Study On Motion Prediction Of Traffic Vehicles For Intelligent Driving

With the continuous improvement of vehicle intelligence,some Advanced Driver Assistant Systems and partially autonomous driving application have already appeared in production cars,and fully autonomous driving function has also entered a large-scale testing stage.The maturity of these vehicle intelligent systems is of great significance to solve the problems brought by traditional automobile industry such as safety,congestion and environmental pollution.Vehicle intelligence has been recognized as one of the inevitable development directions of automobile technology.It is necessary to pay close attention to the change of surrounding circumstance and potential dangers when driving in complex traffic.An intelligent vehicle needs to reasonably predict the change of surrounding circumstance when it makes decisions and plans a future trajectory that is safe,feasible and satisfies the requirements of comfort,economy and driving efficiency.Human drivers can gain the ability to infer the intention of surrounding traffic vehicles and predict their future motion by constantly learning in the traffic environment.However,much information cannot be observed such as the intention of surrounding traffic vehicles for the intelligent driving system,and the probability reasoning needs to be performed with observable variables.Therefore a motion prediction method capable of inferring the driving intention of traffic vehicles is crucial for the intelligent driving system.At present,most of the motion prediction methods of traffic vehicles are based on the physics-based motion models and maneuver-based motion models.These two types of methods only consider the one-way effect between traffic and ego vehicle.In other words,the result of traffic vehicle's motion prediction is used as the input of ego vehicle's decision and planning.In fact,however,the motion of ego vehicle has a great influence on the motion of traffic vehicle.These methods consider each traffic vehicle as an individual when performing motion prediction,and the interaction of traffic participants is neglected.Apparently,this approach will lead to a serious bias of motion prediction in some scenes with the obvious interaction of traffic participants such as highway entrances and congestion roads.On the other hand,most current motion prediction methods are deterministic,which are usually based on the vehicle dynamic and kinematic models.However,the sensed or perceived motion of traffic vehicles can be full of errors or inaccuracy due to the inaccurate and/or incomplete sensing information and the input of traffic vehicle driver is also unknowable.So the future trajectory of traffic vehicles is full of uncertainty.It is obviously difficult to obtain reliable results for such deterministic motion prediction methods.In addition,even some current methods considering uncertainty,it models the uncertainty only by adding Gaussian noise,and it is difficult to accurately express the probability distribution of vehicle motion in a real traffic scene.In view of the deficiencies in the current research,this paper proposes a method of motion prediction that considers the interaction of traffic participants.It mainly includes the following aspects:Firstly,a game theoretical reasoning approach for traffic vehicles' maneuver is proposed.This research model the interaction of traffic participants as a non-cooperative game with incomplete information.Based on the game theory,an algorithm framework of interactionaware maneuver prediction is designed.The payoff function is used to model the intention of driver.The expected utility is employed to express the intention probability of each maneuver,and maneuver prediction is performed by combining the results of maneuver recognition.Through comparison experiments with traditional maneuver prediction methods,the results have demonstrated that the proposed approach can predict the maneuver of traffic vehicles earlier,faithfully represent the interaction of traffic participants,and have higher robustness.Secondly,the maneuver recognition of traffic vehicles is performed based on the continuous Hidden Markov Model.As an important part of the interaction-aware maneuver prediction framework,the maneuver recognition of vehicles through historical trajectories represents an understanding of its current maneuver.This paper uses continuous Hidden Markov Model to model vehicle maneuver as discrete unobserved states.Continuous vehicle states are used to estimate the maneuver probability.The model is trained and tested using NGSIM data.The test results show that the model has high sensitivity and accuracy,and the accuracy of recognition of lane-changing maneuver before 1s can reach more than 95%.Thirdly,a trajectory prediction method based on Gaussian process is proposed.According to the real sample trajectories,Gaussian process is used to model the motion of traffic vehicles with different motion modes,and the uncertainty of traffic vehicles' trajectory is represented more realistically by the probability.At first,according to different driving habits the real sample trajectories are classified by the Gaussian mixture model.Then separate GP-based motion model are built for each category.Finally,the probability of conditional Gaussian distribution is deduced,and the GP-based model is used to predict the motion of traffic vehicles.The final experiment demonstrated that the proposed method is suitable to reflect the uncertainty of traffic vehicles' trajectory realistically and elegantly.The error between the expectation of predicted trajectory and real trajectory is very small,and it has a good performance in long-term prediction.In order to estimate the parameters of payoff function in the interaction-aware maneuver prediction framework and train the HMM-based maneuver recognition model,the trajectory data of NGSIM is preprocessed,and the effective data is extracted to estimate the parameters.Finally,in order to verify the whole motion prediction methods proposed in this paper,three experiments were designed in two scenes.Compared with traditional motion prediction method,the experimental results show that the proposed method has better performance in terms of the advanced degree of prediction,reflecting the interaction and robustness due to considering the interaction of traffic participants and modeling the driver's intention.