Research On Decision And Planning Algorithm Of Intelligent Vehicle For Non-intersection Sections

With the continuous development of new technologies such as electric vehicles,artificial intelligence,sensor technology and 5G communication,intelligent networked vehicles will become an important pillar in major national development strategies such as "Made in China2025" and "Internet+".As one of the core systems of autonomous vehicles,the decision planning system is responsible for guiding the vehicle through various complex roads safely and efficiently,and therefore determines the intelligence of autonomous driving.In this paper,based on the National Key Research and Development Program of the Ministry of Science and Technology,"Research on Key Technologies of Electric Autonomous Vehicles and Demonstration Operation",we study the decision planning algorithm of autonomous vehicles in non-crossing roads,and model the decision system through machine learning.can quickly select the appropriate target lane and plan a safe,legal and comfortable trajectory.In this paper,we first process the public autonomous driving dataset,Next Generation SIMulation(NGSIM),and use the processed data as the original data for the decision network,then clean the original data,mainly including data pre-processing,decision scenario extraction and data generation three processes,so that it meets the requirements of the network input format.Secondly,in the Prescan simulation process,the data from the simulated sensors also need to be processed by the algorithm in order to be input to the decision network for decision making,which includes the use of LIDAR for target detection,and the use of millimeter wave radar for the extraction of motion information of the ambient vehicle.Once the input data for the decision network is ready,the data is then used to make decisions and plan a reasonable path.The decision system mainly adopts three approaches: the decision system based on fully connected feedforward neural network,the decision system based on recurrent neural network and the decision system based on automatic machine learning,and the appropriate network parameters and training parameters are selected for training according to the characteristics of each model,and the optimal network is selected as the decision model by comparing the performance of the three different networks on the data set.The motion planning algorithm first plans a path using the RRT* algorithm based on the target lane in the decision result,and if the planning fails within the specified time,it keeps the current lane driving,and if the planning succeeds,it performs trajectory smoothing and speed planning based on the planned path.Both trajectory smoothing and speed planning will be optimized by setting equation and inequality constraints respectively and selecting the corresponding optimization objective function for optimization.To ensure the continuity of trajectory and speed and simplify the calculation,both trajectory and speed equations are expressed by using cubic polynomials.Finally,a joint simulation environment based on ROS and Prescan is built in this paper,and the optimal decision network is selected and the effectiveness of the motion planning algorithm is verified before the dynamic verification,and then the simulation results of three dynamic working conditions reflect the effectiveness of the decision planning algorithm in this paper.