Planar mobile robot visual positioning method and device without three-dimensional map model

Abstract

The application discloses a planar mobile robot vision positioning method and device without a three-dimensional map model. A query image and a reference image are extracted and paired to obtain 2D-2D matching feature points; a prior kinematic constraint of a planar mobile robot is used to simplify epipolar constraint; based on the simplified epipolar constraint, the 2D-2D matching feature points are used to obtain an absolute pose by using a 2P1P minimum solution method; and the absolute pose passing through translational consistency check is subjected to 6-DOF nonlinear optimization according to the number of inliers. The 2P1P minimum solution method and a multiple check scheme are combined to establish a robust three-dimensional model-free vision positioning framework, and the sensitivity of a positioning system to environmental changes and noise is reduced. The application does not depend on a three-dimensional model established in advance, but realizes positioning through multiple checks and analysis of images and scenes, thereby reducing dependence on the three-dimensional model and enhancing adaptability to complex scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of robot positioning, and particularly relates to a planar mobile robot visual positioning method and device without a three-dimensional map model. BACKGROUND

[0002] Now is an era of rapid development of robot technology, and the functions that can be realized by mobile robots are constantly enriched, and gradually integrated into various aspects of our life. At present, mobile robots are mainly applied in intelligent home, logistics transportation, inspection and unmanned driving fields. In these fields, whether an autonomous mobile robot can efficiently and stably realize high-precision synchronous positioning and map construction has become a very important and key evaluation standard.

[0003] Mobile robot positioning is mainly divided into laser positioning and visual positioning, although the laser positioning technology is more mature, but the cost of laser sensor is much higher than that of camera. In terms of application scene, laser is mainly applied in indoor, and when there are multiple laser devices, there is interference, while camera can work in indoor and outdoor scenes, and cameras do not interfere with each other. Camera can provide intuitive environmental information and has high resolution, which is convenient for feature extraction and description.

[0004] At present, the most accurate visual positioning method is based on three-dimensional information, and the visual positioning method based on three-dimensional scene generally first obtains 2D-3D feature matching between the query image and the pre-constructed three-dimensional scene, and then performs accurate positioning based on absolute pose estimation algorithm, but all are for specific scenes, that is, when facing a new scene, additional time and computing resources are needed to construct a high-precision three-dimensional map. Therefore, the goal of these methods is to realize accurate and robust visual positioning under the condition of pre-establishing a three-dimensional scene model.

[0005] Another more flexible visual positioning method is to directly position based on a 2D scene database, which is a set of reference images with associated camera poses. The most similar database image is retrieved, and the relevant pose in the database is used to approximate the pose of the query image. But this method is limited by the sampling interval when constructing the database, and it is difficult to meet the requirements of tasks with high positioning accuracy requirements. In order to further improve the positioning accuracy, 2D-2D feature matching is used to estimate the pose of the query image.

[0006] However, the existing 2D-2D pose estimation methods have two problems. First, these methods are mostly based on essential matrix to estimate the relative pose between two images, and the essential matrix provides a rotation matrix and a relative translation direction, but not an absolute translation scale. Therefore, they cannot obtain the absolute pose of the camera relative to the map, which is essential for visual localization. Second, most of the relative pose estimation problems are described as a 6-DOF problem, ignoring the prior kinematic constraints in robotics. Mobile robots and ground vehicles usually operate on a plane, such as hazardous chemical warehouse, home, and other scenes. The planar motion constraint can simplify the 6-DOF pose problem to a 3-DOF pose problem, reducing the number of matching features required for estimating the essential matrix, thus reducing the computational burden of the robot, improving real-time performance, effectively improving fault tolerance, and enhancing algorithm robustness. SUMMARY

[0007] The present application aims to overcome the shortcomings of the prior art and provide a planar mobile robot visual localization method and device without a three-dimensional map model. The prior kinematic constraint is used to simplify the epipolar constraint and triangulation. Based on the simplified epipolar constraint, only two feature point matches are needed for essential matrix estimation. Based on the simplified triangulation, only one pair of feature matches in the second reference image can uniquely determine the absolute translation scale, complete the absolute pose estimation, and be called the 2p1p method. The robot pose is estimated with the least number of feature matches during localization, thereby reducing the computational burden of the robot. The absolute translation scale is obtained by triangulation to get the absolute pose, and the application scenarios are more widely used. A multiple check scheme is adopted to improve the robustness and accuracy of localization.

[0008] In the first aspect, the present application provides a planar mobile robot visual localization method without a three-dimensional map model, comprising the following steps:

[0009] S1: Obtain an environment map, which is a two-dimensional image database composed of N reference images with pose annotations; N≥2;

[0010] S2: Obtain a current observed query image i from a camera mounted on a planar mobile robot, and extract and pair feature points from the first K reference images retrieved from the two-dimensional image database, to obtain m pairs of 2D-2D matching feature points, K≥2, m≥3; and find at least 2 pairs of 2D-2D matching feature points between reference image j1 and query image i, and at least 1 pair of 2D-2D matching feature points between reference image j2 and query image i; reference image j1 and reference image j2 are different images, and the transformation matrix T is known.

[0011] S3: Simplify epipolar constraint according to prior kinematic constraint of planar mobile robot; based on the simplified epipolar constraint, take any two pairs of 2D-2D matching feature points existing between reference image j1 and query image i, and calculate current rotation matrix R and scale-free translation vector t~ of the camera;

[0012] S4: Introduce motion characteristics into triangulation, calculate scale of translation vector to obtain absolute pose in reference image j2 by taking any one pair of 2D-2D matching feature points;

[0013] S5: Perform translation consistency check according to the calculated absolute pose;

[0014] S6: Statistically count inliers of the absolute pose passing the translation consistency check, and perform 6-DOF nonlinear optimization on the absolute pose according to the number of inliers to obtain a final positioning result.

[0015] In a second aspect, the present application provides a planar mobile robot visual positioning device without a three-dimensional map model, comprising:

[0016] A power module for supporting hardware power supply;

[0017] A motion module for supporting movement of the robot;

[0018] An acquisition module for acquiring a query image of a current pose of the robot, and acquiring the first K reference images from a two-dimensional image database;

[0019] A matching module for calculating feature points of the query image and the reference image, matching the calculated feature points in the query image with the feature points of the reference image to obtain m pairs of 2D-2D feature matching points;

[0020] A calculation module for estimating an absolute pose of the robot according to the obtained 2D-2D feature matching points;

[0021] An inspection module for performing translation consistency check on the absolute pose;

[0022] An optimization module for performing 6-DOF nonlinear optimization on the absolute pose passing the translation consistency check to obtain a final positioning result.

[0023] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, when the computer program is executed in a computer, the computer executes the method.

[0024] In a fourth aspect, the present application provides a computing device comprising a memory and a processor, the memory having executable code stored therein, and the processor executes the executable code to implement the method.

[0025] Compared with the prior art, the application has the beneficial effects that:

[0026] 1) The application adopts a two-dimensional image data combined with pose estimation method, realizes the robot positioning technology, can avoid establishing a complex three-dimensional model data, and thus simplifies the positioning process.

[0027] 2) The application proposes a 2P1P minimum solution method, uses prior kinematic constraints to simplify the epipolar constraint, and reduces the number of required feature matching points. Only two pairs of feature matching points can accurately estimate the relative pose of the robot, and in the case that there are a large number of false matching in the visual data, using fewer matching feature points can improve the positioning success rate and greatly improve the system robustness.

[0028] 3) The application proposes a 2P1P minimum solution method, creatively introduces the motion characteristics into triangulation, and further reduces the number of required feature matching points. Only one pair of feature matching points of another reference image can accurately estimate the absolute pose of the robot. The traditional method is to calculate two relative poses and triangulate to obtain the absolute pose.

[0029] 4) The application combines the proposed 2P1P minimum solution method with a multiple check scheme to establish a robust three-dimensional model-free visual positioning framework, reduces the sensitivity of the positioning system to environmental changes and noise. The multiple check scheme can verify and correct possible errors through multiple independent check steps, thereby improving the accuracy and stability of positioning. Instead of relying on a three-dimensional model established in advance, the positioning is realized through multiple checks and analysis of images and scenes, thereby reducing the dependence on three-dimensional models and enhancing the adaptability to complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is a flowchart of the plane mobile robot visual positioning method without a three-dimensional map model of the application.

[0031] Figure 2 It is a schematic diagram of a plane motion coordinate system.

[0032] Figure 3 It is a module relationship flowchart. DETAILED DESCRIPTION

[0033] In the following, the technical solutions of the application are further described in combination with the drawings and specific embodiments.

[0034] Figure 1 It is a flowchart of the plane mobile robot visual positioning method without a three-dimensional map model of the application, and the application discloses a plane mobile robot visual positioning method without a three-dimensional map model, including the following steps:

[0035] S1: Obtain an environment map, specifically a two-dimensional image database composed of a set of reference images with pose annotations, without the need to construct a three-dimensional point cloud map model; Figure 2 A schematic diagram of a planar motion coordinate system is shown.

[0036] S2: The camera carried by the planar mobile robot obtains a currently observed query image i, and uses the Netvlad algorithm to retrieve the top five reference images most similar to the query image from the two-dimensional image database to extract and pair feature points, obtaining m pairs of 2D-2D matching feature points, m>=3; and in the above-mentioned 2D-2D matching feature points, at least 2 pairs of 2D-2D matching feature points exist between the reference image j1 and the query image i, and at least 1 pair of 2D-2D matching feature points exist between the reference image j2 and the query image i; the reference image j1 and the reference image j2 are different images, and the transformation matrix T is known.

[0037] The above-mentioned extraction and pairing of feature points of reference images are specifically: traditional sift, brief algorithm or deep learning method R2D2, Superpoint algorithm is used to extract feature points; and then the feature points are matched to obtain 2D-2D feature matching points.

[0038] The two-dimensional image database contains transformation matrix T information between reference images, specifically: the reference camera coordinate systems of the reference image j1 and the reference image j2 are {C j1} and {C j2} respectively, and the transformation matrix Wherein is the rotation matrix from {C j1} to {C j2}, is the translation vector from {C j1} to {C j2}.

[0039] S3: Simplify the epipolar constraint according to the prior kinematic constraint of the planar mobile robot; based on the simplified epipolar constraint, any two pairs of 2D-2D matching feature points between the reference image j1 and the query image i are taken, and the current rotation matrix R and the translation vector t without scale of the camera are calculated Specifically, the following sub-steps are included:

[0040] According to the movement of the robot on the plane, the robot rotates around an axis Y, and the robot translates on the plane composed of X and Z, so the rotation matrix R and the translation vector t can be represented as:

[0041]

[0042]

[0043] where θ is the unknown, representing the angle of rotation of the robot around the Y axis; φ is the unknown, representing the direction angle of the robot movement; and p is the unknown, representing the scale of the robot translation, i.e., the distance of movement.

[0044] According to the essential matrix E = t ∧ R, we have:

[0045]

[0046] where the ∧ symbol represents the outer product operation, i.e., the vector is written as an anti-symmetric matrix

[0047] Since the reference image j1 and the query image i have two pairs of 2D-2D matching feature points, we have:

[0048] where The superscript a in the above equation represents the serial number of the matching point, and the subscript b represents the serial number of the image where the matching point is located; represents the coordinate value of the point on the x-axis of the image, represents the coordinate value of the point on the y-axis of the image.

[0049] Substituting equation (3) into the epipolar constraint we have:

[0050]

[0051]

[0052] Solving the above two equations simultaneously, we have:

[0053] φ = arctan 2 (sinφ, cosφ) (6)

[0054] θ = arctan2 (sin(θ-φ), cos(θ-φ)) + φ (7)

[0055] Solving the above two equations simultaneously, we have θ and unknown, substitute equation (3) to obtain the essential matrix E.

[0056] Performing pose decomposition on the essential matrix E, we can obtain four relative poses where is the normalized translation vector and has R s and R s are related by a 180° rotation around the baseline.

[0057] Using the chirality test to screen out the correct relative pose in , we have the robot rotation matrix R and the scale-free translation vector

[0058] S4: Introduce motion characteristics into triangulation. Calculate the scale of the translation vector by arbitrarily selecting a pair of 2D-2D matched feature points in the reference image j2 to obtain the absolute pose. This includes the following sub-steps:

[0059] Based on the transformation matrix between the two reference camera coordinate systems in the prior map It is known that:

[0060]

[0061] Equation (9) can be rewritten The essential matrix is:

[0062]

[0063] The specific representation of a pair of 2D-2D matching feature points in reference image j2 and query image i:

[0064] Using epipolar constraints We can obtain:

[0065]

[0066] Solving for the unique solution ρ yields a solution ρ. We then determine if the unique solution ρ satisfies ρ > 0. If it does, we consider it to have passed the positive depth check, obtaining the scale of the translation vector and thus the robot's absolute pose (R, t). Otherwise, return to step S3.

[0067] S5: Perform a translation consistency check based on the calculated absolute pose, specifically including the following sub-steps:

[0068] Based on the relative transformation between the query image and the reference image Where R j It is the rotation vector that rotates the reference image to the map, c i It is the camera center for querying images, c j It is the camera center of the reference image. The direction should be the same as that recovered from the essential matrix. The directions are consistent. Use To quantify consistency, a threshold α is set. th If α < α th If so, the consistency check is considered passed.

[0069] S6: Perform 6-DOF nonlinear optimization based on the matched feature points to obtain the final localization result. This includes the following sub-steps:

[0070] 6-1 Select n pairs of 2D-2D matching feature points, where 3 ≤ n ≤ m;

[0071] 6-2 update the absolute pose (R, t) by using steps S2-S5;

[0072] 6-3 calculate the essential matrix E according to the absolute pose (R, t) by using formula (3);

[0073] 6-4 substitute all 2D-2D feature matching points into formula (4) to calculate the reprojection error; If the result is less than a certain threshold, it is an inlier, and if it is greater than a certain threshold, it is an outlier, and the number of inliers is counted; repeat steps 6-2 to 6-4 until the maximum number of iterations is reached;

[0074] 6-5 according to the round with the most inliers in all iterations, the absolute pose (R, t) obtained in this round of iteration is the optimal solution, and the 6-DOF optimization is performed on the optimal solution based on the inliers;

[0075] The 6-DOF optimization is specifically to minimize the sum of epipolar constraints, and a method such as Gauss-Newton is used to optimize and solve to obtain the optimized pose.

[0076] The application also discloses a camera visual positioning device, which comprises a power supply unit, a motion unit, a camera unit and a processing unit. The power supply unit supplies power to other units. The motion unit is used to support the movement of the robot. The camera unit is used to construct a two-dimensional image database and acquire a current image of the robot. The processing unit further comprises a processor, a memory and a program stored in the memory, and the computer program is executed by the processor to realize the steps of the plane mobile robot visual positioning method without a three-dimensional map model.

[0077] The steps of acquiring a current query image of the robot from the camera unit, matching feature points of the current query image and reference images in the database, calculating the pose of the robot according to the matched feature points, checking and optimizing according to the calculated pose, etc. all belong to algorithms and are completed on a computer or a processor carried by the robot.

[0078] The method in the embodiment and the device in the foregoing embodiments are two aspects based on the same inventive concept, and a person skilled in the art can clearly understand the system structure and implementation process in the embodiment according to the foregoing description.

[0079] For the convenience of description, when describing the above device, the functions are divided into various modules and described respectively, and a module relationship flowchart is as shown in Figure 3 Of course, the functions of the modules can be realized in the same software or hardware or multiple software or hardware when implementing the application. The functions of the specific modules are as follows:

[0080] The power module is used to support the power supply of the hardware;

[0081] The motion module is used to support the movement of the robot;

[0082] an acquisition module configured to acquire a query image of a current pose of the robot and acquire K reference images from a two-dimensional image database;

[0083] a matching module configured to calculate feature points of the query image and the reference images, and match the feature points calculated in the query image with the feature points of the reference images to obtain m pairs of 2D-2D feature matching points;

[0084] a calculation module configured to estimate an absolute pose of the robot according to the obtained 2D-2D feature matching points;

[0085] a checking module configured to perform a translational consistency check on the absolute pose;

[0086] an optimization module configured to perform a 6-DOF nonlinear optimization on the absolute pose that passes the translational consistency check to obtain a final positioning result.

[0087] The method of the present application and the conventional 5p5p and 8p8p methods will be compared in the following experiment. The 5p5p refers to using a two-time five-point method to obtain a relative pose of the robot, and then triangulating to obtain an absolute pose of the robot. The 8p8p refers to using a two-time eight-point method to obtain a relative pose of the robot, and then triangulating to obtain an absolute pose of the robot.

[0088] The robot was controlled to move in a hall, and a total of 50 images with a resolution of 1024x540 were collected. The ground truth camera poses were obtained by aligning dense synchronous 3D scans. Four images were selected from the 50 images as reference images of the map. The remaining 46 images were used as query images. The rotation error between the estimated pose and the true pose of each query image was less than 10 degrees, and the distance error was less than 0.5 meters, and the positioning was considered successful. The success rate of the method of the present application was 15.22% higher than that of the 8p8p method and 8.7% higher than that of the 5p5p method.

[0089] Based on the above embodiment description, those skilled in the art can clearly realize that the present application can be realized by a general hardware platform combined with software. Based on such understanding, the technical solutions of the present application essentially embody the contribution to the prior art in the form of a software product. The software product can be stored in various storage media, such as ROM / RAM, disk, etc. It contains programs, so that computer devices can execute the methods described in various embodiments or parts of embodiments of the present application. These computer devices can be personal computers, servers or MCUs, etc.

[0090] The described apparatus embodiments are only schematic and many variations can be made to the described embodiments in their implementation. For example, as those skilled in the art will appreciate, the described embodiments can be implemented in any suitable electronic device, for example, a mobile phone, tablet computer, computer, laptop computer, desktop computer, server computer, or any other suitable electronic device. In its implementation, the described embodiments can also include any sub-combination of the described features, or variations thereof, as well as any suitable implementation of the described embodiments.

[0091] Those skilled in the art will appreciate that the application described herein is susceptible to variations and / or modifications as can be best suited to a particular application and / or environment. The application includes any and all such modifications.

[0092] The application is applicable for a wide range of general or specific computing system environments or configurations. These environments include, but are not limited to, for example: personal computers; server computers; handheld or laptop devices; tablet devices; multiprocessor systems; microprocessor-based systems; set-top boxes; programmable consumer electronics; network PCs; minicomputers; mainframe computers; distributed computing environments that include any of the above systems or devices; and the like. In general, the application can be applied to virtually any computing system environment or configuration, encompassing a wide range of personal devices to large computers and distributed computing environments.

[0093] The above-described embodiments are merely preferred embodiments of the application, and cannot be used to limit the scope of protection of the application. Any non-essential changes and replacements made by those skilled in the art based on the application without departing from the spirit and scope of the technical scheme of the application shall fall within the scope of protection of the application.

Claims

1. A visual localization method for a planar mobile robot that does not require a 3D map model, characterized in that, Includes the following steps: S1: Obtain an environmental map, which is a two-dimensional image database consisting of N reference images with pose annotations; N≥2; S2: The camera on the planar mobile robot acquires the currently observed query image i, and extracts and pairs the feature points with the first K reference images retrieved from the two-dimensional image database to obtain m pairs of 2D-2D matching feature points, where K≥2 and m≥3; In the above-mentioned 2D-2D matching feature points, at least two pairs of 2D-2D matching feature points exist between the reference image j1 and the query image i, and at least one pair of 2D-2D matching feature points exist between the reference image j2 and the query image i; the reference image j1 and the reference image j2 are different images, and the transformation matrix T is known; S3: Simplify the epipolar constraints based on the prior kinematic constraints of the planar mobile robot; based on the simplified epipolar constraints, randomly select two pairs of 2D-2D matching feature points between the reference image j1 and the query image i, and calculate the current rotation matrix R of the camera and the scale-free translation vector for them. S4: Introduce motion characteristics into triangulation, and calculate the scale of the translation vector by arbitrarily taking a pair of 2D-2D matching feature points in the reference image j2 to obtain the absolute pose; S5: Perform a translation consistency check based on the calculated absolute pose; S6: Statistically analyze the interior points of the absolute pose that have passed the translation consistency check, and perform 6-DOF nonlinear optimization on the absolute pose based on the number of interior points to obtain the final positioning result.

2. The planar mobile robot visual localization method without a 3D map model as described in claim 1, characterized in that, The reference camera coordinate systems of the reference image j1 and the reference image j2 are respectively {C j1 } and {C j2 }, then the transformation matrix in For {C j1 } to {C j2 The rotation matrix of} For {C j1 } to {C j2 The translation vector of}.

3. The planar mobile robot visual localization method without a three-dimensional map model as described in claim 1, characterized in that, Step S3 is as follows: Based on the robot's movement in a plane, its rotation around an axis (let's call it the Y-axis), and its translation in the plane formed by the X and Z axes, the rotation matrix R and the translation vector t can be expressed as: Where θ is an unknown quantity, representing the angle of the robot's rotation around the Y-axis; φ is an unknown quantity, representing the direction angle of the robot's movement; and ρ is an unknown quantity, representing the scale of the robot's translation, i.e., the distance it moves. According to the essential matrix E = t ^ R yields: The ∧ symbol represents the outer product operation, that is, the cross product operation of vectors. Write it as an antisymmetric matrix Furthermore, the specific representations of the two pairs of 2D-2D matching feature points in the reference image j1 and the query image i are as follows: in In this context, the superscript 'a' represents the sequence number of the matching point, and the subscript 'b' represents the image sequence number where the matching point is located. This represents the coordinates of a point on the x-axis of the image. This represents the coordinates of a point on the y-axis of the image. Substituting equation (3) into the epipolar constraint We can obtain: Solving the two equations simultaneously, we get: φ=arctan 2(sinφ,cosφ) (6) θ=arctan2(sin(θ-φ),cos(θ-φ))+φ (7) By combining the two equations above, we can obtain θ and Substitute the unknown quantities into equation (3) to obtain the essential matrix E; Pose decomposition of the essential matrix E yields four relative poses. in It is a normalized translation vector and has R s and R s ′Related around the baseline 180°; Using chiral testing in The correct relative pose is selected from the data and used as the robot's rotation matrix R and scale-free translation vector.

4. The planar mobile robot visual localization method without a three-dimensional map model as described in claim 1, characterized in that, Step S4 specifically includes the following sub-steps: Based on the transformation matrix between the two reference camera coordinate systems in the prior map It is known that: Equation (9) can be rewritten The essential matrix is: The specific representation of a pair of 2D-2D matching feature points in reference image j2 and query image i: Using epipolar constraints We can obtain: Solving for the unique solution ρ yields a solution ρ. We then determine if the unique solution ρ satisfies ρ > 0. If it does, we consider it to have passed the positive depth check, obtaining the scale of the translation vector and thus the robot's absolute pose (R, t). Otherwise, return to step S3.

5. The planar mobile robot visual localization method without a three-dimensional map model as described in claim 1, characterized in that, Step S5 specifically includes the following sub-steps: Based on the relative transformation between the query image and the reference image Where R j It is the rotation vector that rotates the reference image to the center of the world, c i It queries the camera center of image i, c j It is the camera center of the reference image j1; The direction and essence matrix recovered The direction is consistent; use To quantify consistency; to determine whether α < α th α th This represents the threshold. If the threshold is met, the consistency check is considered to have passed; otherwise, the consistency check is considered to have failed, and the process returns to step S3.

6. The planar mobile robot visual localization method without a three-dimensional map model as described in claim 1, characterized in that, Step S6 specifically includes the following sub-steps: S6-1 Take n pairs of 2D-2D matching feature points, where 3≤n≤m; S6-2 updates the absolute pose (R,t) using steps S2-S5; S6-3 calculates the essential matrix E based on the absolute pose (R,t) using equation (3); S6-4 Substitute all 2D-2D feature matching points Calculate the result; if the result is less than a certain threshold, it is an interior point; if it is greater, it is an exterior point. Count the number of interior points. Repeat steps 6-2 to 6-4 above until the maximum number of iterations is reached. S6-5 takes the round with the most inliers among all iterations as the optimal solution, and optimizes the optimal solution with 6 degrees of freedom based on the inliers.

7. A planar mobile robot visual positioning device that implements the method of any one of claims 1-6, characterized in that, include: The power module is used to support hardware power supply; Motion module, used to support the robot's movement; The acquisition module is used to acquire the query image of the robot's current pose and to obtain the first K reference images from the two-dimensional image database; The matching module is used to calculate the feature points of the query image and the reference image. Based on the feature points calculated in the query image, the feature points in the reference image are matched to obtain m pairs of 2D-2D feature matching points. The calculation module is used to estimate the robot's absolute pose based on the obtained 2D-2D feature matching points; The inspection module performs a translation consistency check on the absolute pose. The optimization module performs 6-DOF nonlinear optimization on the absolute pose that has passed the translation consistency check to obtain the final positioning result.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-6.

9. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-6.

Images