Research On Automatic Parking Control Strategy Based On Deep Reinforcement Learning

At present,the development center of the automobile industry is gradually shifting from traditional vehicles to intelligent vehicles,and the rapid development of intelligent driving technology is accompanied by it.Automatic parking system is one of the research directions of intelligent driving technology.By giving driving guidance or completely replacing the driver for parking operation during parking,the parking success rate is effectively improved and the probability of traffic accidents is reduced.At the same time,due to the small parking spaces and irregular parking,drivers often face parking difficulties.Therefore,the automatic parking system has become a research hotspot of major car companies and universities.First of all,this paper analyzes and summarizes the research status of automatic parking system,selects the parking control strategy in the automatic parking system as the research focus of this article,and summarizes the existing automatic parking control strategy and existing problems.The parking coordinate system is constructed based on the Gaussian coordinate system and the traditional vehicle coordinate system.Based on the kinematic principles of the vehicle,the car parameters and parking space parameters are simplified.The kinematics model of the car is established based on Ackerman steering principle,Taking parallel and vertical parking as parking scenarios,the kinematics model is analyzed for motion.Taking the vehicle’s extreme turning characteristics as an entry point,the minimum parking space required for parking is studied.Secondly,the deep reinforcement learning algorithm in the field of artificial intelligence is studied.DDPG(Deep Deterministic Policy Gradient)algorithm is one of deep reinforcement learning algorithms,with certain learning ability and generalization performance,this article applies it to automatic parking control strategy to achieve end-to-end vehicle parking control.using Tensorflow as the framework Construct the DDPG algorithm model,taking the vehicle state as the state information and the vehicle control command as the action information,and output the action information through the vehicle state information to control the vehicle to complete the parking.Focusing on the sampling strategy and reward function,an efficient sampling strategy is proposed based on the empirical playback method and data structure,and a reward function based on trajectory learning is designed based on the preview model and Coulomb’s law to complete the DDPG algorithm.Improvements have improved the training efficiency and training effect of the DDPG algorithm.Finally,this paper uses CarSim vehicle dynamics simulation software to build a simulation and algorithm training platform.Through the co-simulation with Python,the vehicle dynamics model and the DDPG algorithm model are trained to verify the effectiveness of the control strategy based on the DDPG algorithm Sex.Design multiple sets of parking generalization test,adjust the parking starting position and parking space size in the test,so as to carry out the parking test,Verify and analyze the generalization of the algorithm.The vehicle interface in CarSim is hardwareized.The hardware-in-the-loop platform uses Lab VIEW to write the steering program and completes the control of the actuator through CAN communication,so as to control the data exchange between Python and CarSim,and conduct the test through the hardware-in-the-loop platform.The test results show that the vehicles can complete parking and the steering wheel rotates smoothly,which verifies the functionality and effectiveness of the automatic parking control strategy designed in this paper.