WO2022257196A1

WO2022257196A1 - Autonomous mobile apparatus positioning method based on point cloud map dynamic loading

Info

Publication number: WO2022257196A1
Application number: PCT/CN2021/102443
Authority: WO
Inventors: 蒋涛; 蔡明希; 许林; 邹艳玲; 张林帅; 李晨
Original assignee: 成都信息工程大学
Priority date: 2021-06-09
Filing date: 2021-06-25
Publication date: 2022-12-15
Also published as: CN113375664A; LU502363B1; CN113375664B

Abstract

An autonomous mobile apparatus positioning method based on point cloud map dynamic loading. The method comprises: step one, calibration, involving: performing online calibration and alignment on an inertial measurement unit (IMU) and a lidar by means of constructing an error model of the IMU; step two, constructing a map, involving: setting a point cloud loading rule, a mobile platform carrying the calibrated lidar and the calibrated IMU to record bags, and performing map construction on the bags on the basis of a loam algorithm of downsampling; step three, performing modular partitioning processing on the constructed map, so as to obtain a plurality of block maps that can be loaded by a robot, wherein a loading area is reserved between adjacent block maps; and step four, performing map mapping. By means of the method, a positioning deviation caused by a human error during the movement and matching of a carrier can be eliminated, and an initial loading speed is greatly increased, thereby ensuring that a movement speed and a matching speed of an inspection robot during the movement process remain stable, and realizing the reliability of full-time positioning.

Description

Localization method of autonomous mobile device based on dynamic loading of point cloud map

technical field

The invention relates to the field of complex environment robot inspection. More specifically, the present invention relates to an autonomous mobile device positioning method based on dynamic loading of point cloud maps used in industrial factories, agricultural parks and structured environments with relatively obvious features.

Background technique

In industrial factories, agricultural parks, and structured environments with obvious characteristics, because the GPS signal is blocked, the positioning effect of a single sensor is poor, and the limitations are large, which cannot meet the sub-centimeter level precision positioning requirements. In order to solve such problems, the positioning method of multi-sensor fusion is often used. However, this method will also bring new problems, such as IMU signal offset, which has not been corrected by GPS, resulting in an increasing cumulative error; lidar has many useless repetitive points in such an environment with obvious characteristics, and traditional laser The point cloud map is too large and the point cloud is repeated, resulting in slow loading speed and matching speed that cannot meet the inspection requirements, resulting in distorted positioning and accidents.

Contents of the invention

An object of the present invention is to solve at least the above problems and/or disadvantages and to provide at least the advantages as will be described hereinafter.

In order to achieve these objects and other advantages according to the present invention, a method for positioning autonomous mobile devices based on dynamic loading of point cloud maps is provided, including:

Step 1, calibration, by constructing the error model of the inertial sensor IMU, online calibration and alignment of the IMU and Lidar;

Step 2: Construct the map, set the point cloud loading rules, record the bag package with the calibrated lidar and IMU on the mobile platform, and construct the map of the bag package based on the loam algorithm of downsampling;

Step 3: Modular partition processing is performed on the constructed map to obtain multiple block maps that can be loaded by the robot, and a loading area is reserved between adjacent block maps;

Step 4, map mapping, maps the modularized map with the downloaded GPS map to obtain a rasterized map for the robot to walk.

Preferably, in step 1, the online calibration and alignment include:

S10, establishing an IMU calibration error model;

S11, stimulating the error term of the error model;

S12, performing an observability analysis on the error term;

S13, using the extended Kalman filter algorithm to estimate the component parameters of the error term to obtain a corresponding calibration matrix;

S14. Jointly calibrate the Lidar and the IMU to obtain a compensation matrix that can be added to map construction.

Preferably, in S10, the establishment of the calibration error model is configured to include:

The gyro error model is established as follows:

In the formula, ε ⁿ =[ε _E ε _N ε _U ] ^T represents the component of the gyro error ε ⁿ in the navigation coordinate system, and refers to the errors of east, north and sky respectively;

is the transformation matrix from the carrier coordinate system to the navigation coordinate system;

is the gyro output, and its components represent the output of the gyro x-axis, y-axis, and z-axis; the gyro scale factor error matrix δK _G =diag[δK _Gx δK _Gy δK _Gz ], and its components represent the output of the gyro x-axis, y-axis, and z-axis Scale factor error; ε ^b ＝[ε _bx ε _by ε _bz ] ^T is the random constant value error of the gyroscope, and its components represent the constant value errors of the gyroscope’s x-axis, y-axis, and z-axis;

Establish the gyro installation error matrix as follows:

Among them, the components of the matrix represent the installation errors of the gyroscope's x-axis, y-axis, and z-axis;

Accelerometer-based scale factor error δK _A , installation error matrix δA, random constant error

The measurement error model of the accelerometer is constructed as follows:

The corresponding calibration error model is established by referring to the parameters in the measurement error model and the gyroscope error model.

Preferably, in S11, the error term excitation is to perform error term excitation on the calibration error model by selecting the motion trajectory setting and/or attitude change of the inertial sensor;

In S12, the implementation of observability analysis is to determine the observability based on the selected excitation method, obtain the energy density spectrum of the state variable through the optimal interpolation method, and construct the Kalman filter Kalman filter equation according to the described error. accomplish;

In S13, use the extended Kalman filter algorithm to analyze the model characteristics of the system equation and the measurement equation, construct the observed quantity and use the state vector as the output, and complete the estimation of the unknown state or internal parameters through time update and measurement update;

In S14, through the joint calibration of Lidar and IMU, the external parameters of Lidar are obtained, and the conversion R-T parameters between Lidar and IMU are obtained through the conversion of IMU internal parameters obtained in S13, and the parameters obtained through IMU’s own error calibration matrix and joint calibration Matrix, inverting the parameter matrix to obtain a compensation matrix that can be added to the map construction.

Preferably, in step 2, the point cloud loading rule is configured to include:

Based on the mapping principle of Loam, all the cubes in the grid are connected into a large point cloud in a circular manner when sliding the window, and then a complete map is released through the topic;

Among them, in the loop, by adding a constant with an initial value of 0 to determine whether it is less than the current number of point clouds, if not, enter the loop to accumulate the laser cloud surround, and the initial constant is also increased by one bit after each loop; When the constant value is not less than the current number of point clouds, jump out of the loop;

The downsampling is configured to include:

In the Loam algorithm, the size of the filter is used to reduce the volume of the point cloud file, reduce the dimension of the feature and retain the effective information;

Wherein, the way of filtering the size is to save the point cloud scanned in the first frame in the map, and then judge whether the overlap rate between the point and the point in the previous frame reaches 50% or more after each frame is added to the map , delete if present, and keep if absent.

Preferably, in step 3, the modularized partition processing of the map is configured to include:

S30, determine the number of modules for roughly dividing the map;

S31. Determine the size of the buffer between modules;

S32. Carry out modular division of the map according to predetermined division rules, so that the difference between the number of point clouds in each module and the preset value does not exceed 10%, and the reserved area between modules meets the size of the buffer.

Preferably, in S30, the division of the number of regions of the map is based on the number of all point clouds contained in the map, and the relationship between the original loading time and the number of divided modules is calculated to obtain:

In S31, based on the driving speed of the inspection robot and the time required for each point cloud module to be loaded on the industrial computer, the length of the area loaded by the next modular map is set between every two adjacent modules as a buffer zone;

In S32, the division rules include:

S320, dividing the map into regions based on the number of modules obtained in S30, and setting each corner in the map as an independent module without further division;

S321. Select a plurality of modules as starting points, so as to construct a buffer zone between adjacent modules in a diffusion manner.

Preferably, the construction method of the buffer is configured to include:

S3210, take each corner as the starting point, set the full loading time as S, then in the map with the number of corners larger than S+1, select several corners with smaller relative positions as the starting point; and for the number of corners less than S+1 1 map, after selecting all the corners, carry out quantity judgment to determine the starting point;

S3211, spread outwards at the same time for each selected starting point, set the number of point clouds on the map as N, then the number of point clouds selected in the diffusion frame reaches

When , the diffusion limit is triggered and the frame selection is stopped;

S3212, compare the number of modules obtained by diffusion with the target number, and if the difference is 1, select a new starting point between the two farthest modules, perform secondary diffusion, and frame the selected point cloud in the secondary diffusion quantity reached

or when the number of point clouds is not reached, but the reserved distance between the diffusion radius and any diffusion radius of two adjacent modules is equal to 3 meters, the trigger stops;

Wherein, the position of the new starting point is

And l is the distance between the starting points of the two farthest modules, r ₁ and r ₂ are the diffusion radii of the two modules respectively;

S3213, integrate all the circles into rectangles according to the radius, and use the K-Means clustering algorithm and the method of probability greed to retain the quantitative point cloud in the module;

S3214, calculate the distance between the boundaries of each adjacent module, if the result is greater than 3 meters, then expand, if the expansion reaches 20% and cannot meet the requirements, perform quantitative offset, so that the interval distance meets the requirements;

If the result is less than 3 meters, the corresponding point clouds of the two modules are forced to be averaged to leave enough buffer area.

Preferably, in Step 4, the rasterized map construction method is configured to include:

S41. Download the 22-level Gaode map, extract the value of the overlapped part of the downloaded Gaode map and this map as a high-resolution grid map, and calculate the respective offsets by transforming the map resolution and the corresponding position of the map amount, and save it to the yaml file;

S42. Cut out the part corresponding to the point cloud map from the grid map, and map it into the point cloud map according to the conversion relationship in the yaml file.

Preferably, it also includes:

Step five, during the inspection operation of the robot, when it is driving on any module, the industrial computer loads the current module and the two buffers connected to it;

When the robot travels to any buffer zone, it judges whether the two module maps connected to it have been loaded, if not, then loads the map, if so, stops the judgment, and when the buffer zone finishes driving and enters the next module area, Delete the previous module area and its unilateral buffer, load the next buffer, maintain the matching speed of the system, and repeat until all maps are traversed.

The present invention at least includes the following beneficial effects: First, the present invention needs to calibrate the laser radar and IMU before establishing a high-precision point cloud map, and unify the coordinate system to ensure the accuracy of later positioning;

Second, when the present invention uses a high-precision algorithm to build a map, it eliminates repeated points through down-sampling, reduces the volume of the map file, and ensures the loading speed;

Third, the present invention divides the point cloud map into regions according to constraints such as the driving speed of the inspection robot and the density of the point cloud map, sets reasonable loading rules, and improves the point cloud matching speed.

Other advantages, objectives and features of the present invention will partly be embodied through the following descriptions, and partly will be understood by those skilled in the art through the study and practice of the present invention.

Description of drawings

Figure 1 shows the error types of INS;

Fig. 2 is the basic steps of online calibration of the present invention;

Fig. 3 is the processing flowchart of the present invention adopting attitude change setting to stimulate;

Fig. 4 is the map structure constructed by the original loam algorithm algorithm;

Fig. 5 is the map structure constructed after the loam algorithm processing of the present invention adopts down-sampling;

Fig. 6 is the map after the present invention adopts modular processing;

Fig. 7 is the corresponding grid map obtained after the mapping of the present invention;

Among them, Figures 4-7 are the inversion processing of the actual renderings.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings, so that those skilled in the art can implement it with reference to the description.

The invention proposes a positioning strategy based on point cloud map modular loading and fast matching. Before building a high-precision point cloud map, it is necessary to calibrate the lidar and IMU and unify the coordinate system; secondly, when using a high-precision algorithm to build a map, the repeated points are eliminated through downsampling to reduce the size of the map file; then, The standard GPS point coordinates are mapped to the point cloud map; finally, the point cloud map is divided into regions according to the constraints such as the driving speed of the inspection robot and the density of the point cloud map, and reasonable loading rules are set.

The positioning strategy based on point cloud map modular loading and fast matching proposed by the present invention includes:

1. Calibration

The inertial navigation system INSNS system is to use the inertial measurement device (such as accelerometer and gyroscope, etc.) installed on the carrier to sense the movement of the carrier, and output the attitude and position information of the carrier. The INS system is completely autonomous, highly confidential, and flexible, and has multi-function parameter output, but there is a problem that errors accumulate rapidly with time, and navigation accuracy diverges with time. Calibration error model).

Analyze and study the errors of inertial sensors, including systematic errors and random errors, establish a mathematical model of sensor errors, use this model to correct the output original data, and improve the accuracy of the original data. The error types output by INS are shown in Figure 1 The system error is the error that comes with the sensor, and its parameters can be obtained from the business trip file, which can be used directly when building the model, while the random error is the error caused by installation or other external factors, and its parameters need to be estimated , to complete the calibration or registration.

As shown in Figure 2, the online calibration and alignment of the INS, the specific steps include:

The gyro error model is established as follows:

The gyro installation error matrix is:

The component represents the installation error of the gyroscope's x-axis, y-axis, and z-axis. According to the above definition, a complete gyro error model can be constructed.

Combining the scale factor error δK _A of the accelerometer, the installation error matrix δA, the random constant error

The measurement error model of the accelerometer is constructed as follows:

In the formula, its components refer to parameters similar to the gyroscope error model. By establishing a calibration error model and selecting appropriate excitation and estimation methods, the calibration of the above error items can be completed.

Error item incentives, that is, through a specific method, the system error caused by a certain error is reasonably amplified, so as to facilitate the subsequent operation process. Appropriate incentive methods can improve the observability of certain error parameters, and the excellence of the incentive effect greatly affects the calibration effect. The selection of the excitation mode includes motion trajectory setting, attitude change setting and so on. The method of attitude change is preliminarily drawn up for error item excitation, and the steps of adopting attitude change setting as the excitation method are shown in Figure 3.

The observability is closely related to the excitation of the error term. Generally, the observability is used to define whether the error term of the SINS is excited in the maneuvering mode. By setting a reasonable trajectory, using observability analysis to determine its degree of observability, the energy density spectrum of the state variable is obtained through the optimal interpolation method, and the Kalman filter equation is constructed according to the described error, so as to realize the observability analysis.

For parameter estimation, the Kalman filter method is usually used to analyze the model characteristics of the system equation and the measurement equation, construct a reasonable observation quantity and use the state vector as an output, and complete the estimation of the unknown state or parameter through time update and measurement update. Although Kalman filtering has a wide range of applications, it mainly solves the problems of linear systems. For nonlinear filtering problems, in-depth improvement research is needed. This topic intends to use Extended Kalman Filter (EKF), considering the balance The error term simplifies tedious calculations. After the completion, fix the installation, use the lidar and IMU for joint calibration, get the external parameters of the Lidar, and convert the R-T parameters between the Lidar and the IMU through the conversion of the IMU internal parameters obtained in the previous step.

The compensation matrix is obtained through the inverse operation of the IMU's own error calibration matrix and the parameter matrix obtained by joint calibration, and the matrix is added to the map construction. So far, the preparation for map construction has been completed.

2. Map construction

The traditional loam algorithm can establish a relatively complete point cloud map, but there are problems such as too small carrier loading area and too many repeated points. First, the algorithm must be improved.

The first step is to set reasonable point cloud loading rules. The mapping principle of Loam is to save the point cloud map accumulated around the robot through the sliding window method. At the beginning, the initial position of the robot initializes a 3D mesh, and each cube in this mesh is represented by a point cloud. For each incoming scan, the mapping operation only updates a small cube area around the current position, and the updated area is connected to a large point cloud, which is finally down-sampled and published through the topic. As the robot moves freely in the grid, once it approaches any boundary of the grid, the grid moves in the corresponding opposite direction, thus losing the map cubes accumulated on the other side of the grid. The size of the grid is specified by the constructor LaserMapping in the Loam algorithm. The original regulation is an area of 50*50*50 meters. If the area is expanded blindly, the entire algorithm will consume a large amount of memory, and even crash directly in severe cases. So you can set up a loop to connect all the cubes into a big point cloud instead of the current update area, so as to publish the complete map through the topic. By adding a constant with an initial value of 0, use it to judge whether it is less than the current number of point clouds. If not, enter a loop to accumulate the laser cloud surround. After a loop, the initial constant also increases by one bit. In this loop, if the constant is not less than the current number of point clouds, then jump out of the loop. In this way, the grid area is increased and memory consumption is avoided.

The second step is downsampling. Although the above operations avoid memory dissipation to a certain extent, they still occupy a large amount of memory due to the disorder of the point cloud and the exponential increase in the number. In order to alleviate the loading pressure, the point cloud is down-sampled. By using an appropriate filter size to reduce the size of the point cloud file, reduce the dimension of the feature and retain effective information, avoid over-fitting to a certain extent, and facilitate transmission and loading. The filter size is designed as the following rules: the map saves the point cloud scanned in the first frame, and each subsequent frame is judged after being added to the map, whether the point overlap rate with the point in the previous frame reaches 50% or more, and if so, then Delete, or keep if none. Usually after downsampling, the volume can be reduced to less than half of the original, but the features will not be reduced. At the same time, due to the characteristics of the matching algorithm NDT used in the later stage, the final matching and positioning effect will not be affected after the down-sampling operation. During the process of building the map, the point cloud file is continuously read, then downsampled, and finally saved to a new file.

The third step is to use the mobile platform to carry the calibrated lidar and IMU to move at a constant speed in the structured road, wake up the radar and IMU, record the corresponding bag package, and end the recording after the path forms a loop.

The fourth step is to use the loam algorithm with downsampling to construct the map of the bag package in the previous step.

As shown in Figure 4, when the downsampling operation is not performed, the bag data is directly used for map construction. It can be seen that the map only retains about a quarter of the size of the original image, and the point cloud is messy, and the key areas The overlap is severe.

As shown in Figure 5, the modified algorithm was used to construct the map, the map was successfully loaded and the loop road was clearly displayed, the point cloud size of each area was uniform, the degree of sparseness was uniform, and the memory occupied was only one-third of that of the map in Figure 4. One, the loading speed has been significantly improved.

3. Map modular processing

Partition the created map. The first step is to determine the approximate number of areas to be divided. Count the number of all point clouds contained in this map, and obtain the relationship between its original loading time and the number of divided regions. For example, the number of point clouds in a map is N, and the average full loading time on the industrial computer is about S seconds. In order to ensure that the ideal loading speed of each module on the industrial computer does not exceed 1 second, that is, the minimum map needs to be divided into S+1 modules are dynamically loaded.

The second step is to determine the size of the buffer area. A loading area must be reserved between every two adjacent modules as a buffer area. Each module is connected to two buffer areas. When the robot reaches one of the buffer areas, it immediately judges whether the maps of the two modules connected to it have been loaded. If not, load the map. If so, stop judging. All in all, the buffer area is used to ensure that the module map is loaded smoothly and completely. In order to ensure that the inspection robot can complete map matching without distortion in the speed-limited areas such as campuses and factories, the maximum speed limit is 5km/h, which is about 1.5m/s after conversion. The ideal loading time for each module is 1 second. In order to ensure that the next module has been loaded before the robot drives out of the loading area, the reserved time is twice the loading time, which is 2 seconds. Therefore, the cache The span of the area is 1.5*2=3 meters.

The third step is to divide the map. Pick each starting point. In order to ensure the safety of the inspection robot during driving, each corner is an independent module and will not be divided separately. Therefore, the starting point can be selected at a special corner. For a map with a number of corners greater than S+1, several corners with smaller relative positions are selected; for a map with a number of corners less than S+1, the quantity judgment is performed after all corners are selected. Taking Figure 5 as an example, the specific operation is as follows: select four corner points as the starting point, and start to spread outward at the same time, when the number of selected point clouds reaches

When , the diffusion limit is triggered and the frame selection is stopped. At this time, the difference between the obtained module and the target number is 1, and a new starting point is selected between the two farthest modules, and the position of the starting point is selected as

Among them, l is the distance between the starting points of the two farthest modules, and r1 and r2 are the diffusion radii of the two modules respectively. There are two diffusion stop conditions for the current module, one is when the number of selected point clouds reaches

The second is that the number of point clouds has not reached, but the reserved distance between its diffusion radius and any diffusion radius of two adjacent modules is equal to 3 meters, and the trigger stops. Either of the two conditions is satisfied. Then, integrate all the circles into rectangles by radius. At this point, it is optimized using the improved K-Means clustering algorithm. K-Means requires that after the central point is quantitatively selected, the distance from each data point to the central point is calculated, and the data point is classified into the category which is closest to the central point. Calculate the center point of each class as the new center point. Repeat the above steps until the center of each class does not change much after each iteration. Therefore, after the center point has been selected and the preliminary clustering has been completed, the iterations have been carried out, and the retained point cloud is selected in a probabilistic greedy manner. The closer to the center point, the greater the probability of being retained, and the farther away from the center point, the smaller the probability of being retained. The probabilistic greedy operation not only retains high-quality points, but also avoids falling into local optimum. The specific operation of probabilistic greed is: exhaustively enumerate the distance between the center point C ₁ and other points and calculate the sum. The probability of each point being retained is:

Use this operation to preserve the quantitative point cloud in the module. Finally, calculate the distance between two pairs of module boundaries. If it is greater than 3 meters, then expand it. If the expansion reaches 20% and cannot meet the requirements, perform quantitative offset to make the separation distance meet the requirements. If it is less than 3 meters, the corresponding point clouds of the two modules are forced to be deleted on average to leave a sufficient buffer area. Finally, ensure that the difference between the number of point clouds of each module and the preset value does not exceed 10%, and each reserved area meets the requirement of 3-meter intervals, and then the map construction can be completed.

In actual operation, the construction of the map may further include: dividing a rough area in the first step. The built map is divided based on the sparseness of the point cloud. In order to ensure the safety of the inspection robot during driving, each corner is an independent module and will not be divided separately. Ensure that the number of point clouds in each module differs by no more than 10%, in order to pursue the uniformity of each time when the module is loaded. For example, this map is evenly divided into five areas according to the point cloud sparseness, which are distinguished by different colors. The second step is to calculate the driving distance of the inspection robot. A loading area must be reserved between every two adjacent modules, which is the black area in Figure 6. In speed-limited areas such as campuses and factories, the driving speed of the robot is usually 5km/h, which is about 1.5m/s after conversion. It takes about 1 second to load each point cloud module on the industrial computer. Therefore, to ensure safe driving, a loading time of 2 seconds is reserved to ensure that the loading of the next module has been completed before the robot drives out of the loading area. Therefore, the length of the loading area is set to 3m in the map. When the robot travels to the loading area, the system starts to judge whether to load two areas adjacent to the current loading area. If not, it starts loading the next area module. If so, it stops. Judgment, keep running. When the loading area is completed, the previous module is deleted to maintain the matching speed of the system, and so on.

4. Mapping

Download the 22-level Gaode map, extract the part of the downloaded Gaode map that overlaps with this map, and use the value as a high-resolution grid map. Select a special point as the origin, which is generally selected as the lower left corner. Through the map resolution and the corresponding position transformation of the map, the offsets are calculated and saved in the yaml file. The finally obtained grid map is shown in Figure 7.

Then cut out the part corresponding to the point cloud map from the completed grid map, and map it to the point cloud map according to the conversion relationship in the yaml file.

So far, a complete cut modular laser point cloud map containing GPS signals has been created.

5. Robot inspection operation based on map

The constructed map is shown in Figure 6. When the robot travels on any module, the industrial computer only loads the current module and the two buffers connected to it. When it travels to any buffer, it immediately judges whether the maps of the two modules connected to it have been loaded. , then load the map, and if so, stop judging. When the robot enters the next area, delete the previous area and its unilateral buffer, load the next buffer, and so on, until all modules are traversed. Adopting the positioning strategy or method based on point cloud map modular loading and fast matching of the present invention has the following advantages:

1. LiDAR and IMU are calibrated to eliminate human error, resulting in reduced map deviation and more complete wall construction, which eliminates the positioning deviation caused by human error when the carrier moves and matches.

2. The map is down-sampled and modularized, resulting in a complete map construction, increased features, reduced interference, reduced memory, and greatly improved the initial loading speed. A map that originally took 5 seconds to load can now be loaded in only 1 second, and the accuracy has not been reduced, but has also been raised to a higher level. Ensure that the moving speed and matching speed of the inspection robot remain stable during the movement process, and realize the reliability of all-time positioning.

The above solution is only an illustration of a preferred example, but not limited thereto. When implementing the present invention, appropriate replacements and/or modifications can be made according to user requirements.

The number of devices and processing scales described here are used to simplify the description of the present invention. Applications, modifications and variations to the present invention will be apparent to those skilled in the art.

Although embodiments of the present invention have been disclosed above, it is not limited to the applications set forth in the specification and examples. It can be fully applied to various fields suitable for the present invention. Additional modifications can readily be made by those skilled in the art. Therefore, the invention should not be limited to the specific details and examples shown and described herein, without departing from the general concept defined by the claims and their equivalents.

Claims

A positioning method for autonomous mobile devices based on dynamic loading of point cloud maps, characterized in that it includes:

Step 1, calibration, by constructing the error model of the inertial sensor IMU, online calibration and alignment of the IMU and Lidar;

Step 2: Construct the map, set the point cloud loading rules, record the bag package with the calibrated lidar and IMU on the mobile platform, and construct the map of the bag package based on the loam algorithm of downsampling;

Step 3: Modular partition processing is performed on the constructed map to obtain multiple block maps that can be loaded by the robot, and a loading area is reserved between adjacent block maps;

Step 4, map mapping, map the modularized map with the downloaded GPS map to obtain a rasterized map for the robot to walk.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 1, wherein in step 1, the online calibration and alignment include:

S10, establishing an IMU calibration error model;

S11, stimulating the error term of the error model;

S12, performing an observability analysis on the error term;

S13, using the extended Kalman filter algorithm to estimate the component parameters of the error term to obtain a corresponding calibration matrix;

S14. Jointly calibrate the Lidar and the IMU to obtain a compensation matrix that can be added to map construction.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 1, wherein in S10, the establishment of the calibration error model is configured to include:

The gyro error model is established as follows:

In the formula, ε n =[ε E ε N ε U ] T represents the component of the gyro error ε n in the navigation coordinate system, and refers to the errors of east, north and sky respectively;
is the transformation matrix from the carrier coordinate system to the navigation coordinate system;
is the gyro output, and its components represent the output of the gyro x-axis, y-axis, and z-axis; the gyro scale factor error matrix δK G =diag[δK Gx δK Gy δK Gz ], and its components represent the output of the gyro x-axis, y-axis, and z-axis Scale factor error; ε b ＝[ε bx ε by ε bz ] T is the random constant value error of the gyroscope, and its components represent the constant value errors of the gyroscope’s x-axis, y-axis, and z-axis;

Establish the gyro installation error matrix as follows:

Among them, the components of the matrix represent the installation errors of the gyroscope's x-axis, y-axis, and z-axis;

Accelerometer-based scale factor error δK A , installation error matrix δA, random constant error
The measurement error model of the accelerometer is constructed as follows:

The corresponding calibration error model is established by referring to the parameters in the measurement error model and the gyroscope error model.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 2, characterized in that in S11, the error item is excited by selecting the motion track setting and/or attitude change of the inertial sensor, and the calibration The error model is used to stimulate the error term;

In S12, the implementation of observability analysis is to determine the observability based on the selected excitation method, obtain the energy density spectrum of the state variable through the optimal interpolation method, and construct the Kalman filter Kalman filter equation according to the described error. accomplish;

In S13, use the extended Kalman filter algorithm to analyze the model characteristics of the system equation and the measurement equation, construct the observed quantity and use the state vector as the output, and complete the estimation of the unknown state or internal parameters through time update and measurement update;

In S14, through the joint calibration of Lidar and IMU, the external parameters of Lidar are obtained, and the conversion R-T parameters between Lidar and IMU are obtained through the conversion of IMU internal parameters obtained in S13, and the parameters obtained through IMU’s own error calibration matrix and joint calibration Matrix, inverting the parameter matrix to obtain a compensation matrix that can be added to the map construction.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 1, wherein in step 2, the point cloud loading rule is configured to include:

Based on the mapping principle of Loam, all the cubes in the grid are connected into a large point cloud in a circular manner when sliding the window, and then a complete map is released through the topic;

Among them, in the loop, by adding a constant with an initial value of 0 to determine whether it is less than the current number of point clouds, if not, enter the loop to accumulate the laser cloud surround, and the initial constant is also increased by one bit after each loop; When the constant value is not less than the current number of point clouds, jump out of the loop;

The downsampling is configured to include:

In the Loam algorithm, the size of the filter is used to reduce the volume of the point cloud file, reduce the dimension of the feature and retain the effective information;

Wherein, the way of filtering the size is to save the point cloud scanned in the first frame in the map, and then judge whether the overlap rate between the point and the point in the previous frame reaches 50% or more after each frame is added to the map , delete if present, and keep if absent.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 1, characterized in that, in step 3, the modular partition processing of the map is configured to include:

S30, determine the number of modules for roughly dividing the map;

S31. Determine the size of the buffer between modules;

S32. Carry out modular division of the map according to a predetermined division rule, so that the difference between the number of point clouds in each module and the preset value does not exceed 10%, and the reserved area between modules satisfies the size of the buffer zone.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 6, wherein in S30, the area number division of the map is based on the number of all point clouds contained in the map, and the calculation of the original loading time and Divide the relationship on the number of modules to obtain:

In S31, based on the driving speed of the inspection robot and the time required for each point cloud module to be loaded on the industrial computer, the length of the area loaded by the next modular map is set between every two adjacent modules as a buffer zone;

In S32, the division rules include:

S320, dividing the map into regions based on the number of modules obtained in S30, and setting each corner in the map as an independent module without further division;

S321. Select a plurality of modules as starting points, so as to construct a buffer zone between adjacent modules in a diffusion manner.
The autonomous mobile device positioning method based on dynamic loading of point cloud maps according to claim 7, wherein the construction method of the buffer zone is configured to include:

S3210, take each corner as the starting point, set the full loading time as S, then in the map with the number of corners larger than S+1, select several corners with smaller relative positions as the starting point; and for the number of corners less than S+1 1 map, after selecting all the corners, carry out quantity judgment to determine the starting point;

S3211, spread outwards at the same time for each selected starting point, set the number of point clouds on the map as N, then the number of point clouds selected in the diffusion frame reaches
When , the diffusion limit is triggered and the frame selection is stopped;

S3212, compare the number of modules obtained by diffusion with the target number, and if the difference is 1, select a new starting point between the two farthest modules, perform secondary diffusion, and frame the selected point cloud in the secondary diffusion quantity reached
or when the number of point clouds is not reached, but the reserved distance between the diffusion radius and any diffusion radius of two adjacent modules is equal to 3 meters, the trigger stops;

Wherein, the position of the new starting point is
and
is the distance between the starting points of the two farthest modules, r 1 and r 2 are the diffusion radii of the two modules respectively;

S3213, integrate all the circles into rectangles according to the radius, and use the K-Means clustering algorithm and the method of probability greed to retain the quantitative point cloud in the module;

S3214, calculate the distance between the boundaries of each adjacent module, if the result is greater than 3 meters, then expand, if the expansion reaches 20% and cannot meet the requirements, perform quantitative offset, so that the interval distance meets the requirements;

If the result is less than 3 meters, the corresponding point clouds of the two modules are forced to be averaged to leave enough buffer area.
The autonomous mobile device positioning method based on point cloud map dynamic loading according to claim 1, characterized in that, in step 4, the construction method of the rasterized map is configured to include:

S41. Download the 22-level Gaode map, extract the value of the overlapped part of the downloaded Gaode map and this map as a high-resolution grid map, and calculate the respective offsets by transforming the map resolution and the corresponding position of the map amount, and save it to the yaml file;

S42. Cut out the part corresponding to the point cloud map from the grid map, and map it into the point cloud map according to the conversion relationship in the yaml file.
The autonomous mobile device location method based on point cloud map dynamic loading as claimed in claim 1, is characterized in that, also comprises:

Step five, during the inspection operation of the robot, when it is driving on any module, the industrial computer loads the current module and the two buffers connected to it;

When the robot travels to any buffer zone, it judges whether the two module maps connected to it have been loaded, if not, then loads the map, if so, stops the judgment, and when the buffer zone finishes driving and enters the next module area, Delete the previous module area and its unilateral buffer, load the next buffer, maintain the matching speed of the system, and repeat until all maps are traversed.