Research On Real-time Detection Method Of Anomalous Behavior Based On Trajectory Data

With the development of mobile device and positioning technology and the improvement of the storage capacity,more and more spatio-temporal data can be collected and stored accurately,which provides a solid foundation for the development of applications based on spatio-temporal data.The trajectory data is a series of spatio-temporal data generated by moving objects.The analysis and mining of trajectory data can bring a series of conveniences to people's daily life,such as navigation,route recommendation,intelligent transportation and so on.Among them,the anomaly detection of trajectory data can find possible abnormal behavior from the massive data,which is of positive significance for regulating the behavior of the data producer.At present,there are some key points for anomaly detection of trajectory data.First,with the improvement of positioning technology and storage capacity,the volume of trajectory data is often very large.Therefore,it is necessary to have an efficient detection framework to deal with the massive trajectory data.Second,trajectory data is usually unlabeled data,and anomalous samples in data are often rare.Hence,the anomaly detection of trajectory can be summed up as a class of unbalanced sample classification problems,which needs an unsupervised or semi-supervised algorithm to deal with.Third,the anomaly detection of trajectory is often application specified.It means different applications need different algorithms to produce good performance.Finally,for the real scenario,the anomaly detection algorithm needs to be designed with the real-time characteristics,which can meet the needs of efficient detection online.In this paper,an unsupervised and a semi-supervised real-time anomaly detection algorithm are proposed for the two specific scenarios,taxi fraud behavior and dredging operation.Specifically,in the anomaly detection of taxi trajectory,a real-time anomaly detection algorithm based on the spatiotemporal law called STL is proposed.The algorithm learns two spatiotemporal models from historical trajectory data,one that represents the relationship between driving distance and displacement,and the other that represents the relationship between driving time and displacement.With the trained model,a point is detected as anomalous if the driving distance and driving time for the corresponding displacement is not within the normal range.In the anomaly detection of dredging operation,a dredging operation anomaly detection method based on AIS data called DOAD is proposed,which is to detect the anomalous dredging behavior during the whole operation.The method first builds a feature system to extract effective behavior features from the trajectory data.Then it uses a t-SNE based neural network and Gaussian mixture model to train a semi-supervised detection model.With the trained model,anomalous behavior can be detected in real time during the dredging operations.The experiment results show that the STL algorithm is more accurate and produces fewer false positives than the existing methods.The DOAD method can effectively fulfill the dredging operation detection task.