A method for simultaneous localization and 3D map construction in dynamic environments

Abstract

This invention discloses a method for simultaneous localization and 3D map construction in dynamic environments, comprising the following steps: S1, filtering feature points extracted from objects marked as dynamic based on semantic information; S2, matching the remaining feature points with the previous frame, and using the matched feature point pairs to calculate a stable fundamental matrix; S3, filtering out all real dynamic feature point pairs through geometric constraints, and retaining static feature point pairs. Compared to traditional methods that directly eliminate prior dynamic feature points based on semantic information, this invention employs a dynamic feature filtering strategy, combining semantic information and geometric constraints to jointly detect moving target objects, filtering out feature points of moving targets in the current frame. This effectively improves the localization accuracy and stability in dynamic environments, enabling the robot to fully perceive its surroundings and operate safely.

Description

Technical Field

[0001] This invention relates to the field of visual SLAM, and in particular to a method for simultaneous localization and 3D map construction in dynamic environments. Background Technology

[0002] Since their inception in the 20th century, robots have developed rapidly and are now applied across various industries, including manufacturing, agriculture, military, and daily services. Their level of intelligence is increasing, and robots are becoming more and more integrated into human life. From the widely deployed automated guided vehicles (AGVs) and household robotic vacuum cleaners to the current industrial research hotspot of autonomous driving, their achievements and development potential are remarkable. For mobile robots, a crucial issue is how to acquire positional information from the environment—in other words, how to create a high-precision map of the unknown environment to implement subsequent path planning algorithms. This is the perception problem, one of the four core problems of robotics. Perceiving the external environment is a vital prerequisite for robots to move stably in the complex environments of human society, avoiding danger and approaching targets while avoiding obstacles. Therefore, robot navigation systems should enable robots to fully perceive their surroundings to operate safely.

[0003] For robots to integrate into human life, they must first be able to move stably within human environments. Human environments are complex, containing many static facilities serving humans, as well as objects in constant motion. This includes not only the movement of individual humans, but also the movement of vehicles used by humans and the movement of animals within human society. These are all challenges that intelligent robots face in operating safely and stably within human environments.

[0004] Simultaneous Localization and Mapping (SLAM) has been a core problem in robot navigation since its inception in the 1990s. It relies on cameras mounted on mobile robots to continuously capture images of the environment, creating a simplified 3D map of the surroundings from the acquired image pairs. Its goal is to construct a map of the environment while simultaneously estimating the robot's pose parameters. Traditional SLAM methods primarily rely on relatively stable distance sensors such as LiDAR. However, the distance data obtained by LiDAR is very sparse, resulting in SLAM-constructed environmental maps containing only a very small number of landmarks. This map can only be used to improve the robot's localization accuracy, but not for other areas of robot navigation such as path planning. Furthermore, the high cost, large size, weight, and power consumption of LiDAR limit its application in certain fields. Compared to sensors like LiDAR, vision can not only obtain distance and depth information of the environment, but also the color of objects, and, combined with relevant algorithms, achieve object recognition. The amount of information it acquires geometrically surpasses that of sensors like LiDAR.

[0005] Currently, there are many mature visual SLAM solutions, such as the well-known ORB-SLAM2 system (a real-time monocular SLAM system based on feature points). However, most of these SLAM solutions are based on a static scene assumption, i.e., a static environment with good texture and lighting. But reality is often a dynamic environment with a large number of objects or people interfering, which brings difficulties to the SLAM problem. Traditional SLAM algorithms only solve the localization and mapping problems in static environments; however, when there are many moving objects in the environment, the robot's localization will produce large errors, or even lose tracking. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of improving the accuracy of localization and mapping of mobile robots in dynamic environments, and to provide a method for simultaneous localization and 3D map construction in dynamic environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for simultaneous localization and 3D map construction in a dynamic environment includes the following steps:

[0009] S1. Filter the feature points extracted from objects marked as dynamic based on semantic information;

[0010] S2. Match the remaining feature points with the previous frame and use the matched feature point pairs to calculate the stable basis matrix;

[0011] S3. Filter out all real dynamic feature point pairs through geometric constraints, and retain static feature point pairs.

[0012] In some embodiments, step S1 includes: using an object detection algorithm to filter feature points extracted from objects marked as dynamic.

[0013] In some embodiments, the filtering of features extracted from objects marked as dynamic using object detection algorithms specifically involves filtering each extracted ORB feature using semantic detection boxes generated by YOLO v5.

[0014] In some embodiments, step S2 includes: traversing all map points in the reference frame, projecting them onto the current frame respectively, and finding a feature point in the current frame that is closest to the descriptor as its matching point.

[0015] In some embodiments, step S3 includes: if the feature point pair is within the dynamic semantic bounding box, identifying the feature point pair with D>λ1θ as a true dynamic feature point pair and removing it; if the feature pair is outside the dynamic semantic bounding box, identifying the feature point pair with D>λ2θ as a true dynamic feature point pair and removing it; where λ1 and λ2 are two different thresholds. It is a preset threshold.

[0016] In some embodiments, λ1 is less than λ2.

[0017] In some embodiments, the preset threshold It can be expressed by the following formula:

[0018]

[0019] The total number of ORB feature points is expressed as follows: , It is the first The distance from each feature to its polar line.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the above-described method.

[0021] The present invention also provides a robot having a computer program stored thereon, which, when executed by a processor, can implement the above-described method.

[0022] The present invention has the following beneficial effects:

[0023] This invention employs a dynamic feature filtering strategy, combining semantic information and geometric constraints to detect moving target objects and filter out feature points of each moving target in the current frame. Compared to traditional methods that directly remove prior dynamic feature points based on semantic information, this effectively improves the positioning accuracy and stability in dynamic environments, enabling the robot to fully perceive its surroundings and operate safely.

[0024] Furthermore, compared to the semantic segmentation networks used in traditional methods, this embodiment of the invention selects a lighter target detection algorithm, which can effectively label objects in the scene, effectively solve the interference problem of non-static object localization and map construction in dynamic environments, and improve the robustness of the system. Attached Figure Description

[0025] Figure 1 This is a flowchart of the simultaneous localization and 3D map construction method in a dynamic environment according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart of the SLAM method for dynamic scenes based on semantic information fusion in an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of polar line constraint in an embodiment of the present invention;

[0028] Figure 4 This is a flowchart of the dynamic feature filtering strategy in an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of marking dynamic feature points in the experimental examples of this invention;

[0030] Figure 6 This is a schematic diagram of preserving static feature point pairs in the experimental examples of this invention. Detailed Implementation

[0031] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0032] In practical applications, intelligent robots cannot operate entirely in static environments, making SLAM algorithms particularly important in dynamic environments. Furthermore, traditional SLAM algorithms only provide the robot with geometric information about the environment; that is, the robot can only acquire the positional information of obstacles. While this information allows the robot to plan a navigation path, it cannot utilize environmental information to perform more advanced navigation tasks. In summary, SLAM technology aims to solve two main problems: first, improving the accuracy of localization and mapping for mobile robots in dynamic environments; and second, acquiring higher-level semantic information from the surrounding environment.

[0033] The existing technology has the following drawbacks:

[0034] 1. Currently, only a few studies involve SLAM methods in dynamic environments. For example, Professor Zhang Guofeng's research group at Zhejiang University introduced the RDSLAM method, which, based on the PTAM method, incorporates an adaptive RANSAC (Random Sample Consensus) method with temporal priors to filter out mismatched points on dynamic objects. This method is suitable for situations where the scene changes slowly in a small-scale indoor desktop environment. CD SLAM is a keyframe-based SLAM method. The paper proposes HoC (Higher Order Components) to filter non-static feature points, thereby maintaining relatively reliable SLAM over a long period. All the above methods are based on the same principle: treating moving objects in the image as outliers and excluding them during pose estimation, relying only on the static parts of the scene. This results in the accuracy of such methods depending on the proportion of feature points in the static parts of the scene. If there are too many or too dense dynamic objects in the scene, it will seriously affect the reliability of pose estimation, and consequently, the accuracy of map construction.

[0035] 2. Another approach is to combine deep learning for feature point discrimination. In recent years, deep learning technology has swept through the field of artificial intelligence research. Deep learning has demonstrated excellent performance in vision-related image recognition and segmentation. Therefore, there has been a growing call to introduce deep learning into the field of robot navigation. Some classic semantic segmentation algorithms, such as YOLACT++, Polarmask, and Mask r-cnn, are highly anticipated in their combination with SLAM. For example, DynaSLAM (a visual SLAM system built on ORB-SLAM2) designed by Bescos et al. introduces Mask r-cnn into ORB-SLAM (a feature point-based visual SLAM based on ORB descriptors), using semantic information obtained from instance segmentation to label feature points and filter out interference from non-static objects in the SLAM system. Its disadvantages are as follows: 1) Most existing solutions utilize deep learning masks to remove all feature points on objects with prior dynamic properties. This leads to the system mistakenly removing feature points on static objects even when the target is static, causing errors in camera pose estimation and resulting in the loss of information about these objects in the constructed static scene map; 2) Existing solutions use semantic segmentation to remove non-static feature points, but semantic segmentation networks have high computational costs and low running speeds, making it difficult to achieve real-time performance when combined with SLAM. On the other hand, semantic segmentation can result in incomplete segmentation and redundant extensions; 3) Existing solutions that use semantic and epipolar constraints to filter dynamic feature points have significant errors in the fundamental moments used for epipolar constraints, affecting algorithm performance.

[0036] like Figure 1 As shown, the simultaneous localization and 3D map construction method in a dynamic environment according to an embodiment of the present invention includes the following steps:

[0037] S1. Filter the feature points extracted from objects marked as dynamic based on semantic information;

[0038] S2. Match the remaining feature points with the previous frame and use the matched feature point pairs to calculate the stable basis matrix;

[0039] S3. Filter out all real dynamic feature point pairs through geometric constraints, and retain static feature point pairs.

[0040] The problems that this invention aims to solve are three issues with existing SLAM algorithms in dynamic environments:

[0041] By using deep learning networks, all feature points on objects with prior dynamic properties are removed. This will cause the system to mistakenly remove feature points on static objects even when the target is static, resulting in errors in camera pose estimation and causing the constructed static scene map to lose information about these objects.

[0042] Existing solutions use semantic segmentation to remove non-static feature points, but semantic segmentation networks are computationally expensive and slow, significantly reducing system real-time performance when combined with SLAM. Furthermore, semantic segmentation can lead to incomplete segmentation and redundant extensions.

[0043] Existing schemes that use semantic and epipolar constraints to filter dynamic feature points have significant errors in the basis matrix used for epipolar constraints, which affects the algorithm's performance.

[0044] This invention proposes a novel dynamic ORB feature filtering method, known as the dynamic scene SLAM method based on semantic information fusion.

[0045] The flowchart of the method in this embodiment is as follows: Figure 2 As shown. First, the semantic detection boxes generated by YOLO v5 (an object detection model) are used to filter each extracted ORB feature. Then, the remaining feature points are matched with the previous frame, and the stable basis matrix is ​​calculated using the matched feature point pairs. Finally, all true dynamic feature point pairs are filtered out through geometric constraints, as detailed below:

[0046] 1. Semantic detection

[0047] This invention employs YOLO v5 to detect objects labeled as dynamic. YOLO v5 can be implemented in real-time on embedded platforms. The method is trained on the PASCAL VOC (Visual Object Classes) dataset and can detect a total of 20 classes. In indoor environments, people are likely to be dynamic objects; therefore, in this invention, people are labeled as dynamic objects. Besides YOLO v5, other high-performance object detection algorithms can also be used to identify and remove dynamic feature points.

[0048] Filtering out all features extracted from objects marked as dynamic is a simple way to improve the accuracy of SLAM systems. However, this approach has two problems. First, in a real indoor environment, people are not the only independent dynamic objects. For example, a chair moved by a person is also a dynamic object. Second, YOLO v5 can only provide one detection box, which may contain many static features. When the detection box occupies more than half of the current frame, there are few remaining features, which may cause the tracking thread in the SLAM system to fail.

[0049] In summary, using semantic information filtering on all features extracted from objects marked as dynamic is insufficient to improve localization accuracy and can lead to tracking failure in some extreme cases.

[0050] In this embodiment of the invention, feature points extracted from objects marked as dynamic are first filtered based on their semantic information. Then, a stable fundamental matrix is ​​calculated based on the remaining features. Finally, all true dynamic feature point pairs are filtered out using epipolar constraints.

[0051] Feature matching is performed between the features and the previous frame: all map points in the reference frame are traversed, and each point is projected onto the current frame. In the current frame, the feature point with the closest descriptor distance is found as its matching point.

[0052] 2. Geometric constraint method

[0053] Given the limitations of semantic constraints, this embodiment of the invention uses geometric constraints to further filter out dynamic features. Epipolar constraints, or geometric constraints, are part of photographic geometry. They exist independently of the external environment between two images and depend only on the camera's internal parameters and relative pose. The epipolar constraint relationship between two adjacent images is as follows: Figure 3 As shown.

[0054] refer to Figure 3 , and These are two images captured from the same scene. It is a point in space, which is and The projections on are respectively and . and It is the optical center of two cameras, and their connections are respectively at... and Place and and intersect. and yes and The polar line can be connected with... , and Obtained by the intersection of the image planes; x2 and p refer to the image point of the spatial point P (unknown) on I1 being x1. (Given), then the image point of world point P on I2 must lie on the polar line l2. For example, if a spatial point is at position p, then its image point on I2 is x2, which is also on the epipolar line l2; this is the epipolar constraint.

[0055] if and Spatial geometric transformations between them are denoted as ,So and It can be represented as

[0056] (1)

[0057] in , yes and The coordinates of P in the camera coordinate system. , It is its depth value. R is the camera's parameter matrix, and R and T represent the rotation matrix and translation vector, respectively. The rotation matrix and translation vector describe the motion of a rigid body in three-dimensional space.

[0058] If normalized coordinates Then we can obtain the following equation:

[0059] (2)

[0060] Multiply the above equation by Then you will get:

[0061] (3)

[0062] Where T represents the transpose of the matrix;

[0063] in It is the antisymmetric matrix of vector t. Substituting equation (1) into equation (3), we obtain the following equation:

[0064] (4)

[0065] Where F is the fundamental matrix, and T represents the transpose of the matrix. The above equation illustrates the correct relationship of matching feature pairs. In other words, if one feature point correctly matches another feature point, it must lie on the epipolar line of its image plane. However, in real-world scenarios, due to the influence of noise and camera calibration parameter errors, the calculated fundamental matrix often contains certain errors, often preventing feature points from accurately falling on the calculated epipolar line. That is, the influence of noise and erroneous matching points caused by the dynamic environment may result in a large error in the calculated fundamental matrix, thus yielding a non-stable fundamental matrix. Therefore, we can use the distance of each feature point to the epipolar line to measure the reliability of the matching point pair.

[0066] if and The homogeneous coordinates are respectively ,but The epipolar line can be calculated using the following equation:

[0067] (5)

[0068] in, , and Let represent the polar equation vector. Then, the equation from point can be obtained using the following formula. Distance to the poles:

[0069] (6)

[0070] in express to the poles The distance. In real-world scenarios, there are typically two situations that can cause this distance to be too large. The first is feature point mismatch, and the second is feature points extracted from dynamic objects. Regarding distance... Constraints can be used to remove mismatched point pairs and dynamic features. If Then the point pair is considered a dynamic feature, where It is a preset threshold, which can be calculated according to equation (7).

[0071] (7)

[0072] The total number of ORB feature points is expressed as follows: , It is the first The distance from each feature to its polar line.

[0073] 3. Dynamic Feature Filtering Strategy

[0074] Using semantic information alone cannot fully filter dynamic features, and using geometric constraint methods alone also has some problems. Specifically, when only some dynamic feature is used to calculate the fundamental matrix, the calculated fundamental matrix will contain a large error.

[0075] This invention combines semantic and geometric constraints to identify dynamic features of a scene:

[0076] First, the extracted ORB features are filtered using semantic detection boxes generated by YOLO v5. Then, the remaining feature points are matched against the previous frame, and the stable basis matrix is ​​calculated using the matched feature pairs filtered by the dynamic object detection boxes. Finally, all true dynamic feature point pairs are filtered out using geometric constraints. A detailed flowchart of the dynamic feature filtering strategy is shown below. Figure 4 As shown, if feature point pair i is within the dynamic semantic bounding box, i.e., the semantic detection box generated by YOLO v5, feature point pairs with D>λ1θ are identified as true dynamic feature point pairs and removed, while static feature point pairs are retained; if feature point pair i is outside the dynamic semantic bounding box, feature point pairs with D>λ2θ are identified as true dynamic feature point pairs and removed, while static feature point pairs are retained; where, Indicates the distance between the feature and the polar line. The result is obtained from equation (7). It can be seen that two different thresholds, λ1 and λ2, are set to filter dynamic features in the algorithm, because features in the detection box marked as dynamic objects are likely to be true dynamic features. Therefore, features in the detection box region are filtered more strictly, which means that λ1 is less than λ2.

[0077] This invention primarily includes a dynamic scene SLAM method based on semantic information fusion, which improves positioning accuracy in dynamic environments. The improvements of this invention compared to existing technologies include:

[0078] Compared to traditional methods that directly remove prior dynamic feature points based on semantic information, this invention employs a dynamic feature filtering strategy that combines semantic information and epipolar constraints to detect moving target objects, filter out feature points of each moving target in the current frame, and estimate the camera pose.

[0079] In the research process, the first step is to select a deep learning network to remove non-static features. For the image deep learning submodule, compared with the semantic segmentation network used in traditional methods, this embodiment of the invention selects a lighter-weight object detection algorithm, which can effectively label objects in the scene, effectively solve the interference problem of non-static object localization and map construction in dynamic environments, and improve the robustness of the system.

[0080] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the above-described method.

[0081] This embodiment also provides a robot that stores a computer program, which, when executed by a processor, can implement the above-described method.

[0082] This invention modifies epipolar constraints by fusing semantic information. Specifically, it filters feature points extracted from objects marked as dynamic based on semantic information, matches remaining feature points with the previous frame, and uses the matched feature point pairs to calculate a stable basis matrix. Semantic detection boxes generated using YOLO (an object detection model) are used to filter extracted ORB (a feature extraction algorithm) features. Then, a stable basis matrix is ​​calculated using the matched feature point pairs. Finally, geometric constraints are used to filter out all true dynamic feature point pairs. This invention improves epipolar constraints through semantic detection and semantic information: it calculates the basis matrix based on features outside the semantic detection box marked as dynamic objects, and employs a dynamic feature point filtering strategy: different filtering thresholds are applied to feature points inside and outside the dynamic object detection box. This solves the problems of large errors in robot localization in dynamic environments and the robot's inability to utilize environmental information to perform more advanced navigation tasks.

[0083] It can produce the following effects:

[0084] 1. Improve pose estimation accuracy in dynamic scenes.

[0085] This invention presents a dynamic environment camera pose estimation method based on semantic SLAM. For dynamic scenes, it proposes an improved pose estimation method based on semantic information and epipolar constraints. By combining semantic information to calculate a more accurate basis matrix, it detects moving target objects through epipolar constraint relationships and filters out feature points of moving targets in a more accurate way. This overcomes the influence of moving objects in the scene on camera pose estimation and improves the accuracy of camera pose estimation in dynamic scenes.

[0086] 2. Improve the speed of pose estimation.

[0087] This invention employs the lightweight neural network YOLO v5. Most existing solutions use semantic segmentation to remove non-static feature points. However, while semantic segmentation networks have low computational cost and high speed, their combination with SLAM significantly reduces system real-time performance. Using YOLO v5 to detect dynamically labeled objects allows for real-time implementation on embedded platforms, greatly improving the speed of SLAM systems in dynamic environments.

[0088] 3. Improved the stability and robustness of mapping.

[0089] The dynamic feature detection module accurately eliminates dynamic feature points, improving the stability and robustness of subsequent mapping and enhancing the map's ability to represent the environment.

[0090] Experimental Example

[0091] refer to Figure 5 This method utilizes YOLO v5 to detect and label dynamic objects. Since people are likely to be dynamic objects in indoor environments, they are labeled as such in this approach. In this example, the identified people are enclosed in a dynamic object detection box, and then a stable basis matrix is ​​calculated based on feature points outside the detection box. (See reference...) Figure 6 Then, geometric constraints are used to filter out all real dynamic feature point pairs, while retaining static feature point pairs. Experimental results are as follows: Figure 6 As shown, all dynamic feature points on the character are removed. Specifically, each extracted ORB feature is filtered using semantic detection boxes generated by YOLO v5. Then, the remaining feature points are matched with the previous frame, and the stable basis matrix is ​​calculated using the matched feature point pairs. Finally, all true dynamic feature point pairs are filtered out using geometric constraints.

[0092] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A method for simultaneous localization and 3D map construction in a dynamic environment, characterized in that, Includes the following steps: S1. Filter the feature points extracted from objects marked as dynamic based on semantic information; S2. Match the remaining feature points with the previous frame and use the matched feature point pairs to calculate the stable basis matrix; S3. Filter out all real dynamic feature point pairs through geometric constraints, and retain static feature point pairs; Step S3 includes: if the feature point pair is within the dynamic semantic bounding box, the feature point pair with D>λ1θ is identified as a true dynamic feature point pair and removed; if the feature point pair is outside the dynamic semantic bounding box, the feature point pair with D>λ2θ is identified as a true dynamic feature point pair and removed; where λ1 and λ2 are two different thresholds. It is a preset threshold, where λ1 is less than λ2; The preset threshold It can be expressed by the following formula: ； The total number of ORB feature points is expressed as follows: , It is the first The distance from each feature to its polar line.

2. The method as described in claim 1, characterized in that, Step S1 includes: using an object detection algorithm to filter feature points extracted from objects marked as dynamic.

3. The method as described in claim 2, characterized in that, The process of using object detection algorithms to filter features extracted from objects marked as dynamic specifically involves using semantic detection boxes generated by YOLO v5 to filter each extracted ORB feature.

4. The method as described in claim 1, characterized in that, Step S2 includes: traversing all map points in the reference frame, projecting them onto the current frame, and finding a feature point in the current frame that is closest to the descriptor as its matching point.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program can implement the method described in any one of claims 1-4.

6. A robot having a computer program stored thereon, characterized in that, When executed by a processor, the computer program can implement the method described in any one of claims 1-4.

Images