车辆轨迹的增量式建模与在线异常检测

Incremental modeling and online anomaly detection of vehicle trajectories

针对智能交通系统中车辆轨迹自动异常检测问题,提出一种基于批处理(batch-mode)模型初始化的增量式轨迹建模,并将其应用到在线异常检测。首先采用改进的Hausdorff距离和谱聚类对初始轨迹集进行分类并建立初始轨迹模型库;然后对提取的新轨迹进行在线异常检测以及轨迹识别,通过增量式(incremental)EM算法更新轨迹类别的隐马尔可夫模型参数;最后进行模型结构更新。户外实际场景监控视频实验结果表明,与经典的batch-mode算法相比,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,同时该算法在异常检测方面具有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。

For vehicle trajectory automatic anomaly detection in intelligent transport system,this paper proposed an incremental trajectory modeling method based on batch-mode model initialize,and applied to online anomaly detection. Firstly,the proposed method utilized improved Hausdorff distance and spectral clustering algorithm to cluster the initial trajectories,and established the initial trajectory models. Then,it performed online anomaly detection and captured the trajectory class recognition once a new trajectory,and updated the hidden Markov model parameters with incremental EM algorithm. Finally it updated the model structure by trimming mixture component. Experiment on outdoor real world surveillance video demonstrates that,comparing with the classic batch-mode algorithm,incremental trajectory modeling can get more accurate trajectory models and faster speed. At the same time,the proposed algorithm has higher detection rate and lower false alarm rate. The algorithm realizes online anomaly detection and it is insensitive to the initial trajectories.