A robot self-positioning accuracy evaluation method and system based on visual sensor

Abstract

The present invention discloses a method and system for evaluating robot self-positioning accuracy based on a visual sensor. The method includes the following steps: S1: collecting motion data of a mobile robot and recording it in ROS package format; S2: extracting the recorded data, fusing the RGB image and the depth image to obtain point cloud information at the current timestamp; S3: calculating the two-dimensional information entropy using each RGB image frame or converting the point cloud information obtained in step S2 into a global descriptor, finding and merging revisited sample pairs through a search algorithm to obtain a sample pair dataset; S4: obtaining the local displacement vector and point set error of the sample pair dataset by accessing the sample pair dataset; S5: collecting estimated trajectory data under the current recording environment; S6: offline registering the data, randomly selecting registration pairs, and repeatedly running the algorithm to calculate the sample error, which is the error of the robot trajectory. The present invention achieves robot self-positioning accuracy evaluation using a low-cost visual sensor.

Description

Technical Field

[0001] The present invention relates to the technical field of robot positioning, and more specifically, to a robot self-positioning accuracy evaluation method and system based on a visual sensor. Background Art

[0002] Robotic SLAM algorithms are increasingly being used in industries such as home care, autonomous driving, and warehousing and logistics. Evaluating the positioning accuracy of mobile robots with ease and low cost is crucial. SLAM algorithms are typically evaluated with the aid of benchmark datasets, artificial landmarks, and expensive equipment such as motion capture systems. However, dataset-based methods are environment-specific. The use of motion capture systems limits spatial coverage, while artificial landmarks pose challenges such as contamination and the need for large deployments. Compared to artificial landmarks, natural objects have natural landmarks. By repeatedly visiting the same location two or more times, high-precision local displacement metrics can be obtained based on all observations collected from these repeated visits. These local displacement metrics can then be used to evaluate the global error of a SLAM positioning algorithm's trajectory. This error can be calculated by using the parameters of a statistical model, enabling the robot to assess its self-localization accuracy during motion. While lidar sensors offer high point cloud accuracy, they are relatively large and expensive, and they suffer from specular reflection issues. Because vision sensors produce less accurate point clouds than lidar sensors, the positional uncertainty of the point clouds is high, making obtaining high-precision local displacement metrics challenging.

[0003] Prior art discloses a method for evaluating robot positioning accuracy based on monocular vision. This method places a calibration plate in the robot's motion environment. The robot observes the same calibration plate from different positions during motion, and multiple sampling is used to quantitatively compare different positioning algorithms in the same environment. Although this prior art also uses a visual sensor, it calculates local displacement vectors based on the robot's different poses under the calibration plate. However, the calibration plates need to be placed in large numbers throughout the environment and are easily contaminated. This makes it impossible to achieve natural robot movement in the environment and perform self-positioning accuracy assessment.

[0004] In order to solve the above problems, the present invention proposes a robot self-positioning accuracy evaluation method and system based on visual sensors from the perspective of natural landmarks. Summary of the Invention

[0005] In order to overcome the defects of the above-mentioned prior art robot positioning accuracy assessment method in terms of high cost and low reliability, the present invention provides a robot self-positioning accuracy assessment method and system based on visual sensors.

[0006] The primary purpose of the present invention is to solve the above technical problems, and the technical solutions of the present invention are as follows:

[0007] A first aspect of the present invention provides a method for evaluating robot self-positioning accuracy based on a visual sensor, comprising the following steps:

[0008] S1: Use visual sensors to collect motion data of the mobile robot and record the data in ROS package format;

[0009] S2: Extract the recorded data from the ROS package, fuse the RGB image and depth image with the same timestamp in the recorded data, and obtain the point cloud information at the current timestamp;

[0010] S3: Calculate the two-dimensional information entropy using the RGB image of each frame in the recorded data or convert the point cloud information obtained in step S2 into a global descriptor, find the revisited sample pairs through the search algorithm, and merge the sample pairs to obtain a sample pair dataset;

[0011] S4: Obtain the local displacement vector and point set error of the sample pair data set by accessing the sample pair data set;

[0012] S5: Collect estimated trajectory data in the current recording environment online or offline using the SLAM algorithm;

[0013] S6: Combine the local displacement vector and point set error of the sample pair data set with the estimated trajectory data to obtain an offline registration pair, randomly select the registration pair, and repeat the operation to calculate the error of the sample, which is the error of the robot trajectory.

[0014] Furthermore, the visual sensor in step S1 is set on a mobile robot, and the mobile robot moves in a random walk or fixed-point navigation manner.

[0015] Furthermore, the visual sensor includes an RGB-D camera, a monocular camera, and a binocular camera.

[0016] Furthermore, the point cloud information at each timestamp includes several frames of point cloud data. For each frame of point cloud data, the maximum point set of approximate features in multiple neighborhoods is extracted or directly downsampled to filter corner points and plane points respectively, and eliminate unstable points caused by motion distortion.

[0017] Furthermore, step S3 also includes finding revisited sample pairs through a loop closure detection method, and the loop closure detection method includes: an extraction method based on an ORB-SLAM2 local map and a loop closure detection algorithm based on deep learning.

[0018] Furthermore, the specific steps of obtaining the local displacement vector of the sample pair data set by accessing the sample pair data set are as follows:

[0019] Determine the feature extraction model, use each frame of point cloud data obtained in step S2 as input data, and perform preliminary feature extraction;

[0020] Perform nearest neighbor search on the point cloud information obtained from the preliminary feature extraction to find the rough registration point pairs;

[0021] The ICP series algorithm is used to further align the coarse registration point pairs. Based on the initial registration matrix, the optimization is performed to obtain the precise registration relationship between the two sets of point clouds, namely the precise registration matrix. The translation vector of the precise registration matrix is ​​the local displacement vector.

[0022] Calculate the average distance between the inner point pairs of the two groups of point clouds after precise matching. If it is greater than the preset threshold, match again or choose to abandon the acquisition;

[0023] Because the uncertainty of a point cannot be calculated for a single frame point cloud, for each point in a single frame point cloud, the entropy of the neighborhood of the point is calculated as s1; the other frame point cloud is converted to the frame point cloud through the registration matrix, and the entropy of the neighborhood of the point is calculated as s2. The uncertainty s of the point is s2-s1. At this time, the uncertainty of the point in the frame point cloud is the average value of the s values ​​of all points∈ pi .

[0024] Furthermore, the feature extraction model includes: FPFH algorithm model and 3dsmooth algorithm model.

[0025] Furthermore, the SLAM algorithms used in step S5 include: AMCL algorithm, Catagrapher algorithm, LOAM series algorithms, and ORB-SLAM series algorithms.

[0026] Furthermore, the specific calculation of the error in step S6 is:

[0027] Define any local translation vector as Vx, find the two poses of the trajectory of the positioning algorithm corresponding to the timestamp, define the translation vector between the two as Vy, form an offline registration pair (Vx, Vy), and the uncertainty of the points between the registration point clouds ∈ p , the calculation formula is:

[0028] -∈1+v y +∈2=∈ p1 +v x -∈ p2

[0029] |v x +∈ p | 2 =|v y +∈| 2

[0030] ∈1, ∈2 represent the errors of the two postures of the trajectory of the positioning algorithm, ∈ p1 Indicates the uncertain value of the point in the first frame of point cloud; ∈ p2 Indicates the uncertainty value of the point in the second frame point cloud, ∈=∈2-∈1 represents the difference between the two positioning errors, ∈ p =∈ p1 -∈ p2 , represents the difference in the uncertainty value of the points in the frame point cloud;

[0031] The estimation function of the GLO algorithm is defined as:

[0032] g σ (∈)=∑(v y +σ*z) 2

[0033] σ represents the mean value, and z represents the Gaussian distribution;

[0034] Loss function positioning l σ for:

[0035] I σ =(∑(v y +σ*z) 2 -∑(v y +∈ p ) 2 ) 2 .

[0036] A second aspect of the present invention provides a robot self-positioning accuracy assessment system based on a visual sensor, the system comprising: a memory and a processor, wherein the memory includes a robot self-positioning accuracy assessment method program based on a visual sensor, and when the robot self-positioning accuracy assessment method program based on a visual sensor is executed by the processor, the following steps are implemented:

[0037] S1: Use visual sensors to collect motion data of the mobile robot and record the data in ROS package format;

[0038] S2: Extract the recorded data from the ROS package, fuse the RGB image and depth image with the same timestamp in the recorded data, and obtain the point cloud information at the current timestamp;

[0039] S3: Calculate the two-dimensional information entropy using the RGB image of each frame in the recorded data or convert the point cloud information obtained in step S2 into a global descriptor, find the revisited sample pairs through the search algorithm, and merge the sample pairs to obtain a sample pair dataset;

[0040] S4: Obtain the local displacement vector and point set error of the sample pair data set by accessing the sample pair data set;

[0041] S5: Collect estimated trajectory data in the current recording environment online or offline using the SLAM algorithm;

[0042] S6: Combine the local displacement vector and point set error of the sample pair data set with the estimated trajectory data to obtain an offline registration pair, randomly select the registration pair, and repeat the operation to calculate the error of the sample, which is the error of the robot trajectory.

[0043] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0044] The present invention uses visual sensors to obtain the robot's motion data, and then fuses the data to obtain point cloud data. Through revisiting and registration, the local displacement vector and point set error are obtained. The trajectory is estimated and offline registration pairs are performed through the SLAM algorithm. The registration pairs are randomly selected to calculate the trajectory error, which reduces the evaluation cost and improves the robot's self-positioning accuracy and the reliability of the evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a robot self-positioning accuracy evaluation method based on a visual sensor according to the present invention.

[0046] Figure 2 This is a block diagram of a robot self-positioning accuracy evaluation system based on a visual sensor in the present invention. DETAILED DESCRIPTION

[0047] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0049] Example 1

[0050] like Figure 1 As shown, the first aspect of the present invention provides a robot self-positioning accuracy evaluation method based on a visual sensor, comprising the following steps:

[0051] S1: Use visual sensors to collect motion data of the mobile robot and record the data in ROS package format;

[0052] It should be noted that the visual sensor in the present invention is installed on a mobile robot, and the mobile robot adopts a random walk or fixed-point navigation mode of movement. More specifically, the visual sensor can be: an RGB-D camera, a monocular camera, or a binocular camera.

[0053] In a specific embodiment, the visual sensor set up by the robot needs to have at least one RGB-D camera, and the sensor is installed according to the algorithm to be evaluated. The present invention is not limited to the evaluation algorithm mentioned above and can be used to evaluate different types of SLAM algorithms.

[0054] The present invention preferably utilizes an RGB-D camera. In specific embodiments, only a single RGB-D camera can be used as the input sensor. This offers practical advantages in terms of cost feasibility and simplicity of robotic design. Full functionality is achieved without the need for ground truth trajectories and can achieve evaluation accuracy on the same order of magnitude as motion capture systems, significantly reducing the expensive purchase cost. Compared to artificial landmarks, this approach overcomes their limitations. Compared to lidar sensors, the uncertainty error calculation of the post-registration points compensates for the lack of visual sensor accuracy, reducing costs. This approach overcomes the difficulties traditional positioning accuracy methods face in requiring ground truth trajectories.

[0055] S2: Extract the recorded data from the ROS package, fuse the RGB image and depth image with the same timestamp in the recorded data, and obtain the point cloud information at the current timestamp;

[0056] It should be noted that the point cloud information at each timestamp includes several frames of point cloud data. For each frame of point cloud data, the maximum point set with approximate features within multiple neighborhoods is extracted or directly downsampled, and corner points and planar points are screened separately to remove unstable points caused by motion distortion. More specifically, for each frame of point cloud, the feature matrix of the point neighborhood area is calculated, and the matrix eigenvalues ​​are further extracted. Points with the minimum eigenvalue below a certain threshold are retained as surface points, and unevenly distributed and unstable points are removed.

[0057] S3: Calculate the two-dimensional information entropy using the RGB image of each frame in the recorded data or convert the point cloud information obtained in step S2 into a global descriptor, find the revisited sample pairs through the search algorithm, and merge the sample pairs to obtain a sample pair dataset;

[0058] It should be noted that, in the present invention, revisited sample pairs can also be found through a loop closure detection method, and the loop closure detection method includes: an extraction method based on an ORB-SLAM2 local map and a loop closure detection algorithm based on deep learning.

[0059] S4: Obtain the local displacement vector and point set error of the sample pair data set by accessing the sample pair data set;

[0060] It should be noted that the present invention obtains the relevant point cloud registration relationship by accessing the sample pair data set and using the point cloud features of the sample pair point cloud, and further uses the registration algorithm for further registration. The local displacement is the length of the translation vector. If the distance of the inner point is greater than a certain threshold, it is an incorrect registration and is registered again. If this process is repeated three times and still an incorrect registration, the sample pair data is abandoned. Because the uncertainty of the point cannot be calculated for a single frame point cloud, for each point in a single frame point cloud, the entropy of the neighborhood of the point is calculated to be s1; the other frame point cloud is converted to the frame point cloud through the registration matrix, and the entropy of the neighborhood of the point is calculated to be s2, and the uncertainty s of the point is s2-s1. At this time, the uncertainty of the point of the frame point cloud is the average value of the s value of all points∈ pi .

[0061] More specifically, the steps for calculating the local displacement vector of the sample to the dataset are:

[0062] Determine a feature extraction model, use each frame of point cloud data obtained in step S2 as input data, and perform preliminary feature extraction; the feature extraction model includes: FPFH algorithm model and 3dsmooth algorithm model.

[0063] Perform nearest neighbor search on the point cloud information obtained from the preliminary feature extraction to find the rough registration point pairs;

[0064] The ICP series algorithm is used to further align the coarse registration point pairs. Based on the initial registration matrix, the optimization is performed to obtain the precise registration relationship between the two sets of point clouds, namely the precise registration matrix. The translation vector of the precise registration matrix is ​​the local displacement vector.

[0065] Calculate the average distance between the inner point pairs of the two groups of point clouds after precise matching. If it is greater than the preset threshold, match again or choose to abandon the acquisition to improve the robustness and accuracy of the local translation vector;

[0066] For each point in a single frame point cloud, the entropy of the neighborhood of the point is calculated as s1; the point cloud of another frame is converted to the point cloud of this frame through the registration matrix, and the entropy of the neighborhood of the point is calculated as s2. The uncertainty s of the point is s2-s1. At this time, the uncertainty of the point of the frame point cloud is the average value of the s values ​​of all points∈ pi , that is, the point set error is obtained.

[0067] It should be noted that the present invention is not limited to the point cloud registration of sample pairs, and can also use multi-visual point cloud registration methods, such as repeatedly visiting multiple frames of point clouds and registering them with each other, followed by further global optimization.

[0068] S5: Collect estimated trajectory data in the current recording environment online or offline using the SLAM algorithm;

[0069] It should be noted that the SLAM algorithms of the present invention include: AMCL algorithm, Catagrapher algorithm, LOAM series algorithms, and ORB-SLAM series algorithms.

[0070] The present invention can be repeatedly used to evaluate different types of SLAM algorithms, and can effectively help test and select the best parameters suitable for a certain SLAM algorithm, effectively reducing the number of parameters required for calculation and the consumption of computing resources.

[0071] S6: Combine the local displacement vector and point set error of the sample pair data set with the estimated trajectory data to obtain an offline registration pair, randomly select the registration pair, and repeat the operation to calculate the error of the sample, which is the error of the robot trajectory.

[0072] The specific calculation of the error is:

[0073] Define any local translation vector as Vx, find the two poses of the trajectory of the positioning algorithm corresponding to the timestamp, define the translation vector between the two as Vy, form an offline registration pair (Vx, Vy), and the uncertainty of the points between the registration point clouds ∈ p , the calculation formula is:

[0074] -∈1+v y +∈2=∈ p1 +v x -∈ p2

[0075] |v x +∈ p | 2 =|v y +∈| 2

[0076] ∈1, ∈2 represent the errors of the two postures of the trajectory of the positioning algorithm, ∈ p1 Indicates the uncertain value of the point in the first frame of point cloud; ∈ p2 Indicates the uncertainty value of the point in the second frame point cloud, ∈=∈2-∈1 represents the difference between the two positioning errors, ∈ p =∈ p1 -∈ p2 , represents the difference in the uncertainty value of the points in the frame point cloud;

[0077] The estimation function of the GLO algorithm is defined as:

[0078] g σ (∈)=∑(v y +σ*z) 2

[0079] σ represents the mean value, and z represents the Gaussian distribution;

[0080] Loss function positioning l σ for:

[0081] I σ =(∑(v y +σ*z) 2 -∑(v y +∈ p ) 2 ) 2 .

[0082] It should be noted that since the covariance matrix of the neighborhood of a well-registered point is spherical, while the distribution of a poorly registered point cloud is ellipsoidal, defining the uncertainty of a point as a Gaussian distribution can effectively describe the registration error of the point.

[0083] Example 2

[0084] like Figure 2 As shown, the second aspect of the present invention provides a robot self-positioning accuracy assessment system based on a visual sensor, the system comprising: a memory and a processor, wherein the memory comprises a robot self-positioning accuracy assessment method program based on a visual sensor, and when the robot self-positioning accuracy assessment method program based on a visual sensor is executed by the processor, the following steps are implemented:

[0085] S1: Use visual sensors to collect motion data of the mobile robot and record the data in ROS package format;

[0086] S2: Extract the recorded data from the ROS package, fuse the RGB image and depth image with the same timestamp in the recorded data, and obtain the point cloud information at the current timestamp;

[0087] S3: Calculate the two-dimensional information entropy using the RGB image of each frame in the recorded data or convert the point cloud information obtained in step S2 into a global descriptor, find the revisited sample pairs through the search algorithm, and merge the sample pairs to obtain a sample pair dataset;

[0088] S4: Obtain the local displacement vector and point set error of the sample pair data set by accessing the sample pair data set;

[0089] S5: Collect estimated trajectory data in the current recording environment online or offline using the SLAM algorithm;

[0090] S6: Combine the local displacement vector and point set error of the sample pair data set with the estimated trajectory data to obtain an offline registration pair, randomly select the registration pair, and repeat the operation to calculate the error of the sample, which is the error of the robot trajectory.

[0091] Example 3

[0092] This embodiment illustrates the above method of the present invention through specific experiments.

[0093] Step 1: Install the RealSense d435i sensor on the mobile robot and move it in an unknown environment. Record the data and save it as a ROS package. The mobile robot needs to revisit the same location at least twice or more.

[0094] Step 2: Feature processing: For each frame of point cloud, the feature matrix of the point neighborhood is calculated, and the matrix eigenvalues ​​are further extracted. Points with the minimum eigenvalue less than a certain threshold are retained as surface points, and points with uneven distribution and instability are removed.

[0095] Step 3: Use the two-dimensional information entropy algorithm to describe each frame of point cloud data, construct a KDTree through the descriptor vector, search for ten areas that are repeatedly visited and meet certain conditions, and merge the first access timestamps of the same five areas that are repeatedly visited to obtain a set of sample pairs.

[0096] Step 4: For each sample pair, use the FPFH algorithm to extract features from the target point cloud and the source point cloud, and obtain a coarse registration pair through the nearest neighbor search; then, use the ICP algorithm to further align the maximum registration pair to obtain an accurate transformation matrix. Under this transformation matrix, if the average value of the distance between the inner point pairs is greater than a certain threshold, then re-registration is performed, and the sample pair is abandoned after repeating this process three times. Because the uncertainty of the point cannot be calculated for a single-frame point cloud, for each point in a single-frame point cloud, the entropy of the neighborhood of the point is calculated as s1; the other frame point cloud is converted to the frame point cloud through the registration matrix, and the entropy of the neighborhood of the point is calculated as s2, and the uncertainty s of the point is s2-s1. At this time, the uncertainty of the point of the frame point cloud is the average value of the s values ​​of all points∈ pi .

[0097] Step 5: Use the ROS package to run the ORB-SLAM2 algorithm offline and save the estimated trajectory of the algorithm.

[0098] Step 6: Define a local translation vector as Vx, find the two poses of the LOAM trajectory at the corresponding timestamp, and define the translation vector between them as Vy to form an offline registration pair (Vx, Vy). Repeat the above process until all offline registration pairs are collected. Randomly select 50,000 pairs from the set of all pairs and calculate them using the GLO algorithm. The average of the calculated values ​​is used to calculate the final error.

[0099] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A robot self-positioning accuracy evaluation method based on a visual sensor, characterized in that: The following steps are involved: S1: Use visual sensors to collect motion data of the mobile robot and record the data in ROS package format; S2: Extract the recorded data from the ROS package, fuse the RGB image and depth image with the same timestamp in the recorded data, and obtain the point cloud information at the current timestamp; S3: Calculate the two-dimensional information entropy using the RGB image of each frame in the recorded data or convert the point cloud information obtained in step S2 into a global descriptor, find the revisited sample pairs through the search algorithm, and merge the sample pairs to obtain a sample pair dataset; S4: Obtain the local displacement vector and point set error of the sample pair data set by accessing the sample pair data set; S5: Collect estimated trajectory data in the current recording environment online or offline using the SLAM algorithm; S6: Combine the local displacement vector and point set error of the sample pair data set with the estimated trajectory data to obtain an offline registration pair, randomly select the registration pair, and repeat the operation to calculate the error of the sample, which is the error of the robot trajectory.

2. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 1, characterized in that: In step S1, the visual sensor is set on a mobile robot, and the mobile robot moves in a random walk or fixed-point navigation manner.

3. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 2, wherein: The visual sensor includes an RGB-D camera, a monocular camera, and a binocular camera.

4. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 1, wherein: The point cloud information at each timestamp includes several frames of point cloud data. For each frame of point cloud data, the maximum point set with approximate features in multiple neighborhoods is extracted or directly downsampled to filter corner points and plane points respectively, and eliminate unstable points caused by motion distortion.

5. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 1, characterized in that: Step S3 also includes finding revisited sample pairs through a loop detection method, and the loop detection method includes: an extraction method based on an ORB-SLAM2 local map and a loop detection algorithm based on deep learning.

6. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 1, characterized in that: The specific steps for obtaining the local displacement vector of the sample pair data set by accessing the sample pair data set are: Determine the feature extraction model, use each frame of point cloud data obtained in step S2 as input data, and perform preliminary feature extraction; Perform nearest neighbor search on the point cloud information obtained from the preliminary feature extraction to find the rough registration point pairs; The ICP series algorithm is used to further align the coarse registration point pairs. Based on the initial registration matrix, the optimization is performed to obtain the precise registration relationship between the two sets of point clouds, namely the precise registration matrix. The translation vector of the precise registration matrix is ​​the local displacement vector. Calculate the average distance between the inner point pairs of the two groups of point clouds after precise matching. If it is greater than the preset threshold, match again or choose to abandon the acquisition; For each point in a single frame point cloud, the entropy of the neighborhood of the point is calculated as s1; the point cloud of another frame is converted to the point cloud of this frame through the registration matrix, and the entropy of the neighborhood of the point is calculated as s2. The uncertainty s of the point is s2-s1. At this time, the uncertainty of the point of the frame point cloud is the average value of the s values ​​of all points∈ pi , that is, the point set error is obtained.

7. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 6, characterized in that: The feature extraction models include: FPFH algorithm model and 3dsmooth algorithm model.

8. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 1, characterized in that: The SLAM algorithms used in step S5 include: AMCL algorithm, Catagrapher algorithm, LOAM series algorithm, and ORB-SLAM series algorithm.

9. The method for evaluating robot self-positioning accuracy based on a visual sensor according to claim 1, characterized in that: The specific calculation of the error in step S6 is: Define any local translation vector as Vx, find the two poses of the trajectory of the positioning algorithm corresponding to the timestamp, define the translation vector between the two as Vy, form an offline registration pair (Vx, Vy), and the uncertainty of the points between the registration point clouds ∈ p , the calculation formula is: -∈1+v y +∈2=∈p1+v x -∈ p2 |v x +∈ p | 2 =|v y +∈| 2 ∈1, ∈2 represent the errors of the two postures of the trajectory of the positioning algorithm, ∈ p1 Indicates the uncertain value of the point in the first frame of point cloud; ∈ p2 Indicates the uncertainty value of the point in the second frame point cloud, ∈=∈2-∈1 represents the difference between the two positioning errors, ∈ p =∈ p1 -∈ p2 , represents the difference in the uncertainty value of the points in the frame point cloud; The estimation function of the GLO algorithm is defined as: g σ (∈)=∑(v y +σ*z) 2 σ represents the mean value, and z represents the Gaussian distribution; Loss function positioning l σ for: l σ =(∑(ν y +σ*z) 2 -∑(n y +∈ p ) 2 ) 2 。 10. A robot self-positioning accuracy evaluation system based on visual sensors, characterized in that: The system includes: a memory and a processor, wherein the memory includes a robot self-positioning accuracy assessment method program based on a visual sensor, and when the robot self-positioning accuracy assessment method program based on a visual sensor is executed by the processor, the following steps are implemented: S1: Use visual sensors to collect motion data of the mobile robot and record the data in ROS package format; S2: Extract the recorded data from the ROS package, fuse the RGB image and depth image with the same timestamp in the recorded data, and obtain the point cloud information at the current timestamp; S3: Calculate the two-dimensional information entropy using the RGB image of each frame in the recorded data or convert the point cloud information obtained in step S2 into a global descriptor, find the revisited sample pairs through the search algorithm, and merge the sample pairs to obtain a sample pair dataset; S4: Obtain the local displacement vector and point set error of the sample pair data set by accessing the sample pair data set; S5: Collect estimated trajectory data in the current recording environment online or offline using the SLAM algorithm; S6: Combine the local displacement vector and point set error of the sample pair data set with the estimated trajectory data to obtain an offline registration pair, randomly select the registration pair, and repeat the operation to calculate the error of the sample, which is the error of the robot trajectory.

Images