CN112149555B

CN112149555B - Global vision-based multi-warehouse AGV tracking method

Info

Publication number: CN112149555B
Application number: CN202010998015.7A
Authority: CN
Inventors: 谢巍; 杨启帆; 廉胤东; 周雅静; 林丹淇; 王锴欣
Original assignee: South China University of Technology SCUT
Current assignee: South China University of Technology SCUT
Priority date: 2020-08-26
Filing date: 2020-09-21
Publication date: 2023-06-20
Anticipated expiration: 2040-09-21
Also published as: CN112149555A

Abstract

The invention discloses a multi-warehouse AGV tracking method based on global vision, which comprises the following steps: 1) Shooting a global image of a warehouse and sending the global image to a control center; 2) Processing the global image, identifying a plurality of AGVs by using a target detection algorithm in a first frame by using a tracking system, and determining the ID and the pose of each AGV according to an AprilTag code at the top of the AGV; 3) Dividing a warehouse into a plurality of areas, and carrying out path planning of each AGV by using a hierarchical planning algorithm after a scheduling system acquires the position and pose information of the AGVs sent by a control center; 4) Combining the path information of each AGV with a multi-AGV tracking algorithm, predicting the position of the AGV, selecting the area of the AGV by using a boundary frame, and determining the information of each AGV; 5) And the control center transmits a speed instruction converted from the path information to the AGV to control the AGV to complete the sorting task of the cargoes.

Description

Global vision-based multi-warehouse AGV tracking method

Technical Field

The invention relates to the field of warehouse logistics for assisting in cargo sorting, in particular to a multi-warehouse AGV tracking method based on global vision.

Background

At present, online shopping has become trend, but the logistics industry is prosperous along with electronic commerce, but the logistics industry is high in song and is also challenged, for example, the large throughput of goods in a logistics warehouse system causes sorting and transportation work in the warehouse to become a problem, but the investment of a large amount of labor cost is also insufficient to solve the problems of low sorting efficiency and high error rate, even the phenomenon of violent sorting can occur, and the heavy work in the logistics warehouse at present cannot be well dealt with only by manpower, so that an automatic warehouse system is gradually prosperous. The intelligent storage system utilizes an intelligent navigation Vehicle (AGV) with strong flexibility to replace manual goods sorting and transportation, not only saves resources such as manpower and the like, but also can improve sorting efficiency and reduce error rate, most of the current intelligent navigation vehicles need to be provided with expensive sensors such as cameras or laser radars to collect information and independently communicate with a control system, and a warehouse also needs to be correspondingly provided with devices such as magnetic nails and magnetic tapes, so that the problems of high cost, complex configuration, limited control efficiency and the like (Lynch L, new T, clifford J, et Al.automated Group Vehicle (AGV) and Sensor Technologies-A Review [ C ]//2018 12th International Conference on Sensing Technology (ICST). 2018.) exist. Some intelligent warehousing systems use global visual monitoring (CN 201911046869.9) to track and schedule by recognizing AGV feature marks, but the problems of high resolution or illumination requirements, low fault tolerance, less information storage amount and the like of the feature marks still exist, meanwhile, the traditional detection and tracking mode has low efficiency (Dnmez E, kocamaz AF, dirik M.A Vision-Based Real-Time Mobile Robot Controller Design Based on Gaussian Function for Indoor Environment [ J ]. Arabian Journal for ence & Engineering,2017 (4): 1-16 ]), so the intelligent warehousing system Based on global Vision is also improved in recognition and tracking.

Disclosure of Invention

The invention discloses a multi-warehouse AGV tracking method based on global vision, which has lower configuration cost and higher operation efficiency compared with the traditional AGV warehouse system. Through installing the camera at the warehouse top and shooting global image control many AGVs, utilize AGV top feature code to detect the discernment, divide into a plurality of regions with the warehouse, dispatch according to the transportation task needs to fuse scheduling algorithm in the tracking process based on target detection, the system operation is smooth, can track and dispatch many AGVs accurately in real time, accomplish letter sorting transportation task high-efficiently.

The invention is realized at least by one of the following technical schemes.

A multi-warehouse AGV tracking method based on global vision comprises the following steps:

1) A camera at the top of the warehousing system shoots a global image of the warehouse and sends the global image to a control center for tracking of the AGV;

2) The tracking system of the control center processes the global image, identifies a plurality of AGVs by using a tiny_yolv3 target detection network in a first frame, and determines the ID and the pose of each AGV according to an april tag code at the top of the AGVs;

3) Dividing a warehouse into a plurality of areas, and performing path planning of each AGV by using a hierarchical planning algorithm after a scheduling system of a control center acquires the position and pose information of the AGVs sent by a tracking system;

4) After the dispatching system plans the path of each AGV, the tracking system directly predicts the position of the AGV by using the path information, then a dynamic window is set to select the area where the AGV is located, the AprilTag code in the area is identified to determine the information of each AGV, and if the detected AGV number is less than the current AGV number in the warehouse, the step 2) is executed again;

5) The control center sends a speed instruction converted from the path information to the AGV, and controls the AGV to reach the target area according to the set route so as to complete the sorting task of cargoes.

Further, in step 1), the top camera is placed at the top of the warehouse and used for monitoring and shooting images of the whole warehouse, wherein a control center of the warehouse system comprises a multi-AGV tracking system and a multi-AGV scheduling system, and the control center is used for determining pose information and specific paths of each AGV.

Further, the step 2) specifically includes the following steps:

firstly, when a warehouse system is started, as pose information and path information of an AGV do not exist, a tiny_yolov3 target detection network is used for multi-AGV identification, and a tiny_yolov3 target detection network is trained by using AGV pictures, pedestrian pictures and common obstacle pictures in a warehouse;

in order to identify the characteristic pattern at the top of the AGV, performing image processing on the AGV area selected by the frame, wherein the specific processing mode is that firstly, the area image is converted into a gray image, then, binarization processing is performed by using Opencv, and then, the object contour is extracted by using a Canny edge detection algorithm and a FindContourr function;

fitting the outline by using an appxpolydp function, and selecting a characteristic pattern which has a proper area and is judged to be square as the characteristic pattern of the current AGV, wherein the characteristic pattern selects an open-source april tag system;

step four, rasterizing the whole square area based on four corner points of the outer contour of the AprilTag code at the top of the AGV, decoding the AprilTag according to the pixel values of each grid in the area, and further obtaining the binary information stored by the current AprilTag code as the ID of the current AGV;

step five, after the tracking system extracts the outline set of the pattern, fitting each side of the square by using a least square method, obtaining a rotation angle through the slope of each side, and then weighting to obtain a proper angle:

wherein alpha is _i For the deflection angle corresponding to the ith side, n is a set of side points (x _j ,y _j ) The current pattern is rotated by alpha degrees to obtain any one forward characteristic pattern, but the direction of the AGV is unknown at the moment, so that the pattern can be sequentially rotated by 90 degrees according to the rotation invariance of the AprilTag to finally obtain the standard characteristic pattern in the family code, and then the angles of the two parts are added to obtain the course angle of the AGV, wherein the course angle is shown in the following formula:

θ＝90°×r+(45°-∝) (2)

wherein alpha is an auxiliary angle for calculating the course angle of the AGV, and the value of the auxiliary angle is equal to the average value of included angles of four sides of the AprilTag code; r is 0-3, which means that the current identification is matched with the family code through r times of 90-degree rotation energy, k _i The weight of the angle is calculated for the ith edge least squares method, where each edge is set to have the same weight.

Further, step 3) includes:

the warehouse is divided into two layers for classification and storage, wherein the second layer is a sorting layer, and the first layer is a centralized storage layer; dividing a sorting area into n areas, wherein each area comprises an intersection and a corridor connected with the intersection, the areas are provided with a plurality of grids which are the most basic units for AGV scheduling, and the grids are divided into an initial area and a target area in a sorting layer according to different division of the areas; in the task execution process, each AGV obtains an optimal path reaching a target area by utilizing a hierarchical planning according to the assigned task, the AGV inputs carried packages into a pipeline after running to the target area according to the path, the goods enter a first layer for centralized storage, and the AGV returns to an idle initial area for next task circulation after the sorting task is finished;

further, the hierarchical planning refers to ensuring the scheduling efficiency while reducing the algorithm complexity when performing AGV path planning, converting a complex multi-objective planning problem into two sub-problems, and calculating by using two algorithms respectively; the first layer is planned to be a plan among areas, an area where the AGV is currently located is set as an initial area, an area where the package needs to be delivered is set as a target area, and a shortest area path set for connecting the initial area and the target area is calculated by using an A-path planning algorithm; and planning among grids for the second time, and distributing grid resources in the same area by utilizing a time window arrangement algorithm, so that the AGV can quickly leave the current area according to the area path set on the basis of no collision.

Further, the step 4) is specifically as follows:

under the premise that all AGVs are known to be located and path information is obtained, the detection is not carried out by using a tiny-yolov3 target detection network, the current AGV position is directly predicted according to the path planned by a dispatching system, a dynamic window is arranged on the basis of the predicted position to frame and select the possible area of the AGVs, the operation of the step 2) is carried out on the area, the AprilTag code at the top of the AGVs is utilized to carry out positioning, after all AGVs are detected, if the number of the identified AGVs is inconsistent with the initial number, the AGVs do not operate according to the established track, at the moment, the tiny-yolov3 target detection network is reused to identify the AGVs in the global image, any of the AGVs can be repositioned as long as the AGVs are still in the current storage environment, and the AGVs are dispatched at the control center to return to the established track again.

Further, step 5) is specifically as follows:

in order to enable a large-scale AGV in a warehouse to stably run, the AGVs are set to run at a constant speed, so that the linear speed sent to each AGV by a control center is a fixed value, in addition, the control center changes the angular speed which is required to be sent to the AGVs in real time according to the course angle and the coordinates of the AGVs calculated by a multi-AGV tracking system, and corrects the angular speed on the basis of the linear speed in such a way that if the current position of the AGVs is far away from a given route, the linear speed value is increased, and if the current position of the AGVs is far away from the given route, the linear speed is reduced, so that the AGVs are controlled not to deviate from the given route.

According to the invention, the opencv is used for processing the global image and the tiny_yov3 is used for target detection, the april tag feature code is used for identifying the AGV, so that the identification rate and fault tolerance are ensured, the target detection algorithm is combined with the path planning algorithm, tracking is performed based on prediction, and the overall planning efficiency is improved.

Compared with the prior art, the invention has the following advantages:

(1) By utilizing a machine vision technology, the central controller acquires all AGV information by processing the global image shot by the top camera, so that the calculated amount of the AGV vehicle-mounted controller is reduced, the information interaction amount is reduced, and the overall efficiency of the system is improved;

(2) The neural network is combined with the traditional image processing technology to carry out multi-AGV tracking, the neural network is utilized to ensure the positioning accuracy of the tracking algorithm, the processing speed of the algorithm is accelerated through the traditional image technology, and finally the purpose of accurately tracking a plurality of AGVs in real time is achieved.

(3) The AGV is identified and positioned by using the april tag system, so that the accuracy is higher;

(4) The fuzzy predicted position is obtained by combining the historical position information of the AGVs and the planned path, the image of the whole warehouse is not required to be processed, meanwhile, whether the AGVs derail or not can be monitored, unnecessary information processing is reduced, and the tracking efficiency is improved;

(5) If the AGV breaks away from the track, the control system can detect the AGV in the global scope again, so that the AGV which breaks away from the established route returns as soon as possible, and manual operation is reduced.

Drawings

Fig. 1 is a schematic diagram of a warehouse system based on global vision according to the embodiment;

FIG. 2 is a Gazebo simulation diagram of the warehouse system based on global vision according to the present embodiment;

FIG. 3 is a diagram of a warehouse system based on global vision according to the present embodiment;

FIG. 4 is a flow chart of the multi-AGV tracking system of the present embodiment;

FIG. 5 is a diagram of a simulation warehouse topology and a regional topology map based on global vision according to the present embodiment;

FIG. 6 is an AGV with an april tag code on top of the present embodiment;

FIG. 7 shows the april tag code of the present embodiment;

FIG. 8 is a tracking result of the multi-AGV tracking system of the present embodiment;

FIG. 9 is a graph comparing deviation of AGV course angle calculated by different tracking modes.

Detailed Description

The present invention will be described in further detail with reference to examples and drawings, but embodiments of the present invention are not limited thereto.

FIG. 1 is a diagram of a global vision based multi-AGV warehousing system including: image acquisition equipment (industrial camera), wiFi, data storage server, computer (control center), storage AGV.

The image acquisition equipment is arranged at the top of the warehouse and is used for acquiring global images of the warehouse, and the acquired global warehouse images are transmitted to the data storage server for storage through a network. The computer is connected with the data storage server through the Internet, and can directly read or delete task information of the stored images and packages in the server; the computer (control center) includes many AGVs tracking system and many AGVs dispatch system, and control center can handle and analyze the image information that reads, tracks the pose information of AGVs in the warehouse system, in sending the dispatch system with the information of all AGVs simultaneously, the control information of every AGVs of output, control command passes through the WIFI transmission to corresponding AGVs in, the AGVs moves according to control command, delivers the express delivery parcel to the pipeline in for centralized storage, and then reaches the purpose of sorting parcel.

Fig. 2 is a global vision-based warehouse system built in Gazebo simulation software. The warehouse is divided into two layers for classification and storage. The AGV transports the package on the correct conveyor and delivers it to the target area. After the sorting task is completed, the AGV will return to the starting area to transport other packages and begin the new task again. In fig. 1, there are four cameras at the top of the warehouse, each of which photographs a corresponding area. Then, the warehouse is divided equally into four parts, and the images photographed by the plurality of cameras are spliced to obtain a global image of the warehouse.

The multi-warehouse AGV tracking method based on global vision as shown in FIG. 3 comprises the following steps:

1) Shooting a global image of a warehouse by using a camera at the top of the warehouse system, and sending the global image to a control center for tracking of the warehouse AGV;

2) The area where each AGV is located is selected by using a tiny_yolv3 target detection network or a scheduling system, and the april tag codes in the areas are sequentially identified to obtain coordinates, heading angles and numbers of the AGVs, as shown in fig. 4, the specific process is as follows:

dividing a simulated warehouse system into a plurality of areas for AGV management and scheduling; the specific process is as follows:

as shown in fig. 5, the current simulation warehouse is divided into 16 areas, respectively (Q ₁ ,Q ₂ …,Q ₁₆ ). Each zone contains an intersection and hallway connected thereto. As shown in the right diagram indicated by the arrow in fig. 5, each area (area 6, area 7, area 10, area 11) containing the crossroads is composed of 4 corridors (Cor 1, cor2, cor3, cor 4) and one intersection, the basic composition unit of each of which is a grid, and in order to prevent a plurality of AGVs from colliding during the dispatching process, the area of the grid is slightly larger than that of the AGVs. The white part in the warehouse map is a free area which can pass through, the area 4, the area 8, the area 14 and the area 16 are the starting area of the AGV task, the black area is a pipeline for transporting cargoes, and the AGV needs to transport the cargoes to the corresponding area and then throw the cargoes into the pipeline to finish the sorting task.

Step two, processing the global image, identifying a plurality of AGVs by using a tiny_yolv3 target detection network in a first frame by using a tracking system, and determining the ID and the pose of each AGV according to an april tag code at the top of the AGV, as shown in fig. 6; the specific process is as follows:

when the warehouse system is started, as no pose information and no path information of the AGVs exist, a tiny_yolv3 target detection network is used for multi-AGV detection. The Tiny_YOLOv3 target detection network is trained with AGV pictures, pedestrian pictures, and pictures of common obstructions within the warehouse. Through the tiny_yov3 target detection network, all AGVs can be identified in the warehouse global image of the first frame, and the system can also detect warehouse staff and other obstacles and send the information of the AGVs to the control center.

In order to identify the characteristic pattern at the top of the AGV, the system performs image processing on the AGV area selected by the frame, firstly converts the area image into a gray level image, then performs binarization processing by using Opencv, and then extracts the object contour by using a Canny edge detection algorithm and a FindContour function;

fitting the contour by using an appxpolydp function, and selecting a characteristic pattern which has a proper area and is judged to be square as the characteristic pattern of the current AGV. The april tag system with the characteristic pattern selection open source has advantages in recognition rate and fault tolerance compared with other characteristic codes, and different types of april tag codes such as 16h5,25h9,36h11 and the like can be selected according to the required stored information quantity and the requirements on the characteristic code recognition rate and fault tolerance. Wherein the first number is the number of bits of information that the april tag code of this type can store, and the latter number is the smallest hamming distance between any two codes in the family code of this type. Taking 36h11 as an example, the number of bits of each codeword in the family code is 36, and the minimum Hamming distance between any two codewords is 11. The family code can correct 5-bit errors and detect 5-bit errors. Meanwhile, the method has rotation specificity, namely when one characteristic pattern rotates for a plurality of 90 degrees, the Hamming distance between the decoding result and the family code is still larger than a set value, and only the characteristic code in a unique direction can be matched with the decoding result of the family code, so that the characteristic can be used for determining the rotation angle of the characteristic code. In addition, unlike two-dimensional codes, the two-dimensional code positioning method is positioned through the peripheral black rectangular frame, and compared with three positioning patterns of the two-dimensional code, the two-dimensional code positioning method has better robustness under the condition of low resolution.

The entire square area is then divided into 64 grids based on the four corner points of the april tag code outline at the top of the AGV, i.e., points a, B, C, D in fig. 7. The april tag is decoded according to the pixel value of each grid in the area (the black grid indicates that the bit is 1, and the white grid indicates that the bit is 0), and further the binary information stored by the current april tag code is obtained and used as the ID of the current AGV.

Because the corner points of the characteristic patterns are easy to blur, the method is different from the conventional method that the april tag uses the corner points to determine the rotation angle of the patterns, and the tracking system uses a least square method to fit each side of the square after extracting the contour set of the patterns. The rotation angle is obtained by the slope of each side, and then the rotation angle is weighted to obtain a proper angle.

Wherein alpha is _i The deflection angle corresponding to the ith edge is shown in fig. 7. n is a set of edge points (x) _j ,y _j ) Is a number of (3). Any positive characteristic pattern can be obtained by rotating the current pattern by alpha degrees, but the direction of the AGV is unknown at the moment, so that the pattern can be sequentially rotated by 90 degrees according to the rotation specificity of the april tag, and finally the standard characteristic pattern consistent with the family code is obtained. And then adding the angles of the two parts to obtain the course angle of the AGV, wherein the course angle is shown in the following formula:

θ＝90°×r+(45°-∝) (2)

wherein alpha is an auxiliary angle for calculating the course angle of the AGV, and the value is equal to the average value of the included angles of four sides of the AprilTag code. r is 0-3, which indicates that the current identification can be matched with the family code through r times of 90-degree rotation. k (k) _i The weight of the angle is calculated for each edge least squares method, where each edge is set to have the same weight.

The effect of the multi-AGV tracking system is shown in FIG. 8, wherein 20 AGVs in the warehouse system can be selected by the frame, and then the course angle and the number of the AGVs are output in real time for the multi-AGV dispatching.

Step three, after the serial number and the attitude information of each AGV are determined, carrying out hierarchical planning on each AGV to obtain the path information of the AGV; the specific process is as follows:

the hierarchical planning refers to that when the AGV path planning is carried out, the complexity of an algorithm is reduced, meanwhile, the scheduling efficiency is guaranteed, the complex multi-objective planning problem is converted into two sub-problems, and the two algorithms are used for calculation respectively. The first layer is planned to be a plan among areas, wherein the current area of the AGV is set as an initial area, the area of the package to be delivered is set as a target area, and a shortest area path set for connecting the initial area and the target area is calculated by using an A-path planning algorithm. And planning among grids for the second time, and distributing grid resources in the same area by utilizing a time window arrangement algorithm, so that the AGV can quickly leave the current area according to the area path set on the basis of no collision.

Predicting a boundary frame of the AGV of the current frame by using path information, directly identifying an ApriTag code in the boundary frame, alarming and recording the missing AGV number by a system when the number of detected AGVs is insufficient, and detecting the AGV by using a tiny_yolv3 target detection network again; the specific process is as follows:

on the premise that all AGV positions and path information are known, detection is carried out without using a tiny_yov3 target detection network, the position of the current AGV is directly predicted according to a path planned by a dispatching system, and a boundary box is arranged on the basis of the position to frame and select an area where the AGV possibly exists, so that the instantaneity of a tracking algorithm is greatly improved.

Subsequently, the bounding box is also subjected to the second operation, and the april tag code at the top of the AGV is used for positioning. After all bounding boxes are detected, if the number of identified AGVs is consistent with the actual number of AGVs, the process is cycled for real-time tracking and scheduling of multiple AGVs. If the number of AGVs identified is not consistent with the initial number, then this means that there are AGVs that are not following the established trajectory, e.g., the AGV is malfunctioning or the AGV is hijacked. At this time, the tiny_yolv3 target detection network is reused to identify the AGVs in the global image, and any derailed AGVs can be relocated as long as the AGVs are still in the current warehouse environment, and are scheduled at the control center to be returned to the predetermined track.

In order to improve the robustness of the system, besides the fault tolerance of the april tag code, the system adopts a hierarchical planning algorithm, and the information of each AGV is registered in each area, so that if the hamming distance of the identified april tag code is larger than a set threshold value, the code has the possibility of being wrongly identified, the hamming distance between the code and other AGV top april tag codes in the area is calculated, and the result with the minimum hamming distance is selected as the ID of the current AGV, so that the probability of the false identification of the AGV is greatly reduced.

For example, when the AGVs are located in zone 1, which registers 4 AGVs numbered 1 (d 5d 628584), 2 (d 97f18b 49), 3 (dd 280910 e), and 4 (e 479e9c 98), respectively. In this region, the minimum hamming distance of 4 codewords is 14, the minimum hamming distance of the other bits is 16 and 17, and the error detection and correction capability is improved from original 5,5 to 6,6. And if a 7 bit error occurs. To be misrecognized as 3, the error must be at the intersection of 1 and 3, which is a very low probability, thus greatly increasing the recognition rate of the AGV.

3) The control center pertinently instructs each AGV according to the AGV information obtained by the tracking system and the scheduling system, and controls all AGVs to move according to the appointed path; the specific process is as follows:

and controlling a plurality of AGVs by utilizing the ROS robot operating system, and sending a control instruction to each AGV in real time by performing distributed communication based on the ROS system through WIFI. In order to enable the large-scale AGVs in the warehouse to stably run, the AGVs are arranged to run at a constant speed. Therefore, the linear speed transmitted to each AGV by the control center is a fixed value, and besides, the angular speed which is transmitted to each AGV is changed in real time according to the course angle and the coordinates of the AGV calculated by the multi-AGV tracking system by the control center and corrected on the basis of the linear speed. If the current position of the AGV is far away from the established route, the linear speed value is increased, and if the current position of the AGV exceeds the established route, the linear speed is reduced. In this way, the AGV is controlled not to deviate from the prescribed route.

This example has the following advantages:

1. compared with the use of triangle built-in numbers as the feature patterns, the april tag code adopted by the embodiment has stronger fault tolerance, and even partial pattern missing can be correctly identified. Meanwhile, the april tag code has stronger information storage capacity, and 587 or even more AGVs can be marked at the same time.

2. Compared with the method of directly using an image processing technology to perform multi-AGV positioning, the embodiment utilizes the tiny-YOLOv3 to perform AGV positioning, has higher instantaneity, and can adapt to different light sources and storage environments. Meanwhile, common obstacles in a warehouse, such as warehouse management personnel or falling goods, can be identified in real time through the tiny-YOLOv3, and collision of the AGV is avoided.

3. The output of a conventional multi-target tracking scheme is the type and coordinates of the target. The multi-AGV tracking algorithm provided in this embodiment has a function of outputting the target rotation angle in real time, which is necessary for AGV navigation.

4. Compared with the method for calculating the course angle of the AGV by using the two-dimensional code and the triangle as the characteristic patterns, the algorithm provided by the embodiment has higher calculation accuracy. By way of illustration, the most easily identifiable version of a two-dimensional code 1 (21 x 21 module) and a triangle with a built-in number are placed as a feature pattern on top of the AGV. The two-dimensional code determines the rotation angle through three positioning codes, and the triangular pattern calculates the rotation angle through extracting three angular points of the triangular pattern. And the same speed instruction is sent to the AGV, the top camera shoots a global image in the moving process of the AGV, and then the deviation values of the course angle of the AGV and the true course angle calculated by different algorithms in continuous 100 frames of images are compared, and the result is shown in figure 9. By comparison, the multi-AGV tracking method provided by the invention has higher accuracy in calculating the course angle of the AGV.

The above examples are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above examples, and any other changes, modifications, substitutions, combinations, and simplifications that do not depart from the spirit and principle of the present invention should be made in the equivalent manner, and the embodiments are included in the protection scope of the present invention.

Claims

1. The multi-warehouse AGV tracking method based on the global vision is characterized by comprising the following steps of:

2) The tracking system of the control center processes the global image, identifies a plurality of AGVs by using a tiny_yolv3 target detection network in a first frame, and determines the ID and the pose of each AGV according to an april tag code at the top of the AGVs; the method specifically comprises the following steps:

θ＝90°×r+(45°-∝) (2)

wherein alpha is an auxiliary angle for calculating the course angle of the AGV, and the value of the auxiliary angle is equal to the average value of included angles of four sides of the AprilTag code; r is 0-3, which means that the current identification is matched with the family code through r times of 90-degree rotation energy, k _i Calculating the weight of the angle for the ith edge least square method, wherein each edge is set to have the same weight;

2. The multi-warehouse AGV tracking method based on global vision according to claim 1, wherein in step 1), a top camera is placed at the top of the warehouse for monitoring and capturing images of the whole warehouse, and a control center of the warehouse system comprises a multi-AGV tracking system and a multi-AGV scheduling system for determining pose information and specific paths of each AGV.

3. The multi-warehouse AGV tracking method according to claim 1, wherein the step 3) includes:

the warehouse is divided into two layers for classification and storage, wherein the second layer is a sorting layer, and the first layer is a centralized storage layer; dividing a sorting area into n areas, wherein each area comprises an intersection and a corridor connected with the intersection, the areas are provided with a plurality of grids which are the most basic units for AGV scheduling, and the grids are divided into an initial area and a target area in a sorting layer according to different division of the areas; in the task execution process, each AGV can obtain an optimal path reaching a target area by utilizing a layering plan according to the assigned task, the AGV runs into a carried package input pipeline after reaching the target area according to the path, cargoes can enter a first layer for centralized storage, and the AGV can return to an idle starting area for next task circulation after the sorting task is finished.

4. The multi-warehouse AGV tracking method based on global vision according to claim 3 wherein the hierarchical planning means that when the AGV path planning is performed, the scheduling efficiency is ensured while the algorithm complexity is reduced, the complex multi-objective planning problem is converted into two sub-problems, and the two sub-problems are calculated by two algorithms respectively; the first layer is planned to be a plan among areas, an area where the AGV is currently located is set as an initial area, an area where the package needs to be delivered is set as a target area, and a shortest area path set for connecting the initial area and the target area is calculated by using an A-path planning algorithm; and planning among grids for the second time, and distributing grid resources in the same area by utilizing a time window arrangement algorithm, so that the AGV can quickly leave the current area according to the area path set on the basis of no collision.

5. The multi-warehouse AGV tracking method based on global vision according to claim 1, wherein the step 4) specifically comprises the following steps:

6. The multi-warehouse AGV tracking method based on global vision according to claim 1, wherein the step 5) specifically comprises the following steps: