Research And Application Of Long Trajectory Embedding And Prediction Model Based On Context And Time

Human trajectory prediction plays an important role in many aspects,such as urban planning,urban management,business decisionmaking and so on.With the development of the Internet,it is possible to obtain the human trajectory data throughout the city.Nowadays,the common mobile trajectory data include three types.One type is the data collected by volunteers with recording devices.These devices are used to record volunteers’ mobile behaviour.The second type is the user’s active check-in data collected by websites that provide check in function.The third type is the mobile phone signalling data collected by users when they use mobile phones to connect to the operator’s network.Most of the existing trajectory prediction models usually use user’s check in data or data collected by volunteers.These two types of data are relatively sparse and difficult to represent the actual mobile behaviour of users.A small number of trajectory prediction models uses mobile phone signalling data collected from the operator network,but because the mobile phone signalling data collected by the operator network are usually very dense and contain a lot of noise data,when people use this kind of data,they usually set the staying time threshold artificially to clean the data.Such behaviour will cause a large number of data containing effective trajectory information to be deleted,resulting in the reduction of prediction performance.The existing trajectory embedding model usually works better when using longer dense trajectory.However,due to the limitation of video memory resources,it can only artificially shorten the dense trajectory and cannot embed longer dense trajectory.In view of the above problems,this thesis focuses on how to improve the performance of trajectory embedding and prediction methods for long trajectories and the practical application of long trajectory embedding and prediction models.This thesis mainly includes the following contents:1)Design and implement a context-based timeaware long trajectory embedding model.This model not only takes up relatively low video memory resources for training in long trajectory scenarios,but also has better performance in both long and short trajectory scenarios.2)Design and implement a long trajectory prediction model based on staying time.This model not only solves the problem that the prediction performance is reduced in the process of manually setting the residence time to clean the long trajectory data,but also has a good prediction accuracy.3)The population flow monitoring system was designed and implemented,and the long trajectory embedding and prediction model was applied to practice,providing a convenient query and monitoring channel for urban managers and epidemic prevention and control workers at that time.