Research On Autonomous Vehicles Trajectory Prediction Method Based On LSTM

Vehicle trajectory prediction is one of the key technologies to realize automatic driving,which provides an important guarantee for the safety,reliability and intelligence of vehicles in the process of automatic driving.However,due to problems such as complex road conditions,uncertain environment,diversity of environmental targets,and faster speeds in highway environments,there is a large deviation between the predicted trajectory of the vehicle and the actual trajectory.Therefore,how to overcome the influence of the above-mentioned problems on trajectory prediction and achieve high-precision and high-robust trajectory prediction is an urgent problem in automatic driving.At present,the method based on Long Short-Term Memory(LSTM)network can realize the trajectory prediction of the vehicle to a certain extent.However,this type of method does not consider the temporal and spatial interaction between the data,the information relevance in the data,etc.,so the effect is not good and the accuracy is insufficient.Based on the above-mentioned issues,this thesis has carried out research,and the main work includes the following:(1)Preprocessing of vehicle trajectory data based on Spatio-Temporal(ST)graphsThe current vehicle trajectory prediction focuses on the extraction and fusion of the interactive information between the target vehicle and surrounding vehicles(data interaction in time series and space,as well as semantic relations).In a complex dynamic environment,the accuracy of traditional methods for vehicle trajectory prediction is not ideal.poor effect.In response to these problems,this thesis proposes a preprocessing of the original trajectory data based on the Euclidean structure into a non-Euclidean trajectory data based on the ST graphs.First,fill the original trajectory data of the vehicle into the grid according to certain specifications;then process each grid into the form of a graph;finally,process each graph into the data format of a ST graph based on time series.It explicitly captures the ST interaction between all vehicles in the scene,and obtains more comprehensive data interaction characteristics,which helps to improve the accuracy of vehicle trajectory prediction.(2)Vehicle trajectory prediction model of LSTM based on ST graphsBased on the vehicle trajectory data based on the ST graph format,a vehicle trajectory prediction model based on LSTM is designed and implemented.The model is an Encoder-Decoder framework.The encoder is a hierarchical LSTM network component,with a total of three layers,which respectively encode vehicle trajectory data based on the ST graph format for simultaneous extraction the ST characteristics of vehicle trajectory data.The decoder cascades the characteristic data encoded by the three-layer encoder and then decodes it,and finally outputs the predicted trajectory data.This model effectively improves the accuracy of vehicle trajectory prediction.(3)Vehicle trajectory prediction model of LSTM based on the Attention Mechanism(AM)of ST graphsOn the basis of the LSTM vehicle trajectory prediction method based on the ST graphs,an AM is introduced,and an LSTM vehicle trajectory prediction model based on the AM of the ST graphs is designed and implemented.The model that introduces the AM can calculate the degree of correlation between the state vector between two vehicles at each time step,and simulate the degree of attention of the vehicle to the state of other vehicles,so that the trajectory prediction of the vehicle is more focused on processing the input data and the current output Significantly relevant key information,thereby improving the accuracy of vehicle trajectory prediction.This thesis is experimentally verified on the US-101 and I-80 data sets in NGSIM.The results show that compared with other algorithms,the algorithm in this thesis has obvious improvement and advantages in terms of the accuracy of vehicle trajectory prediction.