The Research Of Topological Space Based Behaviour Decision And Trajectory Planning Methods For Intelligent Vehicles

The development and promotion of Intelligent Vehicles have enormous value to safe traffic,transportation efficiency and intelligent transportation systems.Presently,environment perception,behavior decision and motion control still have some drawbacks in the intelligent driving system.The behavior decision and trajectory planning module has been the critical sub-module of intelligent vehicles,the core symbol of its intelligent level.Due to the uncertainty of traffic participants’ behavior,the massive varying scenarios in open traffic environments,present behavior decision and trajectory planning methods have problems in consistent decision and planning,real-time planning in complex environments.Present behavioral decision and trajectory planning methods can hardly meet people’s requirements for intelligence and safety of autonomous vehicles.Therefore,this study focuses on the intelligent vehicles and the critical problems of the behavioral decision and trajectory planning in complex traffic.The idea of decision planning based on topology is proposed.Topological path planning is put forward firstly.Then,topology is extended to the synchronous behavior searching and trajectory planning in complex dynamic traffics.The topological features are also used in the behavior decision self-learning algorithm for autonomous vehicles.This dissertation aims at making breakthroughs in behavioral decision and trajectory planning in environments with irregular static obstacles or complex dynamic traffics.Finally,the proposed algorithms are tested on the autonomous vehicle platform that combines the virtual simulation and on-vehicle test.The main work and creative ideas include the followings:(1)For the difficulty in extracting and representing the drivable area and boundary constraints,a method solving the extracting and optimizing of topological routes in two-dimensional space with disordered obstacles is proposed.The topological routes are searched from the segments,which are split from the reference paths by static obstacles.The high-order penalty functions with double boundary constraints are proposed instead of hard constraints for optimization.A topological path is optimized by square quadratic programming numerically.Experiments on a road placed with movable stakes indicate that this planner can find several topological paths in a discrete manner.Comparisons with main planners show that this method cannot only plan smoother and reasonable paths but also compute with low costs.(2)For the complexity of planning in heavy traffic,the synchronous behavior searching and trajectory planning method(SMSTP)is proposed based on the topological space.This method constrains the planning space in a narrow space-profile of trajectories.The spatial-temporal space is decomposed into sub-spaces using the predicted trajectories of other traffic participants,and the reunion of sub-spaces generates different topological routes.The final trajectory of one single route is solved by decomposed optimization in longitudinal and lateral direction successively.Simulation experiments in heavy traffic and highway dropping demonstrate that SMSTP is skilled in planning in dynamic traffic.Comparison with traditional behavior-aware planners indicates SMSTP has large advantages over those methods in several performance indexes,such as topological route’s generation,high-resolution trajectory,computing efficiency,etc.(3)To deal with the behavior decision problem in heavy traffic,the topological features based behavior decision(TFBD)using Deep Reinforcement learning is proposed.Combining the features of the trajectory features in SMSTP algorithm,self-states,and lane information,the Markov decision process(MDP)model to the behavior decision of intelligent vehicles is established.A Deep Reinforcement learning algorithm is designed to learning optimal decision policy based on the off-line twinned deep deterministic policy gradient algorithm.Simulation experiments are tested in several scenarios,including the six-lane road,the curvy road with zebras,the merging road,where the trained intelligent vehicle makes reasonable behavior decisions with proper trajectories.Comparisons with the expert system and the policy without full features of trajectory profiles indicate that the standard TFBD algorithm has larger possibilities to learn an optimal policy.(4)A quantitative evaluation method based on Extension theory and Analytic Hierarchy Process is proposed for the comparison of motion planning algorithms.Six performance indices,including smooth,consistency,accuracy,computing efficiency,safety and reachability,are considered.By adding virtual dynamic and static traffic elements(including traffic participants,road structures,etc.),a new simulation platform is formulated for the behavior decision and motion planning system of Intelligent vehicles.The platform facilitates the quick testing and iteration of algorithms by combining real scenarios and virtual environments.Supported by algorithms in this study and the simulation platform,the autonomous vehicle’s intervention ratio is less than 1.5% in complex 2000 km autonomous driving test.The second,third and fifth parts of this study have been used on the autonomous vehicle ”HQ3”.The fourth part has been tested on ”HQ3”.