Research On AGV Localization And Autonomous Navigation Based On Multi-sensor Fusion

With the development of intelligent manufacturing,AGV industry has achieved leapfrog development.The increasingly complex and diversified industrial flexible production lines put forward higher requirements for the application of AGV.There are common or special problems in AGV localization and map building,path planning,obstacle avoidance in moving obstacle environment,and motion balance in the loading process.Therefore,these problems are studied in this paper.In the aspect of AGV localization and map building,the main research contents include the error analysis of the original A-LOAM algorithm odometer in the linear constraint experiment;Wheel odometer factor experiment and calibration,and error analysis of the IMU sensor;Then,based on the wheel odometer,IMU and the laser＿odom in the A-LOAM algorithm,the EKF algorithm is fused to obtain the fused odometer and map are obtained.At the same time,to overcome the shortcomings that the EKF fusion algorithm needs to determine the statistical characteristics of sensor noise in advance,and the statistical characteristics of noise are prone to interference in complex and changeable environments,this chapter also studies the training model based on 1D-CNN neural network,uses the highprecision map obtained by the above EKF fusion multi-sensors,uses the map odometer to participate in the construction of training set,and fuses the wheel odometer(odom)The position and attitude(AMCL)and the angle of inertial unit(IMU)are located by Monte Carlo algorithm of two-dimensional lidar to achieve two-dimensional high-precision localization.In the aspect of AGV path planning,aiming at the problem that the map area of a*algorithm increases and the efficiency decreases,and the planning path makes AGV easy to friction and collision near obstacles,inspired by the canyon landform and river trend represented by Colorado Canyon,a variable scale Canyon algorithm is proposed,which comprehensively uses the variable grid-scale and a* algorithm to obtain the final global path.The experimental results show that the planned path can improve the problem that the shortest path of a* algorithm is close to obstacles,abandon some narrow paths,let the AGV preferentially choose the safe path that is biased towards the center of the path,and improve the security of path planning without significantly increasing the path length.At the same time,cubic B-spline local path optimization is carried out for the planned safe path to obtain a smooth optimized path.Aiming at the situation that the omnidirectional gear train AGV is prone to interference and collision at a specific narrow corner,a double line constrained turning path planning algorithm is proposed to optimize the turning path.By constraining the turning process,the center points of the front and rear end faces of the AGV body lie on two specific lines parallel to the channel,so that the vehicle body does not deviate to one side of the channel during the turning process.The experimental results show that the AGV turning at the narrow and dangerous corner can be controlled,Steering safety can be ensured.In view of the shortcomings of the traditional DWA obstacle avoidance algorithm in the environment of moving obstacles,this chapter puts forward the idea of calculating the critical speed obstacle avoidance: first,identify the moving obstacles,use the yolov3 target detection algorithm to train the image data set marked with moving obstacles,obtain the moving obstacle recognition network model,and calculate the angle of the moving obstacles relative to the obstacle avoidance AGV according to the recognition frame angle;The angle and frame angle of moving obstacles are combined with the point cloud collected by lidar,and the velocity and direction of moving obstacles are calculated through the operation of adjacent frame point cloud;At the same time,in order to verify the accuracy of moving obstacle detection and angle detection in the critical speed obstacle avoidance algorithm,as well as the accuracy of moving obstacle speed calculation,the yolov3 neural network model target detection and angle detection algorithm verification experiments and moving obstacle speed calculation experiments are carried out.The results show that the experimental accuracy can meet the requirements of subsequent algorithms;Finally,the collision model between AGV and moving obstacles is established to solve the relationship of parameters and the critical obstacle avoidance speed of AGV,and the critical speed is used to avoid obstacles.The results show that the critical obstacle avoidance velocity can avoid obstacles by accelerating or decelerating through the collision area without changing the direction of AGV movement.The last chapter carries out innovative research on the dynamic balance of cargo AGV during operation.In view of the possible cargo slip and overturn during the operation of the cargo carrying AGV,a multi-sensor safety monitoring system platform for the cargo carrying AGV is built.By establishing the motion balance model of the cargo carrying AGV,the critical acceleration values of the cargo slip and overturn under the straight,rotation,turning and straight uphill motion states are analyzed,and the center of gravity position of the cargo affecting the stability of the cargo is calculated by using the pressure sensor and the inertial measurement unit;Then the neural network limit learning machine training model is used to classify the odometer and inertial measurement unit data,monitor the AGV motion state,feedback the acceleration in real time and constrain the acceleration.The results show that the platform can monitor the running state of AGV in real time and restrict the acceleration of the system to ensure the safety and stability of cargo transportation.