Research On Vehicle Trajectory Prediction Based On Long-Short Term Memory Encoder-Decoder Model

Predicting traffic participants’ trajectories is the premise of decision-making for automatic driving systems.It is also one of the critical issues in deploying autonomous driving systems.However,due to the interaction between traffic participants,the multi-modality of intentions of traffic participants,and the influence of road geometry and traffic rules,the correlations between the historical trajectory of the target vehicle and its future trajectory contain some uncertainty.The uncertainty brings significant challenges to building the trajectory prediction model.Traditional physical mechanisms use physical laws to predict the future trajectory of vehicles.This method ignores the behavioral interaction between traffic participants,resulting in a significant deviation between the predicted and actual trajectories.Traditional deep learning trajectory prediction methods find the relationship between the vehicle’s historical trajectory and future trajectory through data training,so as to predict the vehicle’s future trajectory based on the current state.Although this method considers the interaction between the road users,it ignores the difference in the influence of each road user on the prediction.The Long-Short Term Memory(LSTM)network also has the problem of accumulation error.Therefore,overcoming the effects of the above issues on trajectory prediction and achieving accurate and robust vehicle trajectory prediction is an urgent and critical problem for autonomous driving.The following are the main contributions of this thesis in view of the above problems.(1)To model the interaction between vehicles,a vehicle trajectory prediction model based on the Convolutional Social Long-Short Term Memory(CS-LSTM)network is adopted to simulate the real road scene better.Aiming at the cumulative error problem of the LSTM network trajectory prediction model,the Exponentially Weighted Convolutional Social Long-Short Term Memory Network(EW-LSTM)model is proposed based on the CS-LSTM model.The EWLSTM model proposes an exponential,smoothed L1 loss function to correct the accumulated error.The simulation results show that compared with the traditional baseline model under the squared loss function,the exponentially weighted smooth L1 loss function effectively reduces the accumulated error and improves the accuracy of vehicle trajectory prediction.(2)To model the importance of the surrounding vehicles on trajectory prediction and make the correction of the accumulative error adaptive,this thesis proposes the Dual-Attention Convolutional Social Long-Short Term Memory Network(DA-LSTM)model.On the basis of EW-LSTM model,the DA-LSTM model utilizes the input spatial attention mechanism to simulate the surrounding vehicles’ importance to the target vehicle trajectory prediction and assign the corresponding weight to the related trajectory coding features,so that the model focuses on the feature extraction of important vehicles.The output attention mechanism weights the predicted trajectory from LSTM network,which avoids the problem of prediction performance degradation caused by the increase of output length,reduces the cumulative error caused by LSTM network,and further improves the accuracy of vehicle trajectory prediction.(3)To model the influence of surrounding static environment and multi-modal vehicle intention on vehicle trajectory prediction,a Multi-modal Vehicle Trajectory Prediction Based on Static Constraints(MTPSC)model is proposed.Based on the CS-LSTM model,the MTPSC model uses the static environment constraint feature extraction module to convert the static environment around the vehicle into constraint features,and uses the Expert Mixture Model as the decoder to predict the future trajectory of the vehicle,reducing the error of the predicted trajectory relative to the actual trajectory with environmental constraints.