Research On Trajectory Prediction Based On Deep Learning

Taxi trajectory prediction can improve urban traffic safety,which is part of intelligent transportation.For traffic management,it can effectively improve the control and utilization of taxis as a public transportation resource.At the same time,it will help alleviate the urban congestion,improve the utilization of urban roads,and bring tangible benefits to the general public.In this paper,the taxis’ historical trajectory data is used to predict the driving trajectory of a single taxi from a microcosmic level to obtain the road to be driven.Firstly,due to the characteristics of the infinite domain of urban plane anchor points,map matching is used to match the anchor points of taxi tracks to urban roads,and each track is represented by a set of roads.Then,the time series of trajectories is processed with an encoder-decoder model based on the attention mechanism.The model consists of three parts:the encoding part,consisting of a Long Short-Term Memory(LSTM),to process the known part of each trajectory;the decoding part,consisting of another LSTM,for decode the result generated by the encoding part,and take the decoded new road as the prediction result.Attention mechanism part,this part mainly solves the decisive problem of the key road.When decoding,we can comprehensively consider the known part of the trajectory and then more accurately infer the appropriate future road based on the special sections of the trajectoryFurther,the taxi trajectory has a strong purpose.So Multi-Task Learning(MTL)is used to explore the role of the destination in the trajectory prediction.In the Multi-Task Learning model,one LSTM is constructed based on the known part of the trajectory to prediction the destination firstly,and then another LSTM is used to predict the future path of the trajectory.Next,the two LSTMs are made to achieve parameter sharing with dynamic combination,so that the LSTMs are could be effected by another.As a result,the two LSTMs can predict the destination and the future road respectively.In this paper,the taxi data set in Xi’an is used to analyze the impact of the known part of the trajectory on the future path and the destination’s prediction.Compared with the existing prediction methods,the models in this paper have a good performance in the prediction of the trajectory.In addition,through comparison,it is found that the use of the attention mechanism can significantly improve the prediction effect of the encoder-decoder model.At the same time,comparing with the trajectory prediction of single data,the prediction containing the destination can significantly improve the result of prediction.