基于冗余分组的轨迹摘要算法

摘要

为了对视频监控设备采集到的轨迹数据进行聚类和异常检测,提出了一种新的轨迹摘要算法.使用了Jensen-Shannon Divergence(JSD)度量方法实现了轨迹数据的重采样,使得计算轨迹间相似度的准确性有所提高,并为后续滤波过程提供了必要的等采样点个数的轨迹数据;自适应地确定轨迹相似性的阈值,并采用非局部的思想,将轨迹数据进行冗余分组,同时识别出异常轨迹数据;从信号处理的角度对分组后的轨迹数据进行硬阈值滤波,经过合并得到摘要轨迹;此外,不受轨迹输入顺序的影响,并且提供了可视化的多尺度轨迹摘要结果.与具有噪声的基于密度的聚类(DBSCAN)算法的异常检测效果进行对比,所提算法在准确率(Precision)、召回率(Recall)以及F1指标上均有所提升.%In order to cluster and detect anomalies for the trajectory data collected by video surveillance equipment,a novel trajectory abstraction algorithm was proposed.Trajectories were firstly resampled by utilizing the Jensen-Shannon Divergence (JSD) measurement to improve the accuracy of similarity measurement between trajectories.Resampled trajectories in equal length,i.e.with the same number of sampling points,were required by the following non-local denoising.The similarity thresholds of the trajectory were determined adaptively,and the non-local means were used to cluster the trajectory data and identify the abnormal trajectory data.From the perspective of signal processing,the grouping trajectory data was filtered by the hard-thresholding method to get the summary trajector.The proposed algorithm was insensitive to the order of input trajectories and provides visual multi-scale abstractions of trajectory data.Compared with the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm,the proposed algorithm performs better in terms of precision,recall and Fl-mearsure.