analyzing and comparing the existing moving target detection algorithms, this paper proposes a threshold adaptive Surendra background differential algorithm for traffic surveillance video, and combines it with the three-frame difference method to detect moving vehicles. Finally, the experimental results show that the improved algorithm can combine the advantages of background differential method and frame difference method, and has strong ability to resist environmental interference, and can restore the real target area of vehicle by taking into account the requirements of real-time and stability of traffic monitoring system. It provides vehicle area target information for vehicle tracking. The existing CamShift algorithm can realize the tracking of moving targets in video, but there are some problems such as the need to select the tracking region manually and the poor ability to resist occlusion. In order to solve the above problems and optimize the tracking effect, this paper inputs the moving vehicle detection results into the initial steps of the CamShift algorithm, and introduces the Kalman filter to predict the moving state of the vehicle. A CamShift vehicle tracking algorithm based on the Kalman filter prediction is proposed. The search range of the target vehicle in the next frame is reduced and the computational complexity of the CamShift algorithm is reduced. The tracking failure caused by the occlusion is analyzed. The prediction value of the Kalman filter is used to replace the target position calculated by the CamShift algorithm. The Kalman filter is updated as an observation. Experiments show that the improved algorithm can effectively resist the tracking failure caused by target occlusion. And the automatic tracking of moving vehicles is realized by using the result of vehicle detection in Chapter 3 when initializing the search target. By tracking the vehicle in real time, the coordinate of the moving center of the vehicle can be obtained from the external rectangular frame of the target. After that, the moving track of the vehicle is obtained by curve fitting. In this paper, the track data are analyzed in depth, and several criteria for distinguishing the motion behavior of vehicles are put forward, including the identification of the moving direction of the vehicle and the judgment of the changing track, turning head, retrograde and so on. The experimental data show that the method proposed in this chapter can be widely used in the recognition of vehicle violations, and the algorithm is easy to implement and has high stability.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495

【相似文献】