Research On Modeling And Prediction Of Vehicle Trajectory In Internet Of Vehicles

With the development of Internet technology,communication technology and artificial intelligence,the level of vehicle intelligence increases constantly.At present,several driver assistance systems have been used widely,and automatic driving systems are also studied.For driver assistance systems and automatic driving systems,vehicles need to make predictions of future trajectories of other vehicles in the driving environment,based on perceptual data information.On the other hand,in vehicular networks,vehicles will continuously report their real-time locations,based on which the networks can predict the vehicles’trajectories for the demanding vehicles.However,the current trajectory prediction accuracy cannot meet the needs of assisted or automatic driving vehicles.Therefore,this thesis takes advantage of the space-time trajectory information of the nearby vehicles of the targeted vehicle,and proposes two vehicle trajectory prediction models based on multi-sequence fusion and multi-task learning to support practical applications.The core idea of the vehicle trajectory prediction model based on multi-sequence fusion is to fully mine and efficiently fuse the vehicle’s own and environmental state information to improve the prediction performance of vehicle trajectory coordinates.Firstly,using the observable temporal features of the vehicle,the vehicle state containing context information of driving behavior can be represented through shallow neural network and pretrain task based on Word2Vec.Secondly,based on the vehicle state information,the vehicle’s own state sequence can be obtained;on the other hand,the vehicle’s environmental state can be represented by building a response area and using methods such as convolution to obtain the environmental state sequence.Furthermore,based on the state sequence of itself and the environment,a sequence-to-sequence model framework is adopted,and the two sequences are respectively encoded by the LSTM(Long Short-Term Memory)encoder,and the encoding results are fused by the attention mechanism to predict the future trajectory sequence through the decoder.Finally,the experiments on the real dataset indicate that the pretrain task,environmental state extraction methods and attention-based fusion methods designed in the model can effectively improve the prediction performance;compared with some existing models,the model has better prediction performance in different prediction time steps.Based on the previous studies,the trajectory generation part is optimized,and vehicle trajectory prediction model is proposed by leveraging multi-task learning.Firstly,the trajectory coordinate prediction task is divided into active region prediction and regional coordinate regression tasks;all network structures in the modeling part of the vehicle trajectory prediction model based on multi-sequence fusion are used as the underlying shared network,and LSTM decoders are respectively constructed for the prediction of active region and regional coordinates,and the prediction results are fused to get the final prediction result of trajectory coordinates to reduce the difficulty of trajectory prediction task.Secondly,the motion intentions of the vehicle in the latitudinal and longitudinal directions are defined,and a variety of parameter sharing methods are used to construct an auxiliary task based on the prediction of vehicle motion intentions to improve the accuracy of the active area prediction task.Finally,through experiments on real dataset,the optimal dimensional configuration of the active region is obtained,the results shows that the vehicle trajectory prediction model based on multi-task learning can reduce the mean value of prediction errors,and then effectively improve the final trajectory prediction performance;meanwhile,it is verified that the introduction of auxiliary task based on vehicle motion intentions prediction can bring gain to the final trajectory prediction performance.