Research And Application Of Intelligent Car Decision Planning Method Combining Dynamic Scene Information And DDPG Algorithm

The goal of the decision-making and planning module used in self-driving cars is to cooperate with other modules so that the vehicle can perform autonomous driving in a safe,comfortable and complying with traffic rules.The decision-making and planning module is usually implemented by two sub-modules: the decision-making module,which is used to generate high-level decisions to give the behavior of the vehicle;the planning module,which generates a smooth and drivable vehicle trajectory.Although the research work of decision-making and planning has made a lot of great progress,these modules are usually developed separately and the applicable scenarios are relatively simple.The result is that using the existing public decision-making and planning methods,it is easy to fail in scenes with many dynamic obstacles,complex roads,and changeable weather environments,and it will be impossible to output safe and stable behavior trajectories.In order to solve this problem,this paper combines the current network learning algorithm’s strong adaptability and traditional methods with good interpretability and smooth results,and proposes a smart car decision planning algorithm based on a dynamic environment,focusing on the vehicle in a straight line.Decision-making and planning tasks under driving,turning and dynamic traffic,with the goal of traffic efficiency and driving safety.Unlike most existing design ideas that only solve decision-making problems or planning problems,or only consider static scenes or the movement of a single obstacle,this article proposes decisions in dynamic scenarios from the perspectives of decision-making and planning,Plan the solution,and finally combine the decisionmaking method with the planning method,and conduct a joint test of the decision-making and planning algorithm in a dynamic scenario.The main research of this article is as follows:(1)Use deep reinforcement learning methods that integrate dynamic scene information to replace traditional decision-making methods.This paper uses the modelfree and self-learning Deep Deterministic Policy Gradient(DDPG)algorithm as a framework to integrate four kinds of dynamic information: images,obstacle perception results,global path planning results,and vehicle real-time status to achieve behavioral decision-making tasks for smart cars.(2)Optimize and improve traditional planning algorithms in combination with the characteristics of dynamic scenarios.In this paper,the motion planning algorithm based on polynomial curve is used to realize the trajectory planning task of the smart car,and the evaluation function of the trajectory is optimized and redesigned in combination with the motion characteristics of the own vehicle and obstacles in the dynamic scene.(3)Use the simulation environment and real vehicle platform to test and verify the decision-making planning system and analyze the experimental results.First,let the results of the decision module proposed in Chapter 3 be expressed in the Frenet coordinate space,and then use the decision results in the Frenet coordinate system as the constraints of the planning algorithm in Chapter 4 to generate the trajectory of the vehicle in the Frenet coordinate system.Then use the Carla simulator to generate a high-simulation simulation environment and a real vehicle platform to test and verify the decision-making planning system of this article.