摘要
针对视觉监控中基于运动轨迹的目标行为分析问题,提出了一种基于隐马尔科夫模型(HMM)聚类的轨迹分布模式提取和异常行为检测算法。首先为每一条运动轨迹训练一个HMM,并通过这些模型来计算轨迹两两之间的距离;然后对该距离矩阵采用主元分析法(PCA)降维并以降维后的每一行作为对应轨迹的特征进行模糊C均值聚类,接着为聚类后的每一类轨迹训练一个HMM作为其分布表达模型;最后利用这些HMM模型来检测给定轨迹所表示的目标行为是否异常。对不同场景的轨迹分析实验表明了方法的有效性。
A novel motion trajectory pattern learning and anomaly detection method based on HMM clustering was put forward for the problem of moving target＇s behavior analysis in visual surveillance system.Firstly,one HMM was trained for each trajectory in the training set and the pair-wise distance between the models was calculated to measure the difference of the trajectories;then,the rows of the pair-wise distance matrix,after processed by PCA,were taken as the features of the corresponding trajectories and clustered through fuzzy C-means method and different HMMs were trained for each cluster of the trajectories to represent their distribution patterns;finally a mechanism was given to detect the anomaly through the learned models.The experiment on the trajectories of different scenes shows its effectiveness.
出处
《华北电力大学学报（自然科学版）》
CAS
北大核心
2010年第6期90-94,共5页
Journal of North China Electric Power University：Natural Science Edition
基金
中央高校基本科研业务费专项基金资助(10QG21)
关键词
轨迹模式分析
隐马尔科夫模型
异常检测
视觉监控
聚类
trajectory pattern analysis
hidden Markov model
anomaly detection
visual surveillance
clustering
