Research On Trajectory Anomaly Detection Method Based On Spatio-temporal Information

With the development of GPS,wireless sensors and other positioning devices,trajectory data mining has become an increasingly important research topic,of which trajectory anomaly detection is an important direction in trajectory data mining,and has wide applications in urban traffic detection and road network planning,disaster weather warning,social public health and safety management,and other fields.However,the traditional metric-based trajectory anomaly detection algorithm does not fully consider the time dimension,location noise and sporadic sampling of trajectories.Based on the manual extraction of trajectory features,the design of features is highly dependent on expert knowledge,and the features often need to be reselected and optimized for different application scenarios.To address the above problems,this paper conducts research in several aspects,such as improving anomaly detection accuracy,reducing algorithm scale and anomaly localization.The main research of this paper includes the following points.(1)In this paper,a trajectory anomaly detection method based on spatio-temporal similarity is proposed.Firstly,the city is divided into equal-sized grids,and the first-order Markov model is used to model the probability that the trajectory of a moving object is located in a grid at a certain moment.Considering that the location information provided by GPS devices cannot accurately represent the specific location of the object,and a Gaussian probability distribution is proposed to simulate the real distribution of trajectory points considering the influence of noise on the trajectory points.And based on the velocity distribution of the trajectory itself,the transfer probability of the moving object between two locations is estimated.The average co-location probability of two trajectories is taken as the spatio-temporal similarity between trajectories.Finally,a trajectory clustering algorithm is used to extract representative trajectories and set weights according to the number of trajectory clusters to compare with the query trajectories to find abnormal trajectories.Experiments using multiple datasets show that the algorithm can discover anomalies in the temporal dimension,making the detection results more meaningful in practice.(2)This paper proposes an anomaly detection method based on the Seq2 Seq model,which first grids the pre-processed trajectory data to mine its temporal and velocity characteristics.Then reconstructs the corresponding high-sampled trajectories on this basis.To address the problem that different points(inflection points,non-inflection points)in the training process contribute equally to the model training,an attention mechanism based on point and velocity features is proposed for adaptive selection of key points and contextual factors for learning the trajectory representation.In order to measure the temporal and velocity differences between output cells and real cells,temporal,velocity weights are introduced for each grid cell.And based on this,a loss function based on spatio-temporal and velocity information is proposed.Experiments show that the average F1 value of the method reaches 0.876 and the training time is reduced by nearly half compared to other methods.