增量式建模下的车辆轨迹识别与在线异常检测研究

发布时间：2018-11-27 18:29

【摘要】：车辆轨迹识别与在线异常检测是智能交通系统中的一个重要研究方向,它对现实生活中发生的交通事故及时有效地救援和后期处理具有很大的辅助作用;同时它能减少因为交通事故引起的交通延误和二次事故;它还可以为城市交通监控与安全管理提供重要的信息凭证。通常情况下,车辆轨迹识别与在线异常检测是基于轨迹建模的方法进行的,而轨迹建模方法分为统计模型和运动模型两种。无监督学习是近几年提出的一种新的轨迹建模方法,该方法能在训练数据中排除异常轨迹,并且在训练数据比较多的情况下有效地进行轨迹识别及异常检测,但是对于正常收集的含有异常轨迹的初始轨迹集及含有少量轨迹的初始轨迹集,该方法的识别率及异常检测精度比较低。针对这些问题,本文将增量式(incremental)EM算法应用到无监督轨迹建模上,提出一种基于批处理(batch-mode)模型初始化的增量式轨迹建模方法,并将其应用到车辆轨迹识别与在线异常检测上。首先采用改进的Hausdorff距离测量初始轨迹集中轨迹之间的相似度,再用谱聚类算法对初始轨迹集进行聚类,提取轨迹的分布模式。对于初始轨迹集中聚类完成的每一类样本轨迹,利用多观察序列训练方法建立每类轨迹的隐马尔科夫模型,依此来得到初始轨迹模型库。对于从视频图像中提取的新轨迹,使用最大后验估计寻找新轨迹最可能的正常轨迹类,再用自动阈值法进行在线异常检测,然后识别新轨迹所属的轨迹模型,通过incremental EM算法更新新轨迹所属类别的隐马尔科夫模型参数。为了增加本文方法的适应性,最后需要对模型结构进行更新。应用本文方法对户外实际场景拍摄视频进行测试,实验结果表明,与经典的无监督批处理轨迹建模算法相比,本文方法即增量式轨迹建模可以得到更加准确的轨迹模型库、更快的运算速度;该算法在初始轨迹集包含异常轨迹的情况下能大幅度提高轨迹识别率;在异常检测方面得到了更高的检测率和更低的虚警率,实现了在线异常检测,具有对初始轨迹集不敏感的特点。
[Abstract]:Vehicle trajectory identification and on-line anomaly detection are important research directions in intelligent transportation system. They can help traffic accidents in real life to rescue and deal with them in a timely and effective manner. At the same time, it can reduce traffic delays and secondary accidents caused by traffic accidents, and it can also provide important information evidence for urban traffic monitoring and safety management. Generally, vehicle trajectory identification and online anomaly detection are based on trajectory modeling, and trajectory modeling methods are divided into statistical model and motion model. Unsupervised learning is a new trajectory modeling method proposed in recent years. This method can eliminate abnormal trajectory in training data, and can effectively identify and detect the trajectory when there are more training data. However, the recognition rate and the detection accuracy of the method are low for the normal collection of initial trajectory sets with abnormal trajectories and the initial trajectory sets with a small number of tracks. To solve these problems, this paper applies incremental (incremental) EM algorithm to unsupervised trajectory modeling, and proposes an incremental trajectory modeling method based on batch processing (batch-mode) model initialization. It is applied to vehicle track recognition and online anomaly detection. Firstly, the improved Hausdorff distance is used to measure the similarity between the trajectories of the initial locus, and then the spectral clustering algorithm is used to cluster the initial locus to extract the distribution pattern of the locus. The hidden Markov model of each kind of trajectory is established by using the training method of multiple observation sequences for each kind of sample locus which is clustered by the initial locus, and the initial locus model library is obtained according to this model. For the new trajectory extracted from the video image, the maximum a posteriori estimation is used to find the most probable normal trajectory class, and then the automatic threshold method is used for on-line anomaly detection, and then the trajectory model to which the new trajectory belongs is identified. The hidden Markov model parameters of the new trajectory are updated by incremental EM algorithm. In order to increase the adaptability of this method, we need to update the model structure at last. The experiment results show that, compared with the classical unsupervised batch trajectory modeling algorithm, the incremental trajectory modeling method in this paper can get more accurate trajectory model base. Faster computing speed; The algorithm can greatly improve the trajectory recognition rate when the initial trajectory set contains abnormal trajectory. Higher detection rate and lower false alarm rate are obtained in anomaly detection, and on-line anomaly detection is realized, which is insensitive to the initial trajectory set.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】