Research On The Prediction Of Crowd Movement Trajectory Based On Deep Learning

Exploring human mobility has long been concerned by many researchers in various disciplines such as nature,economy and society.On the one hand,the daily activities of human tend to be unpredictable.On the other,it has significant application value for behavior monitoring and population migration.The most important task in it is the trajectory prediction,which directly reflects the human lifestyle and mobility patterns.Now,the rapid development of deep learning and big data technology enables us to analyze deeply.In this thesis,we use neural network-based methods to analyze the mobility patterns of users' historical trajectories,and then predict future trajectories.We consider temporal characteristics of the historical trajectories,spatial characteristics of the historical trajectories and long-term dependencies under spatiotemporal characteristics.The main works and innovation are as follows:1.We propose a neural network based on encoder-decoder model to predict the future trajectories.This model uses a Seq2 seq architecture from machine translation to model users' whole trajectories,which considers the joint impact of historical trajectories and current trajectories.Specifically,the encoder encodes the historical trajectories and generates context vectors to save the mobility patterns,then the current trajectories and the generated context vector are used as inputs of the decoder.It finally decodes the current trajectories and generate the future trajectories.In addition,this model considers the effect of the information loss caused by long trajectories.We use the inverse of data and the Bi-LSTM two methods to alleviate this problem.2.Based on the encoder-decoder model,a graph convolution architecture is introduced.This model is able to consider not only the dual effects of historical trajectories and current trajectories,but also spatial relationships of the historical trajectories.Through the time-order historical trajectories,we establish a direct graph and an adjacency matrix of each location.Then approximate linear model of the graph convolution network is used to model the spatial relationships between the locations,so that this model can consider the spatial characteristics of trajectories.3.We propose a prediction model based on the Convolutional Neural Network(CNN),which solves the affect caused by long trajectories.We involve a 1D CNN to capture the short-term patterns of historical trajectories,then the Average-Pooling method is used to represent each sub-trajectory of whole trajectories,a fully-connected neural network is utilized to capture the mobility patterns among such sub-trajectories to represent the long-term patterns of historical trajectories subsequently.The future trajectories are predicted by the captured long short-term mobility patterns.Finally,numerous experiments show that our models have better performance in predicting future locations compared with the existing models.Although they are used for predicting human mobility,they can easily be extended to various prediction tasks.