Automated guided vehicle (AGV) (automated guided vehicle) refers to an unmanned transport vehicle that uses various electrical, magnetic, acoustic, and optical sensors as automatic guiding devices and can follow the preset guiding path . With the continuous development of smart factories and smart logistics, heavy-duty AGV is expected to become the key role of smart warehousing logistics system, realize the automatic transportation of materials in each production link, and ensure the efficient operation of the entire production line [2-3].
AGV's navigation and positioning accuracy and path adjustment ability are the bottlenecks restricting its application in the industrial field . The current navigation methods mainly include magnetic navigation [5-6], inertial navigation , laser navigation , visual navigation  and so on.
Magnetic navigation uses the principle of electromagnetic induction, and its navigation elements are diverse, such as landmark magnetic nails , eddy current coils and radio frequency devices , etc. Although magnetic navigation is widely used in the AGV industry, its laying costs are relatively high. It is not convenient for later maintenance and adjustment, and it is difficult to meet the needs of heavy-duty transfer in the modern flexible production process; inertial navigation technology has high positioning accuracy and strong flexibility, but it has high requirements for control algorithms and is easily affected by the surrounding environment; laser navigation Laser reflection plates need to be installed around the AGV driving path, which has very precise requirements on the installation angle and position. The cost is high, and it is susceptible to environmental interference. It is not suitable for complex factory environments; visual navigation uses image processing technology to navigate It is used in industrial heavy-duty AGVs with low economic cost and strong practicability. However, most of the traditional visual navigation methods use ribbon guidance and scanning code positioning. In practical applications, there are problems such as complex path laying and ribbons that are susceptible to environmental interference. During actual operation, the AGV's trajectory is prone to deviations due to non-linear factors and internal and external disturbances.
In order to improve the control accuracy of the AGV system, literature  combines PID and fuzzy control, and uses fuzzy rules to adjust PID control parameters online. Although the system has certain robustness, it is more adaptable to complex and changing application environments. difference. The adaptive inversion sliding mode control method proposed in literature  makes the system response faster, more robust, and has good instantaneous performance, but it is prone to chattering phenomenon for heavy-load AGV, which directly affects the control effect. Based on PID control, ADRC (active disturbance rejection control) technology considers non-linear factors and internal and external disturbances as total disturbances, and constructs an expanded state observer to estimate and compensate the total disturbances in real time to eliminate each The influence of various uncertain factors  has the advantages of few control parameters, fast convergence speed and good error compensation effect.
Based on this, this article intends to use three independent high-speed monocular cameras to improve the traditional visual navigation method based on ribbon guidance and scanning code positioning. In order to improve AGV's movement flexibility and navigation accuracy.
1 System design
The navigation flexibility and accuracy of the AGV during operation are the key factors to evaluate its system performance. The flexibility of navigation is related to the navigation method and AGV structure, and the navigation accuracy is directly related to the AGV control algorithm.
This paper improves the heavy-duty AGV structure that adopts the traditional visual navigation method, changing 1 or 2 monocular high-speed cameras in the ribbon guidance and scanning and positioning navigation to 3 independent monocular high-speed cameras (3C). The laying method is optimized. The improved AGV does not require ribbon guidance. It can be navigated only by scanning the code, and the AGV operates flexibly. In the AGV control algorithm, the ADRC strategy is used to compensate the disturbance in real time to eliminate the influence of various uncertain factors, so that the AGV runs stably and quickly responds, which can be applied to the complex workshop environment. The improved 3C vision navigation heavy-load AGV structure is shown in Figure 1, where the AGV center axis ①-③ position is installed with a high-speed monocular vision camera, and the center distance between camera 1 and camera 2 and between camera 2 and camera 3 The center distance is equal. The camera adopts a PGV optical camera, which can realize 360 ° turning on the spot through the design of dual steering wheel drive. To avoid conflicts, install a laser obstacle avoidance radar at the ④—⑦ position.
The structure of the AGV control system is shown in Figure 2. It is mainly composed of a battery management module, an obstacle avoidance module, a 3C visual navigation module, a servo drive and steering module, an on-board control module and a host computer control module. The control system adopts distributed control and is composed of two levels of microcomputers. The on-board control module adopts Siemens S7-1200 PLC. The host computer control module adopts industrial control computer. It can realize independent operation of single AGV and simultaneous operation of multiple AGVs.
1.1 3C visual navigation design
According to the structural characteristics of the distribution of the three cameras on the improved AGV, the laying method of the data matrix codes is designed. The center distance between two adjacent data matrix codes is equal to the center distance between the two cameras. The schematic diagram of AGV navigation is shown in Figure 3, in which the numbers 5-8 are the data matrix codes laid on the ground. When the AGV moves from station 5 to station 6, the AGV operation is divided into two steps: 1) Camera 1 and camera 3 simultaneously scan the information of station 6 and station 5, AGV slows down; 2) When the center position of camera 2 and station 6 When the coordinates of the center position coincide, the AGV stops.
In the course of traveling, AGV will inevitably have trajectory errors, so in order to ensure that AGV trajectory errors can be corrected in time, 3 independent cameras are running between stations (that is, from the matrix code entering the scanning area to leaving the scanning area), constantly scanning Site error, and transmit scanning error information to the host computer in real time. The control algorithm of the host computer revises its trajectory and sends it to the AGV for trajectory tracking, so as to realize AGV navigation. The vision camera is on the central axis of the AGV, and the AGV offset is composed of the data matrix code deviation value and the offset angle. During the operation of AGV, the deviation and offset angle when the camera scans a certain data matrix code at a certain time are shown in Figure 4. Figure 4 establishes a local coordinate system with the camera reading area and global coordinates with the work shop; the square area composed of 1, 2, 3, 4 is the data matrix code, and the deviation angle is α 0 . During the operation of AGV, as long as a camera scans the data matrix code, the front and rear steering wheels will automatically perform error correction at the same time. This correction method is more flexible and requires less computation.
The camera reads the center pose of the area as O (x 0 ', y 0' , α 0 ') in global coordinates, and transforms it Is the AGV deviation pose O (x 0 , y 0 , α 0 ). The center pose of the data matrix code is O '(x r ', y r' , α r '), which is converted into the AGV expected posture O' (x r , y r , α r ). The AGV expected trajectory kinematics model is
Among them, v r is the desired velocity, ω r is the desired angular velocity, and the value of α r can be 0 °, ± 90 °, 180 °. Converting the pose error under global variables to the AGV pose error formula under local variables  is
The differential equation of AGV pose error in local coordinates is
Among them, v 0 is AGV running speed, and ω 0 is AGV running angular speed.
1.2 ADRC tracker design
In the AGV system, ADRC is the control algorithm used for the host computer control module. The coordinate information collected by 3C visual navigation is transferred to the host computer, corrected by the ADRC tracker in the host computer, the corrected coordinate command is issued to the PLC, and then the AGV is controlled by the PLC.
1.2.1 ADRC mathematical model
ADRC tracker mainly includes differential tracker, expanded state observer and error feedback system. The differential state tracker (TD) linear equation of state is
Among them, r 0 is the tracking speed factor, the larger the r 0 , the faster the tracking speed; f 0 is the input signal; f 1 , f 2 is the tracking signal of f 0 .
The linear error equation of the expanded state observer (ESO) is
Among them, u is the input of the controlled system; y is the output of ADRC; z 1 , z 2 , z 3 are the estimated values of system state variables; β 01 , β 02 and β 03 are setting parameters; b is the control quantity coefficient. The error between the target and the output value in the error feedback control system and its differential error signal and the input of the controlled system, the corresponding calculation formulas are:
Among them, e 1 is error, e 2 is differential error, b 0 is adjustable parameter, u 0 The control law of the system.
1.2.2 AGV trajectory tracking controller design AGV trajectory
The structure of the tracking controller is shown in Figure 5. The AGV trajectory attitude error values [x e , y e , α e ] T is ADRC input, ADRC output is [x e , y e , α e ] T , The system output is the actual AGV running attitude [x 0 , y 0 , α 0 ] T , at this time The system is three input and three output, so three independent ADRC controllers are needed, and the trajectory and posture errors are regarded as three single input and single output systems.
The control laws of 3 independent ADRC u 0x , u 0y , u 0α are
Among them, k 1 and k 2 are proportional and differential control gains respectively; e 1x , e 1y , e 1α is the attitude error of three independent ADRC; e 2x , e 2y , e 2α are three independent ADRC differential attitude error.
2 Simulation results and analysis
In order to verify the performance of the ADC-based 3C visual navigation heavy-load AGV system designed in this paper, Matlab was used as the platform for simulation. Set the AGV speed to 1m / s, ADRC control parameters r 0 = 10, b 0 = 1, w = 10, β 01 = 30, β 02 = 300, β 03 = 1000, k 1 = 5, k 2 = 3. The circular and linear trajectory tracking curves and pose error curves are shown in Figure 6 and Figure 7, respectively. The horizontal and vertical coordinates of the trajectory tracking curve in the figure represent the position of the AGV in global coordinates; the horizontal coordinate of the pose error curve represents the AGV Running time, the ordinate represents the pose error of AGV in global coordinates.
It can be seen from Figures 6 and 7 that the starting position of the AGV is set outside the track. From the initial moment, the AGV in the circular track can successfully track the given reference track at 1.9S; The AGV in the linear trajectory can successfully track the given reference trajectory at 3.9s, indicating that the 3C visual navigation heavy-load AGV system based on ADRC has a faster response speed. After the operation is stable, the ideal tracking position error in the circular trajectory is less than 1mm, and the ideal offset angle error approaches 0; the AGV ideal tracking position error and ideal offset angle error in the linear trajectory approach 0. Therefore, the 3C vision navigation heavy-load AGV system based on ADRC can effectively realize the real-time tracking of the trajectory and the navigation accuracy is high.
3 Practical application results and analysis
In order to verify the operation of the 3C visual navigation heavy-load AGV system based on ADRC in practical applications, the actual operation test was carried out on site. The length of the AGV for testing is 1.8m, and the center distance between two adjacent cameras installed on the vehicle is 0.75m, so the center distance between two adjacent data matrix codes laid on the ground is also 0.75 m. In the set straight-line path, the path is 15m long, and a total of 22 data matrix codes are laid. To further test its turning performance, a zero-radius turn is performed at the starting position in the path. At the speeds of 0.5m / s, 1.0m / s, 1.5m / s, 2.0m / s, 2.5m / s and 3.0m / s, record the forward (backward) and left-turn (right-turn) of the AGV respectively The maximum navigation error and maximum deviation angle are repeated 50 times and then averaged. The results are shown in Table 1 and Table 2.
It can be seen from Tables 1 and 2 that in actual operation, the maximum navigation error when traveling straight is 7.44mm, and the maximum offset angle is 0.89 °; when turning 90 ° in place, the maximum navigation error is 7.21mm, and the maximum offset angle is 0.92 °. This result has a certain gap with the ideal accuracy in the simulation results, which is related to the actual ground flatness, smoothness, AGV car body manufacturing process and other factors. It can be seen that the absolute value of the maximum navigation error of the 3C vision navigation heavy-load AGV system based on ADRC is less than 8mm, and the absolute value of the maximum offset angle is less than 1 °. The navigation mode of this system is simple, the data matrix code is easy to lay, the navigation accuracy is high, and the AGV operation is stable and flexible.
In this paper, the traditional visual navigation heavy-load AGV structure based on ribbon guidance and scanning code positioning is improved, and the 3C visual navigation heavy-load AGV system based on ADRC is designed. The design uses 3 independent monocular cameras, without ribbon guidance, only need to lay the data matrix code to achieve navigation, and ADRC is used in the navigation control algorithm to effectively eliminate various external interference. The simulation and practical application results show that the AGV is stable and flexible, with fast response speed, the absolute value of the maximum navigation error is less than 8mm, and the absolute value of the maximum offset angle is less than 1 °. Similar products with low manufacturing cost and high engineering practical value. Based on the research results, the next step will be to introduce artificial intelligence algorithms in ADRC control, use intelligent algorithms to automatically adjust the constant parameters that need to be set in ADRC, and turn them into dynamic adjustment variables to further improve the convergence speed of ADRC and reduce the need Parameter setting to further improve AGV operation response speed.
 BACIK J, DUROVSK YF, BIR0S M, et al. Pathfinder-development of automated guided vehicle for hospital logistics [J]. IEEE Access, 2017, 5: 26892.
 Guo Jinchao, Zhao Haiyang, Jiang Zhengke, etc. Two-way heavy-duty intelligent autonomous navigation vehicle system design [J]. Journal of Light Industry, 2017, 32 (2): 97．
 Gao Yu, Guo Jinchao, Cui Guangzhao. An improved multi-robot path planning adaptive artificial potential field method [J]. Journal of Zhengzhou University of Light Industry (Natural Science Edition), 2013, 28 (6): 77.
 Xiao Xianqiang, Cheng Yabing, Wang Jiaen. Research and design of AGV based on inertial and visual composite navigation ［J］. China Mechanical Engineering, 2019, 30 (22): 1．
 LEE S Y, YANG H W. Navigation of automated guided vehicle suing magnet spot guidance method [J]. Robotics & Computer Integrated Manufacturing, 2012, 28 (3): 425.
 LU S, XU C, ZHONG R Y. A RFID-enabled positioning system automated guided vehicle for smart factories [J]. Journal of Manufacturing Systems, 2017, 44: 179.
 JUNG K, KIM J. Intelligent autonomous systems [M]. Berlin Heidelberg: Springer, 2013: 807-816.
 OSMAN K, GHOMMAM J, SAAD M. Vision based lane reference detection and tracking control of an automated guided vehicle [C] ∥IEEE Control Systems Society. 2017 25th Mediterranean Conference on Control and Automation. Piscataway: IEEE, 2017: 595.
 XU Z, HUANG S, DING J. A New positioning method for indoor laser navigation on under determined condition [C] ∥IEEE Control Systems Society. 2016 Sixth International Conference on Instrumentation & Measurement, Computer, Communication and Control. Piseatamay: IEEE, 2016: 703.
 Wang Linhua. Design of the control system of the magnetic conduction AGV automatic navigation vehicle [D]. Changsha: Changsha University of Science and Technology, 2013.
 Jiang Yafeng, Wang Binbin, Yuan Mingxin, etc. Omnidirectional AGV motion control based on adaptive inversion sliding mode ［J］. Computer Simulation, 2019, 32 (2): 348.
 Han Jingqing. Auto disturbance rejection control technology—control technology that estimates and compensates for uncertain factors [M]. Beijing: National Defense Industry Press, 2008: 197-207.
 Luo Rui. Trajectory tracking of mobile robot based on auto disturbance rejection control [D]. Tianjin: Tianjin University of Technology, 2018.