An Automatic Learning Model for Trajectory Outlier Detection
Qingwen Fu, Jiahui Zhu, Yuepeng Chen, Jintao Wan, Bin He · 2020
The rapid development of global positioning system has given birth to a large number of spatial-temporal data, and there are many outliers of points obviously in these trajectory data. It is very important to detect the outliers in the trajectory to improve the data quality and accuracy of trajectory mining. In this paper, we propose a trajectory outlier detection algorithm based on bi-directional long short-term memory model and attention mechanism. Firstly, an eight-dim eigenvector is extracted from each point of trajectory, and then a two-layer bi-directional long short-term memory model is constructed. Finally, representing the trajectory points in an interactive way which is called attention mechanism. The input of the model is the trajectory point with a certain length, and the output is the type of the trajectory point. The model can automatically learn the difference between the normal point and the adjacent abnormal point with motion features. Experimental dataset based on real trajectory data of taxi from Beijing, and results showed that the performance of this algorithm is significantly better than constant speed threshold method or classical machine learning classification. Especially the precision and recall reaches 0.93 and 0.90 separately, which proves the effectiveness of this algorithm.