Studies On Trajectory Planning Considering Motion Uncertainties Of Traffic Vehicles

Vehicle intelligence has been well regarded as one of the key technologies on various problems faced by automotive industry,such as safety,traffic jam and environmental pollution.Trajectory planning is considered as one of the core functions for vehicle intelligent driving,which is based on accurate threat assessment on the motion of traffic vehicles.Since future motion of traffic vehicle is in general unknown,its trajectory is literally a random variable under a random process.There is no absolute safe state in the trajectory planning process,and only collision probability can represent threat assessment result accurately.Therefore,in order to realize safe driving,trajectory planning should calculate collision probability of the planned states accurately and then take it into consideration.So,trajectory planning module must predict future trajectories of the traffic vehicles.Meanwhile,to develop the safety performance of the intelligent vehicles,the probabilistic properties of the threat assessment should be modeled and further be carefully considered in the trajectory planning module.Maneuver-based motion model is effectively used in long-term trajectory prediction of the traffic vehicles.Traditionally,one kind of driving maneuver is corresponding to one motion model.However,in reality,there exists several patterns in one maneuver because of different driving styles among drivers,and there are large differences between trajectories occurred by different driving maneuver pattern.Thus,to improve the precision of the trajectory prediction of the traffic vehicles,the recognition of the driving maneuver pattern should be added based on the driving maneuver recognition.The Support Vector Machine-based classifier has been used as an effective recognition method in previous researches.However,the inherent traffic vehicle parameters and the driver and vehicle states are difficult to be acquired by on board sensors of the intelligent vehicle.As a result,the recognition of maneuver patterns based on a single classifier and a single sample are no longer precise.Under the maneuver-based motion model framework,Gaussian process-based motion model is an effect method to depict the randomness of the vehicle motion and the basis for vehicle future trajectory predication.Traditional methods usually utilize motion model itself to describe probability properties of future trajectory.However,this kind of methods overlooks priori vector which corresponding to the real-time trajectory.Up to now,there is few research keeps an eye on this problem.Rapidly exploring random tree is a popular approach to solve intelligent vehicle trajectory planning problem because of it probabilistic completeness and rapidly exploring ability.Traditional researches mainly focus on problems such as vehicle dynamic or kinematics constraints.They overlook the problem of collision probability.Thus,the risk assessment of planned state is not accurate and the safety of the planned trajectory cannot be guaranteed.Some researches proposed a collision probability threshold to assess the safety level of the main vehicle.But they do not further research the influence of the collision probability properties on the planning process,which results in some blindness to deal with the uncertainties in trajectory planning.In order to solve the problem proposed above,in this thesis,the problem of trajectory planning considering motion uncertainties of traffic vehicles has been studied.The main contents are shown below:Firstly,a maneuver pattern recognition method was proposed to improve the recognition precision.The structure of this method is based on “one-vs-all” error correcting output codes framework.Under this framework,the maneuver pattern recognition problem is decomposed into several binary classification problems.Then,a probability estimates model based on generalized Bradly-Terry model is built.The minimum relative entropy is used as the optimal criterion to calculate real probability.Thus,the single classifier has been replaced by multi-classifier.The relationship between continuous probability estimation results and final recognition result is described by Bayesian inference model.The experimental results show that the proposed recognition method can improve the accuracy rate effectively.Secondly,a trajectory prediction method based on Gaussian process based motion model is proposed in this paper.The pattern clustering of the motion trajectory is established based on the Gaussian Mixture Model.After this,Gaussian process based motion model is built.Then,the matching relationship between real-time motion trajectory and motion model is effectively established by priori vector computing method based on mahalanobis distance.The probability properties of the future traffic vehicle motion is predicted based on conditional Gaussian distribution.The experimental results show that the prediction method can precisely derive the dimensionality of the priori vector and the probability properties of the future traffic vehicle motion trajectories.Thirdly,a novel trajectory planning method which can tackle the uncertainties explicitly was proposed.The proposed method,Probabilistic CL-RRT(PCL-RRT),is a probabilistic extension of the famous CL-RRT method.This method mainly consists of three parts,including sampling strategy,node evaluation and goal state insertion.Under the framework of CL-RRT,the mentioned sampling strategy can generate random node according to the road geometry.Thus,it can improve search efficiency by reducing ineffective nodes.Node evaluation,which models uncertainties as costs of the corresponding node,is the core part of the method.Based on node evaluation,the sampling node always tries to connect the tree node with minimum cost,i.e.the safest node among all the tree nodes.At last,the goal state insertion guarantees goal navigation ability of the method.In order to further verify the trajectory planning method proposed in this paper,the real vehicle experiments are also conduced.Firstly,to establish the real vehicle platform,the structure of the software and hardware,communication system and power supply system are all designed and conducted on a real car.Besides,the path tracking controller based on single-point preview and velocity tracking controller based on acceleration preview are also loaded in the real vehicle.The parameters of the real vehicle are also estimated before experiments.Finally,the experimental results verifies the effectiveness of the trajectory method proposed in this paper.