| In order to ensure the safety of traffic participants,the intelligent driving system should have a strong ability to perceive the environment and be able to predict the trajectory of surrounding traffic vehicles more accurately.The trajectory prediction of the surrounding traffic vehicles by the autonomous vehicle can define a usable space for the planning module,and the planned path in this space has high safety.Applying the trajectory prediction function to the driving assistance system can remind the driver of potential hazards in the road,and even intervene in vehicle manipulation,so as to actively avoid the collision danger.Aiming at the problem of traffic vehicle trajectory prediction,this paper proposes a long-term and short-term trajectory prediction fusion prediction method based on vehicle behavior recognition for expressway scene.The main research content includes the following aspects:Firstly,this paper proposes a lane-changing feature dimensionality reduction method based on the attention mechanism and a lane-changing behavior recognition model based on LSTM.First of all,find the required data set from the existing open source data,the data set needs to contain detailed vehicle trajectory information and vehicle-related status information.At the same time,the data needs to have high accuracy.Choose the high D data set according to the above requirements.The current behavior of the vehicle is hidden in its historical data set,so a certain length of vehicle trajectory information is selected for each training sample,and the vehicle behavior at the last moment of the sample is used as the label.Using the Attention mechanism has the ability to get the most effective information from a high-dimensional array and reduce the dimension of the data.The lane-changing behavior recognition model is established based on LSTM to give full play to the advantages of LSTM in time series modeling.The lane-changing behavior recognition model is verified by the data sets before and after dimensionality reduction.The results show that the recognition accuracy of the model is high,and the impact of dimensionality reduction on it is small.Secondly,in view of the uncertainty of trajectory prediction,this paper proposes an uncertainty trajectory prediction method based on Gaussian process motion model.Select the required lane-changing trajectory from the high D data set,use Gaussian mixture clustering to cluster it,use the indicators AIC and BIC to determine the number of clusters,and obtain the prototype trajectory from the completed trajectory of the clustering.The Gaussian process motion model is used to model the prototype trajectories of each category,and the trajectory prediction method is derived according to the conditional probability properties.The known vehicle trajectories are calibrated in time and space,which is consistent with the coordinate system adopted by the Gaussian process model.The distance index is used to obtain the matching relationship between the historical trajectory and the prototype trajectory.An example is used to verify that the trajectory prediction method has advantages in long-term uncertainty trajectory prediction.Thirdly,the trajectory prediction method based on Gaussian process has defects in the short-term trajectory prediction,and the prediction method based on kinematics model has higher accuracy in the short-term prediction,but poor prediction effect in the long time domain.Therefore,this paper proposes an interactive multi-model trajectory prediction fusion method.Kinematics prediction adopts CTRV model and uses extended Kalman filter to predict trajectory in a recursive manner.Interactive multi-model dynamically assigns the weights of the two models according to the prediction error.Because of the abnormal curvature of the predicted trajectory in the transition stage of the main model,a polynomial fitting smoothing process is adopted.Use relevant indicators to compare the forecast errors of different forecasting methods in the long and short time domains. |
PDF Full Text Request