Research On Behavior Recogniton And Prediction Method Of Vehicles Surrounding Unmanned Vehicle

The method of recognizing and predicting the behavior of vehicles surrounding unmanned vehicle is an important module to realize unmanned.This module accepts the driving information of surrounding vehicle obtained by the perception module,uses algorithms to recognize the current behavior of surrounding vehicles and predict their future behavior and trajectory.The recognition and prediction results are input into the decision and planning module as reference information for the unmanned vehicle to make decisions and plan future paths.Therefore,accurate recognition and prediction of the behavior of surrounding vehicles can ensure that unmanned vehicle can drive safely and efficiently in real traffic scenarios.This paper focuses on the behavior recognition at the current moment and the future behavior and trajectory predicting of vehicles surrounding unmanned vehicle on a structured road.(1)Two natural driving vehicle trajectory datasets in the field of vehicle behavior recognition and prediction are compared.The advantages and disadvantages of the two datasets are analyzed to provide data support for the research of this paper.The classification of vehicle behavior is described.The lateral behavior of vehicle is the key issue of this paper.The method of making the behavior labels corresponding to the observation sequence is determined.The concept of trajectory segment coordinate system is introduced to bring the coordinate system of the original dataset and the global coordinate system of the unmanned vehicle to the same coordinate system,and points out the influence of neighbor vehicles on the behavior of target vehicle,and extracts the corresponding behavior samples in the original data set.Then,the vehicle behavior recognition data set,vehicle behavior and trajectory prediction dataset which needed in this paper are established.(2)Existing research methods ignore the influence of neighbor vehicles on target vehicles.Therefore,this paper proposes a composite model combining HMM(Hidden Markov Model)and MLP(Multi-Layer Perceptron),this model uses HMM's time series modeling capability to model the vehicle's lateral displacement,lateral speed,vertical speed and other continuous observation variables.HMM is used to make preliminary recognition of vehicle behavior.The output of each HMM behavior recognition model is combined with the neighbor information of the target vehicles,and the combined information is inputted into the MLP to complete the final behavior recognition of the surrounding target vehicles.The effectiveness of the algorithm is verified by the performance of the model on the test set.(3)A method for predicting the behavior and trajectory of surrounding vehicles based on the attention mechanism is proposed.The attention mechanism is added to the long short term memory network encoder-decoder architecture.The influence of historical observation information on the vehicle trajectory at each predicting time step is different,and the attention mechanism can adjust the weight of historical observation information in the model at each predicting time step according to the influence.The proposed multi-task training method for vehicle behavior prediction and trajectory prediction enables the model to complete the future behavior prediction and future trajectory prediction of surrounding vehicles at the same time.The algorithm pays more attention to the micro-behavior prediction task which is the future trajectory prediction of surrounding vehicles.The effectiveness of the algorithm is verified by the performance of the model on the test set.(4)The effectiveness and reliability of the method for recognizing and predicting the behavior of vehicles surrounding the unmanned vehicle are verified through the performance of the algorithms in real scenarios based on the "UJS intelligent driving vehicle" experimental platform.