Trajectory outlier detection based on partition-and-detection framework

Liang Bao, Shanshan Wu, Weizhao Chen, Zisheng Zhu, Fan Yi

Abstract: Outlier detection has been a popular data mining task. However, there is a lack of serious study on outlier detection for trajectory data. In this paper, a trajectory outlier detection based on local outlier fraction algorithm (TODLOF) is proposed to detect outliers in the trajectory dataset based on the partition-and-detection framework. When partitioning a trajectory, a minimum description length principle (MDL) based method is adopted. The local outlier factor (LOF) is used as the basis for judging the outlier in the detection stage, which improve the accuracy of the anomaly detection. Finally, experiments were carried out, with hurricane trajectory data and animal migration data as inputs, to prove that this algorithm can detect anomaly trajectories efficiently. And an online version of this algorithm is also presented to meet the requirements of real-time application.