IBTOD: An Isolation-Based Method to Detect Outlying Sub-Trajectories on Multi-Factors
Kaixi Hu, Pan Duan, Bei Hu, Qichang Duan · 2018
With the increasing use of location-aware devices, massive spatial trajectory data can be generated. As a result, researchers devoted their effort to study on trajectory outlier detection. In most existing work, researchers focus their eyes on the spatial information of the tracing points. However, people have more interest on the outlying sub-trajectories instead of the points. Besides, temporal information is also a significant factor. The movement of an object always relates to its velocity and we can find more useful results by taking temporal information into consideration. In our research, we first propose a method to partition a long and complicated trajectory into a series of base units in spatial and temporal levels. Second, we use a k-nearest neighbor queried strategy to transform the abstract relationship into a concrete form and obtain a set of features. Finally, we use an iForest model to detect outlying sub-trajectories that are susceptible to a mechanism called isolation. Based on these three steps, we develop an isolation-based trajectory outlier detection algorithm IBTOD and use box plot to set the parameters automatically. When verifying our algorithm with different trajectory data and comparing with TRAOD algorithm, the detecting results show the efficiency of IBTOD.