Semantic Trajectory Prediction Based On Attention Mechanism

With the rapid development of positioning technology,many devices,such as mobile phones and devices based on Global Positioning System（GPS）.These devices can completely obtain the trajectory data generated by mobile objects（human,animal,vehicle,etc.）.In this case,it is becoming more and more necessary to predict the trajectory of moving objects,which is also of great research significance.In this thesis,semantic trajectory prediction is taken as the research issue.The main research work is as follows.（1）Trajectory Prediction Based on Attention MechanismIn trajectory prediction,the traditional long-Short Term Memory（LSTM）is usually used to construct the prediction method,but the prediction method has a sequence dependence on the time when calculating the state value of the hidden layer.In other words,the state value?of the previous moment is needed when calculating the state value?_-1 of the current moment,and it is neglected when predicting the position of the next moment.The influence of the state value of each cell in the hidden layer is different in output results.A trajectory prediction method based on Attention Mechanism（TPM-AM）is proposed in this thesis.Firstly,the original trajectory of the moving object is transformed into a trajectory sequence which is acceptable to the prediction model.The trajectory sequence is divided into trajectory sequences according to the number of the moving object.The trajectory sequence is divided into input vectors according to the size of sliding window and the length of time step,and the input vectors are input into the prediction model in batches.Secondly,SRU（Simple Recurrent Units）is used to construct the prediction model.When calculating the input trajectory vectors,SRU is a slightly cyclic unit,which makes each state dimension independent,solves the sequence dependence of LSTM on time,speeds up the calculation process and reduces the prediction time.Finally,Attention mechanism is applied to SRU to distribute the output weight of each unit in the hidden layer,because Attention mechanism not only calculates the matching degree of the state value of the hidden layer,but also conducts selective learning,thus improving the accuracy of prediction.Experiments on the open data set of the Metropolitan Transport Administration show that TPM-AM is 3.4 times faster than LSTM,and the accuracy is improved by 1.8%.（2）Semantic Trajectory Prediction Based on Spatial-Temporal-Semantic Graph Convolutional NetworkOn the basis of trajectory prediction based on Attention mechanism,the related research on semantic trajectory prediction is carried out.In trajectory prediction,most of the existing methods are based on the temporal and spatial characteristics of trajectories,ignoring other high-level semantic characteristics of moving objects,such as the speed,direction and state of moving objects.Semantic trajectories have more abundant information than ordinary trajectories.Moreover,traditional methods often neglect the interdependence between the spatial-temporal characteristics of trajectory data,especially when the amount of trajectory data is too large,the trajectory points of multiple moving objects will coincide at different times,in this case,the prediction accuracy will be reduced,and the prediction efficiency will also be reduced.Therefore,a Semantic Trajectory Prediction Method based on Spatial-Temporal-Semantic Graph Convolution Network（STPM-STSGCN）is proposed in this paper.Firstly,in order to extract the high-level semantic characteristics of trajectory more fully,the trajectory sequence is combined with relevant semantic information,such as the velocity,direction and motion state of the moving object embedded in longitude and latitude,so as to generate the semantic trajectory sequence with semantic information.Then,the trajectory sequence with semantics is transformed into spatial-temporal-semantics graph,that is,multiple trajectory sequences are divided into several spatial-temporal semantics maps according to time,and the graph convolution network is used to extract the deep features of spatial-temporal-semantics graph,that is,the spatial-temporal-semantics convolution block is used to convolute the spatial-temporal-semantics graph.Finally,TPM-AM is used for prediction analysis.Experiments show that the prediction speed of STPM-STSGCN is 0.17 times faster than that of TPM-AM,and the accuracy is improved by 1.19%.（3）Semantic Trajectory Prediction Prototype SystemOn the basis of theoretical research,this thesis designs and implements a prototype system of semantic trajectory prediction.The system has the functions of trajectory data input,trajectory data preprocessing,prediction model training,semantic trajectory prediction and prediction results display.The prototype system can easily perform training and prediction operations,provide a good visualization effect for input data and output results,can easily present the prediction results.