Research On Autonomous Driving Decision-making At Park Intersections

With the continuous development of automatic driving technology,many automatic driving assistance functions under highway conditions have become popular,such as lane departure warning,adaptive cruise,etc.As one of the common traffic conditions in urban conditions,it is more difficult for self-driving cars in urban conditions than a single scene element on the highway.In urban working conditions,traffic accidents that occur at intersections account for a large proportion of the total traffic accident scenes,and on the basis of T-shaped intersections without signal lights in the park,steering vehicles are added on this basis.In order to avoid scratches and other reasons,there may be cases where turning by the opposite lane or returning to the main lane after turning is too slow.There are certain potential safety hazards for vehicles going straight through the intersection,which affects the efficiency and safety of road traffic.It is an urgent problem to study the automatic driving decision-making system in this scenario.Therefore,this paper conducts autonomous driving decision-making research on different scenarios of going straight through the T-shaped intersection without signal lights in the park.The main research contents are as follows:1.For the problem of modeling the trajectory of the steering car.First of all,through analysis,6 kinds of routes that will affect the straight-going vehicles when the turning vehicles turn are summarized.Then this paper adopts the method of driving simulator to let the driver imitate the steering operation of the turning vehicle in the scene of T-shaped intersection without signal lights in the park.Finally,through the method of least nonlinear square fitting,it is determined to use the quintic polynomial trajectory to model the trajectory of the steering vehicle.2.In view of the small number of test scenarios,this paper proposes a method based on variational Bayesian Gaussian mixture clustering and ADASYN adaptive oversampling to perform scene generalization on the intersection scene without signal lights in the park.The method of variational Bayesian inference is used to solve the Gaussian mixture model,which solves the problem that it is difficult to determine the number of components of the Gaussian mixture model when using the traditional EM algorithm.Then,under the comprehensive comparison of the number of scenes generated,the effect of trajectory generation and the complexity of the scene,the ADASYN oversampling algorithm was determined to complete the oversampling of sample data,and complete the trajectory generation of steering vehicles under different routes based on the quintic polynomial.3.Aiming at the problem of low learning efficiency in reinforcement learning.This paper proposes a deep reinforcement learning algorithm based on maximum entropy for the decision-making system of through-traffic traffic at unsignalized intersections in the park,and builds an end-to-end deep reinforcement learning framework.Firstly,after convolution processing,the camera image is input into the maximum entropy deep reinforcement learning network model based on the long short-term memory network,and finally the corner angle of the self-driving vehicle and the opening of the accelerator and brake pedals are output.Compared with the traditional maximum entropy deep algorithm and the deep deterministic policy gradient algorithm,this method can make the convergence speed of the model training faster.4.A test platform is built based on the automatic driving simulation software CARLA.After discretizing the trajectory of the steering vehicle that completed the scene generalization in Chapter 2,import it into CARLA to generate a test scene library.Then access the algorithm model built in Chapter4 to complete the test of the automatic driving decision-making system at the T-shaped intersection without signal lights in the park.The effectiveness of the algorithm is evaluated from two aspects of vehicle transit time and road traffic efficiency.Finally,the simulation results show that the algorithm has good performance in both vehicle transit time and road traffic efficiency.