Deep Learning-Based Outliers Detection in Compressed Trajectories

Nowadays anomalous trajectory detection has a prominent place in many real-world applications. In this paper, firstly, we have proposed a comparison between two Deep Learning models, Gaussian Mixture Variational Sequence Auto-Encode (GM-VSAE) and Anomalous Trajectory Detection using Recurrent Neural Network (ATD-RNN) using a public taxi service dataset. The goal is to compare the performances of these two recent models which have addressed the problem differently, but which have already overcome traditional anomalous trajectory detection methods. Furthermore, we have dealt with trajectories compression which allows to reduce the size of data, to cut down the memory space and to improve the efficiency of transmission, storage and processing. Compression of trajectory data is crucial since currently there is an exponential increase of the amount of spatial and temporal information that trace a moving object's path. Therefore, we focused on the GM-VSAE model, which gave the best performance, and applied several compression algorithms to our trajectories data to evaluate the impact of the compression on the performance's metrics, such as the AUC and the execution time of GM-VSAE. Results show the superiority of Uniform Sampling compared to the other compression algorithms. Moreover, compressing data trajectories significantly reduced the training, and the evaluation time while keeping a relatively high value of the AUC (close to 1).