Research On Autonomous Obstacle Avoidance Method Of Field Intelligent Agricultural Vehicle Based On Fusion Navigation And Reinforcement Learning Algorithm

As a branch of agricultural machinery automation,automatic navigation and obstacle avoidance of agricultural machinery are of great significance for improving agricultural production efficiency,saving labor costs and promoting agricultural intelligent management.However,there are still many limitations in the autonomous navigation system of agricultural machinery at this stage,and it cannot respond to the dynamic environment during operation.Based on the existing navigation system based on predefined paths,this paper studies the autonomous obstacle avoidance control method of intelligent agricultural machinery in the field,and conducts verification experiments in the field.The main research contents and research results are as follows:(1)An intelligent agricultural unmanned mobile platform has been constructed.The mobile platform includes four subsystems: satellite positioning system,visual perception system,decision control system and motion actuator.The satellite positioning system and visual perception system can realize multisource information perception of agricultural machinery and the environment,and the host computer in the decision control system is real-time Obtain information such as body position,body heading,obstacle position,etc.,and send motion control instructions to the lower computer through the motion decision model.The lower computer controls the motion actuator according to the desired control instructions to achieve desired driving requirements.The development of this intelligent agricultural unmanned mobile platform provides a platform guarantee for the development of navigation and obstacle avoidance control algorithm research.(2)A steering control subsystem and method based on fusion navigation are developed.According to the crop row images collected by the visual perception system,a navigation line extraction method is proposed,and a PID controller is designed to calculate the steering control amount.At the same time,the desired operation path is planned according to the characteristics of farmland planting,and the steering control amount is calculated based on the pure tracking algorithm according to the position information and heading information obtained by the satellite positioning system.Finally,the steering control quantities calculated by the two methods are fused into the final steering control quantity based on the linear fusion model.In addition,for the complex paddy field environment,a method for identifying the rows of paddy fields under strong interference environment based on the three-class Otsu method and the multidimensional clustering algorithm was proposed,and its feasibility was explored.The test results show that the average attitude error can be kept within 0.02°,and the average distance error can be kept within 10 pixels.This method provides data support for reliable vision-based navigation in complex paddy field environments.(3)An autonomous obstacle avoidance control subsystem based on reinforcement learning is developed.The YOLOv4 algorithm is used to identify moving obstacles such as field laborers and animal forces,and the spatial position information of the obstacles is extracted from the point cloud information obtained by the visual perception system.Based on the continuous temporal and spatial changes of obstacles,the motion state discrimination of obstacles is realized.The verification experimental results show that the obstacle detection system can provide positioning accuracy within 0.11 m within a 10 m field of view.In addition,the system summarizes the different encounter situations between moving obstacles and intelligent agricultural machinery,and builds a reinforcement learning model based on the Double DQN framework,which determines the state space,action space,and reward and punishment functions of the model.Based on the Mu Jo Co physics engine,a simulation environment is constructed to simulate different encounter states of obstacles,and to carry out training and testing of obstacle avoidance models.The test results show that the obstacle avoidance success rate of the constructed obstacle avoidance model under different encounter situations in the simulation environment is 97%,96%,and 98%,respectively,and the return rates are 12.57,12.06,and 12.34,respectively,which proves that the model has good performance.Stability and better obstacle avoidance control strategy.(4)The verification experiment of field automatic navigation and autonomous obstacle avoidance was carried out.The experimental results of automatic navigation show that the average heading deviation is 4.7° and the average lateral deviation is 8.65 cm.The steering control subsystem based on fusion navigation can meet the needs of most field operations.The experimental results of autonomous obstacle avoidance show that the obstacle avoidance control system based on Double DQN reinforcement learning can stably avoid obstacles when faced with three different encounter situations: cross encounter,small angle encounter and large angle encounter.The length of the obstacle avoidance trajectory,the average shortest obstacle avoidance distance and the average obstacle avoidance time are 0.9m,2.58 m and 6.1s,respectively.Compared with the traditional scheme,it has advantages in space and time utilization.The field experiment verifies the feasibility and effectiveness of the steering control subsystem and obstacle avoidance control subsystem designed in this paper,and provides a reliable improvement reference for the existing agricultural machinery automatic navigation system.