Research On Trajectory Classification Method Based On Deep Learning

With the rapid development of mobile Internet and mobile positioning technology,a huge amount of trajectory data is available.Trajectory classification is a crucial method in spatio-temporal data mining.In recent years,some scholars have already applied deep learning to trajectory classification and made some achievements.However,the main difficulty still lies in processing temporal and spatial information in trajectory data simultaneously.This thesis aims at improving the accuracy of trajectory classification from the perspective of enhancing spatio-temporal information and building neural networks,and also studies the supervised and unsupervised methods for trajectory classification due to that more datasets have no label.The main works of this thesis are listed as follows:1.This thesis studies the method of enhancing spatio-temporal information.In terms of the behavior patterns of moving objects presented by trajectory motion features are not comprehensive and accurate enough,this thesis proposes two auxiliary features to enhance the spatial information,including stop state and turn state.In order to make the temporal information and periodic features clearer,a recurrence plot based method for temporal information enhancement is proposed.2.For the supervised classification task,this thesis first builds a dual autoencoder to preliminarily learn the high-level features of input data by considering the structure of input data.Second,based on the predefined class centroids generation algorithm,a classification layer and its corresponding calculation of loss function are proposed toward the distribution in high-level data space,which can make the high-level future more helpful for supervised classification.For the unsupervised classification task,an unsupervised model is improved from the supervised model.Also based on the predefined class centroids generation algorithm,two loss functions of various statistical distances are introduced to build a clustering layer,which can ensure the model learns the features that contribute more to unsupervised classification.3.As for model training,the two-stage model training method of pretraining followed by joint training is proposed.The training strategy of dynamic loss function weights is also put forward,which enables the training process to converge faster while the model has good performance.Besides,for two different classification tasks,different strategies are developed to make the model stop training at a proper time.4.To find out if the proposed methods effective or not,extensive comparison experiments are conducted on two real-world trajectory datasets,including Geolife and SHL datasets.Besides,ablation experiments are also conducted,along with the analysis of model complexity and sensitivity of data sampling rate.Overall,for the supervised task of classifying trajectories into 5 classes,the highest supervised accuracy for the Geolife and SHL datasets can be 89.5% and 89.6%,respectively.As for the unsupervised task of classifying trajectories into 3 classes,the highest unsupervised clustering accuracy for both datasets can be 64.1% and 83.5%,respectively.All results show that the proposed method has various improvements to recent representative methods.