Research On Bird ’s-eye View Vehicle Trajectory Prediction Method

Intelligent vehicle has great development potential in saving labor cost,improving traffic safety and efficiency,and is one of the strategic directions of the global automotive industry upgrading.The vehicle trajectory prediction method from the bird ’s-eye perspective studied in this paper has become an important part in the study of intelligent driving schemes as an important basis of vehicle decision-making module.Existing methods only rely on the error fitting between real trajectory points and the generated one,but do not consider the confidence information of generated trajectory.At the same time,because of not input the map information into the model,the prediction method can only predict vehicle trajectorys in the traffic scenes which similar to the scenes in dataset.Thus,it is difficult to ensure the robustness.This paper is mainly supported by the sub-project of the National Key R&D Program "Research on Vehicle Peripheral Multi-target Behavior Recognition and Prediction Algorithms".The vehicle trajectory prediction from the bird’s-eye view is studied.The main research work is as follows.1.Model the vehicle trajectory prediction task from bird ’s-eye view,including trajectory modeling and maneuvering modeling.At the same time,the advantages of NGSIM,High D and Argoverse data sets are deeply analyzed,and the role of different data sets in the process of model training is studied.Finally,the data processing of the dataset is introduced and the generating logic of the sample set is defined.2.A trajectory prediction method for one vehicle from bird’s eye view,CAE-GAN,was proposed.The encoder which constructed by the convolution social pool,maneuver prediction module and LSTM and LSTM decoder were used as the model infrastructure.Speed,acceleration and lane information was added as the model input.The multi-loss joint training scheme of MSE loss,Generative Adversarial Networks(GAN)loss and classifier loss improves the Generative trajectory accuracy,confidence and prediction maneuver matching rate respectively.At the same time,the loss function of GAN was improved.This method solves the problems of low confidence and accuracy of the generated trajectory and difficult convergence of GAN training.In the validation set,the RMSE accuracy of CAE-GAN in 1-5 seconds is 0.4317 m,1.0690 m,1.8287 m,2.7744 m,3.9556 m respectively.Compared with other algorithms,the average accuracy is improved by 14.6%.3.A trajectory prediction method for muti-vehicles from bird’s eye view,Ti Sp-Map AT,was proposed.The method uses the model encoder based on space and time interaction model and the map interactive attention.Build the time-space graph and lane graph through the establishment the vehicle node and lane node.The temporal and spatial fusion module and lane information extraction module are used to complete the feature interaction of time-space and transfer of node information.Meanwhile,a multi-modal trajectory generation method is established to predict multiple trajectories and the confidence degree of each trajectory.This method solves the problem that the existing methods are not suitable for each traffic scene,and solves the problem that the prediction efficiency of one vehicle prediction method is low.In the validation set,ADE,FDE and MR reach 1.68 m,3.70 m,0.57 when K=1,and 0.83 m,1.32 m,0.15 when K=5.Compared with other algorithms,the average accuracy is improved by 10.8%.4.The two algorithms are deployed in hardware.The corresponding nodes and logic of trajectory prediction system are designed based on ROS.The algorithm is tested in real traffic scenarios.Experimental results show that the two methods proposed in this paper can make excellent prediction of future trajectory under the real vehicle experiment,which can meet the needs of intelligent driving vehicles.