Research On Moving Pattern Discovery Method In Trajectory Data Based On Deep Learning

In recent years,with the continuous development of new technologies such as the mobile Internet and the Internet of Things,a large amount of spatio-temporal trajectory data has been generated,which contains the mobility of moving objects.Mining the moving patterns hidden in the trajectory data can find novel and valuable information and rules,and improve its application value in public transportation,environmental monitoring and public safety.The trajectory data has spatio-temporal attributes and multi-source.Traditional machine learning methods often do not comprehensively consider spatio-temporal attributes and other features of trajectories.Therefore,how to quickly and effectively,automatically and accurately extract useful information from the trajectory data,and find the implicit moving pattern in the trajectory data is very important for trajectory research.In this thesis,the trajectory data of the moving object is taken as the research object,the deep learning is taken as the research method,the discovery of the moving pattern in the trajectory data is taken as the research target,and the moving pattern is discovered and studied from the three perspectives of trajectory clustering,trajectory classification and traffic flow prediction.The research content is as follows:Trajectory clustering clusters similar trajectories into one cluster.The existing trajectory clustering method generally extracts mobile behavior features capable of representing trajectory data,and then clusters the trajectories by a clustering method based on similarity metrics.However,one obstacle hindering their wide usage is that limited sensor devices,communication errors,sensor errors,or sensor vacancies can cause trajectory data to be noisy or missing.In this regard,this thesis proposes a Robust Deep Attention Auto-encoder model(Robust DAA)to solve the noise problem of trajectory data in trajectory clustering,and obtain high-quality denoising low-dimensional feature representation.Specifically,the model introduces the attention mechanism into the traditional deep auto-encoder to form a deep attention auto-encoder,which can enhance feature propagation and feature selection.The deep attention auto-encoder are trained by the proximal method,back propagation and the Alternating Direction of Method of Multipliers(ADMM),so that Robust DAA can further reduce the influence of noise in the trajectory data.Finally,the obtained low-dimensional denoising representation is input into the traditional clustering algorithm to obtain the clustering result.In this thesis,experiments are carried out on artificial and real datasets.The experimental results show that Robust DAA is superior to the current model in accuracy,precision,recall and f1-score.Although there have been some studies on trajectory classification,yet they either require manual feature selection or fail to fully consider the impact of time and space on classification results,it cause reduction of the classification effect,which is not suitable for the task of this thesis.In order to solve the above problems,this thesis proposes a Deep Multi-scale learning Network(MslNet).MslNet is built from different spatial and temporal dimensions,so that the effects of time granularity and spatial granularity on trajectory classification can be fully considered.Finally,the feature representations of the two parts of the model are merged,and the final classification result is output.The model designed in this thesis is based on the latest image classification network structure DenseNet,incorporates attention mechanism and residual learning.This model is able to fully capture local and spatial features so as to enhance feature propagation and capture long-term dependence.Moreover,the number of network structure parameters is also reduced.This thesis evaluates the model on two real data sets(Geolife,Ningbo AIS data).The results show that MslNet is superior to the state-of-the-art models in accuracy,precision,recall and f1-score.Furthermore,the classification results can help to understand mobility accurately.The research of traffic flow prediction is mainly to predict traffic flow between regions at the city level.As traditional time series forecasting models focus on only temporal attributes,they are unable to predict traffic flow accurately.Nowadays,the state-of-the-art traffic flow forecasting model has taken into account spatio-temporal attributes and other factors that affect the traffic flow.The weakness is that it makes model complicated and less universally applicable.In particular,when some data is missing,the prediction results will be quite unreliable.To solve this problem,this thesis proposes a Deep Spatio-Temporal ResNet-LSTM Network(DSTRL-Net)structure combining Long Short-Term Memory(LSTM)and Deep Residual Network(ResNet)to predict the traffic flow between regions.This thesis designs an end-to-end structure of ResNet based on unique properties of trajectory data,which is used to mine and process spatial properties of trajectory data,and uses LSTM to process temporal attributes.It should be noted that ResNet and LSTM are only used to find how the traffic flow in a region is affected by space and several recent time intervals.Finally,for the results of dealing with temporal and spatial attributes models,this thesis proposes a threshold-based fusion algorithm that outputs the traffic flow between regions.Experiments are conducted on traffic flow data in Chengdu,New York City and Ningbo.The results show that DSTRL-Net is superior to currently well-known methods in accuracy and versatility.