Research On Communication-aided Behavior Planning And Motion Planning Of Autonomous Vehicles

Behavior planning and motion planning are crucial parts of autonomous vehicles,their real-time performance and robustness directly affect the power and ride comfort of autonomous vehicles.However,current decision-making and planning algorithms generally only consider single-vehicle intelligence.V2 X communication is not considered,the writing of detailed rules leads to long code and low efficiency,the coverage of scenarios is limited,and the algorithm cannot be updated in real time.In response to the above problems,this thesis relies on the national key research and development plan to carry out research on behavioral decision-making and motion planning of autonomous vehicles based on communication assistance,focusing on the fusion of GCN and DQN method for lane changing and the speed planning method under the improved space discretization,and were verified on the simulation and vehicle platforms respectively.First,a lane-changing decision method based on the fusion of GCN and DQN is constructed.Aiming at the problem of insufficient feature extraction capability of traditional obstacle representation methods,a graph representation algorithm for dynamic obstacles and passable areas is established by fully considering the characteristics of dynamic obstacles and passable areas.On this basis,it combines the advantages of deep learning’s strong ability to extract features and reinforcement learning that does not require a large number of data sets for training.The dynamic obstacle map is extracted and passed through a fully connected layer,graph convolutional neural network and deep Q-learning Network,and finally get the Q value of different lane-changing actions.The model proposed in this chapter has been trained and tested in the simulation environment of straight expressway and intersection,respectively.The experimental results show that the proposed method of lane-changing decision-making has the characteristics of short training speed,flexible lane-changing in congested sections,etc.Requirements for real-time lane change decisions for autonomous vehicles.Then,in view of the shortcomings of the traditional global planning algorithm A*algorithm,such as low search efficiency and too simple heuristic function,an optimized A* algorithm based on jump point search is proposed.By formulating pruning rules,a large number of symmetrical path nodes are skipped,and only the nodes are expanded.When searching for jumping points,the algorithm speed is greatly accelerated.In local motion planning,considering the high dimension of motion planning optimization and the non-convexity of the problem,motion planning is decoupled into path planning and velocity planning.The differential flatness of the kinematics model of autonomous vehicles is proved,and the path planning problem is expressed as path planning based on dynamic programming and path smoothing based on quadratic programming satisfying the vehicle kinematics.On this basis,the expression of speed planning under time discretization and space discretization is analyzed.Aiming at the problem that they cannot deal with nonlinear constraints,a local speed planning algorithm based on improved space discretization is proposed,which can effectively deal with autonomous vehicles.The speed,acceleration and dynamic obstacle constraints that need to be met when driving on the road,while taking into account the safety thresholds of the autonomous vehicle and dynamic obstacles during driving.By linearizing the nonlinear constraints,the speed programming problem is transformed into a convex quadratic programming problem,which can be quickly solved by the OSQP method to meet the real-time requirements of autonomous vehicles.Finally,the real-time performance and robustness of the motion planning algorithm under the assistance of communication are verified by L4 autonomous vehicles,and the real-vehicle experiments are carried out in the scene of obstacle avoidance in straight lane,right turn at intersection and congested ring road scene respectively.The lattice algorithm is compared with the planning algorithm based on MPC.The experimental results show that the designed decision-making method based on the fusion of GCN and DQN has more flexible lane changing ability in congested roads,and the motion planning algorithm of path speed decoupling can meet the requirements of autonomous vehicles.Compared with other motion planning algorithms,it has higher traffic efficiency and ride comfort.