视频序列下的车辆轨迹异常行为识别
发布时间：2019-01-26 18:11
【摘要】：交通管理系统是对交通系统中包括人员、车辆、道路及环境等要素的统筹系统,交通系统中的要素一旦发生异常,就会导致拥堵或事故。及时获取交通信息上报交通管理系统处理,能够降低拥堵或事故的发生率,交通监控系统是交通管理系统对交通要素实时信息获取的重要手段。基于视频的车辆异常行为分析作为智能交通监控系统的核心技术,对于提高道路运行效率、降低事故率及保障交通人生命财产安全有有重要的意义和应用价值。本文以轨迹作为车辆行为信息的载体,围绕车辆异常行为识别这个目的,进行了运动车辆检测、跟踪及异常轨迹识别的研究。针对交通监控中无法采集大量负类样本,并且需要实时检测的特点,本文提出基于自适应单类支持向量机的车辆异常行为检测方法。该方法将车辆轨迹映射为高维空间中的向量点,选取一段正常车辆轨迹,由支持向量机算法学习该样本轨迹后建立支持向量模型,实现对待检测轨迹进行异常监测。 本文所做的工作总结起来主要为以下几个方面： (1)针对交通场景中实时性及精确性的要求,采取混合高斯模型对监控场景进行背景建模提取运动车辆,较好地抑制了噪声,并且检测目标精确。 (2)利用MeanShift算法对检测到的运动车辆实现轨迹跟踪,为车辆轨迹异常行为识别获取基础信息。 (3)利用改进的Hausdorff距离和比较置信度度量轨迹间的带权相似度,并用谱聚类算法对其进行聚类,获得场景中车辆运动行为模式。 (4)以车辆轨迹为研究目标,用单类支持向量机对其进行异常识别,并在单类支持向量机中加入自适应参数,实现支持向量模型的实时更新,以满足长期监控的需求。
[Abstract]:Traffic management system (TMS) is an integrated system which includes personnel, vehicles, roads and environment. Once the elements of traffic system are abnormal, it will lead to congestion or accidents. Getting traffic information in time to report to traffic management system for processing can reduce the incidence of congestion or accident. Traffic monitoring system is an important means for traffic management system to obtain real-time information of traffic elements. As the core technology of intelligent traffic monitoring system, video based abnormal behavior analysis of vehicles has important significance and application value in improving the efficiency of road operation, reducing the accident rate and ensuring the safety of people's lives and property. In this paper, the track is used as the carrier of vehicle behavior information, and the research of moving vehicle detection, tracking and abnormal track recognition is carried out around the purpose of vehicle abnormal behavior recognition. In view of the fact that a large number of negative class samples can not be collected in traffic monitoring and need to be detected in real time, an adaptive single-class support vector machine based vehicle anomaly detection method is proposed in this paper. In this method, the vehicle trajectory is mapped to a vector point in the high-dimensional space, a normal vehicle trajectory is selected, and the support vector model is established after learning the sample trajectory by using the support vector machine algorithm, and the abnormal detection trajectory is monitored. The main work of this paper is summarized as follows: (1) aiming at the requirement of real-time and accuracy in traffic scene, the mixed Gao Si model is used to model the background of the monitoring scene to extract the moving vehicle. The noise is suppressed and the target is detected accurately. (2) MeanShift algorithm is used to track the track of the detected moving vehicle, and the basic information is obtained for the identification of the abnormal behavior of the vehicle trajectory. (3) using the improved Hausdorff distance and the comparative confidence degree to measure the weighted similarity between the trajectories, we use the spectral clustering algorithm to cluster it, and obtain the vehicle motion behavior pattern in the scene. (4) taking the vehicle trajectory as the research object, the single class support vector machine (SVM) is used to identify the anomaly, and the adaptive parameters are added to the single class support vector machine to realize the real-time updating of the support vector model to meet the needs of long-term monitoring.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U491;U495
本文编号：2415762
[Abstract]:Traffic management system (TMS) is an integrated system which includes personnel, vehicles, roads and environment. Once the elements of traffic system are abnormal, it will lead to congestion or accidents. Getting traffic information in time to report to traffic management system for processing can reduce the incidence of congestion or accident. Traffic monitoring system is an important means for traffic management system to obtain real-time information of traffic elements. As the core technology of intelligent traffic monitoring system, video based abnormal behavior analysis of vehicles has important significance and application value in improving the efficiency of road operation, reducing the accident rate and ensuring the safety of people's lives and property. In this paper, the track is used as the carrier of vehicle behavior information, and the research of moving vehicle detection, tracking and abnormal track recognition is carried out around the purpose of vehicle abnormal behavior recognition. In view of the fact that a large number of negative class samples can not be collected in traffic monitoring and need to be detected in real time, an adaptive single-class support vector machine based vehicle anomaly detection method is proposed in this paper. In this method, the vehicle trajectory is mapped to a vector point in the high-dimensional space, a normal vehicle trajectory is selected, and the support vector model is established after learning the sample trajectory by using the support vector machine algorithm, and the abnormal detection trajectory is monitored. The main work of this paper is summarized as follows: (1) aiming at the requirement of real-time and accuracy in traffic scene, the mixed Gao Si model is used to model the background of the monitoring scene to extract the moving vehicle. The noise is suppressed and the target is detected accurately. (2) MeanShift algorithm is used to track the track of the detected moving vehicle, and the basic information is obtained for the identification of the abnormal behavior of the vehicle trajectory. (3) using the improved Hausdorff distance and the comparative confidence degree to measure the weighted similarity between the trajectories, we use the spectral clustering algorithm to cluster it, and obtain the vehicle motion behavior pattern in the scene. (4) taking the vehicle trajectory as the research object, the single class support vector machine (SVM) is used to identify the anomaly, and the adaptive parameters are added to the single class support vector machine to realize the real-time updating of the support vector model to meet the needs of long-term monitoring.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U491;U495
【参考文献】
相关期刊论文 前4条
1 陈继东;孟小峰;赖彩凤;;基于道路网络的对象聚类[J];软件学报;2007年02期
2 王孝艳;张艳珠;董慧颖;李媛;李小娟;;运动目标检测的三帧差法算法研究[J];沈阳理工大学学报;2011年06期
3 袁冠;夏士雄;张磊;周勇;;基于结构相似度的轨迹聚类算法[J];通信学报;2011年09期
4 李英姿,张飞舟,林耀海;智能交通系统中地理信息系统的研究[J];中国公路学报;2000年03期
，本文编号：2415762
本文链接：https://www.wllwen.com/kejilunwen/jiaotonggongchenglunwen/2415762.html
