基于轨迹聚类的公共安全异常检测

Anomaly detection of public safety based on trajectory clustering

公共安全异常检测的需求越来越迫切,监控中基于轨迹聚类的检测方法越来越流行,但是现有方法在处理高维不等长轨迹数据时效果并不理想。提出一个新的轨迹聚类方法,该方法通过组合动态时间弯曲和密度峰算法实现。动态时间弯曲用于度量轨迹间的距离,密度峰算法根据距离进行聚类。前者可直接度量不等长轨迹聚类,后者是近年提出的非球体分布数据聚类算法,以局部密度和最近邻聚类组合实现。实验在PETS2006监控视频数据集上进行,测试结果表明该方法有效地发现了异常的轨迹行为模式。

The demand of public security anomaly detection is becoming more urgent. The approaches based on trajectories clustering become more popular in surveillance, but the existing methods are not good at high-dimensional and unequal-length trajectories. So this paper presents a new approach to cluster trajectories by combining the dynamic time warping and density-peak algorithm. It measures the distance between trajectories by dynamic time warping, and then clusters trajectories by density-peak cluster algorithm. Dynamic time warping can be directly used to measure the distance of trajectories through non-uniform sampling. Density-peak algorithm is a recently proposed cluster algorithm for non-spherical distribution data by combining the local density and the nearest distance. Experiments are conducted on PETS 2006 surveillance video datasets, and results prove that the proposed approach has an effective ability to discover anomaly patterns.