Visual slam method and system in dynamic environment

Abstract

The application discloses a visual SLAM method and system in a dynamic environment, and belongs to the technical field of visual images. The method comprises the following steps: acquiring image information and depth information acquired during movement of an RGB-D camera, so as to acquire a visual image and distance values of each pixel point; performing feature extraction on the visual image, so as to acquire a feature point set; performing target detection on the visual image, so as to acquire a potential dynamic region set; performing three-dimensional reconstruction on the feature points, so as to acquire three-dimensional space points; acquiring comparison results of the feature points and the three-dimensional space points and the potential dynamic region, so as to acquire dynamic suspicious points; updating a dynamic confidence degree, performing optimization demand verification on the dynamic suspicious points, so as to determine and suppress dynamic suppression points, and acquiring static feature points; performing pose optimization on the static feature points, so as to acquire stable static features, and performing three-dimensional reconstruction, so as to acquire a high-quality map. The method solves the problems of insufficient real dynamic suppression precision and easy misdeletion of static features in the prior art, and improves the construction quality of the map.

Description

Technical Field

[0001] This application belongs to the field of visual image technology, specifically relating to a visual SLAM method and system in a dynamic environment. Background Technology

[0002] With the development of mobile robots, augmented reality devices, and unmanned systems, these devices need to achieve autonomous localization and environmental mapping in unknown environments. To address this, a technique called SLAM (Simultaneous Localization and Mapping) is widely adopted, enabling devices to estimate their own position while simultaneously building a spatial model of their surroundings during movement. Among various SLAM techniques, visual SLAM is widely used due to its ability to acquire rich environmental information solely through a camera and its relatively low cost. Visual SLAM typically extracts feature points from images, matches them across consecutive image frames, estimates camera motion using geometric relationships, and simultaneously reconstructs stable features into 3D map points, thus forming an environmental map. However, most traditional visual SLAM methods are based on the premise of a "static environment," assuming that the positions of objects in the scene do not change. In real-world applications, such as indoor environments with people moving around or roads with vehicles passing through, there are often moving objects in the scene. The image features on these moving objects change position over time. If these features are mistakenly treated as static environmental features for pose estimation and mapping, it can lead to fluctuations in localization results, trajectory drift, and unstable map structures.

[0003] To address these issues, existing technologies primarily employ two approaches. One approach introduces semantic recognition technology, identifying objects such as pedestrians and vehicles through object detection or semantic segmentation, and then treating these areas as dynamic regions for removal. While this method can quickly remove obviously dynamic objects, it is prone to mistakenly deleting static background structures or temporarily stationary parts of objects because the detection results only reflect the object category rather than the actual motion state, leading to a reduction in usable features in the map. The other approach uses geometric constraints for dynamic discrimination, identifying dynamic points by analyzing whether feature points satisfy geometric consistency relationships across multiple frames. This type of method does not rely on semantic information, but when the direction of motion of a dynamic object is close to the direction of camera motion, or when the object moves along the epipolar line, dynamic points may still satisfy geometric relationships, making them difficult to identify correctly.

[0004] Therefore, in dynamic environments, how to effectively identify and suppress real dynamic features without accidentally deleting a large number of static features, while ensuring the integrity of the map structure and the stability of positioning, has become a technical problem that urgently needs to be solved in the field of visual SLAM. Summary of the Invention

[0005] To address the issues of insufficient accuracy in suppressing realistic dynamic features and the susceptibility to accidental deletion of static features in existing technologies, this application discloses a visual SLAM method and system for dynamic environments, specifically: First aspect: A visual SLAM method for dynamic environments, the method comprising: Acquire images captured during camera movement to obtain visual images; Feature extraction is performed on the visual image to obtain a set of feature points; Target detection is performed on the visual image to obtain a set of potential dynamic regions; The feature points are reconstructed in three dimensions to obtain three-dimensional spatial points; The comparison results of feature points and 3D spatial points with potential dynamic regions are obtained to identify dynamic suspicious points; Update the dynamic confidence level, perform optimization requirement verification on the dynamic suspicious points, determine and suppress dynamic suppression points, and obtain static feature points; Static feature points are pose optimized to obtain stable static features, and then 3D reconstruction is performed to obtain a high-quality map.

[0006] Optionally, acquiring images during camera movement to obtain visual images includes: Acquire all images captured during the camera's movement and arrange them according to time frames; The images are quality checked and images with insufficient quality are removed to obtain a visual image.

[0007] Optionally, the step of extracting features from the visual image to obtain a set of feature points includes: Obtain key image representation information within the visual image to obtain feature points; Obtain all feature points within each visual image to obtain a feature point set.

[0008] Optionally, the step of performing target detection on the visual image to obtain a set of potential dynamic regions includes: The visual image is segmented to obtain a segmented image; Information within the segmented image is obtained and compared with the established specific category meaning library; When the information in the segmented image matches any specific category meaning in the specific category meaning library, the segmented image is a potential dynamic region; Acquire potential dynamic regions within all visual images to construct a set of potential dynamic regions.

[0009] Optionally, the step of performing three-dimensional reconstruction of feature points to obtain three-dimensional spatial points includes: Obtain depth maps and fixed parameters from the camera to obtain 3D reconstruction parameters; Based on the 3D reconstruction parameters, feature points are back-projected from the 2D image coordinates to the 3D space to obtain 3D spatial points.

[0010] Optionally, the step of obtaining the comparison results of feature points and three-dimensional spatial points with potential dynamic regions to obtain dynamic suspicious points includes: The feature point is compared with the potential dynamic region in terms of position. If the feature point is in the potential dynamic region, the feature point is determined to be a static point. When a feature point is located in a potential dynamic region, its 3D spatial point after back-projection is obtained, and the reference depth of the 3D spatial point and the potential dynamic region is compared. If the judgment criteria are not met, the feature point is judged as a dynamic suspicious point. The judgment criteria are as follows: , in, z i Indicates the first i The depth value of a three-dimensional spatial point; z c Indicates the reference depth of the potential dynamic region; i Represents a sequence index of points in three-dimensional space; This indicates the threshold for the comparison result.

[0011] Optionally, the step of updating the dynamic confidence level, optimizing the requirements of the dynamic suspicious points, determining and suppressing dynamic suppression points, and obtaining static feature points includes: The dynamic confidence level is updated using the following equation: , in, Indicates the dynamic confidence level of the current frame; Indicates the historical weighting coefficient; S （t-1） Indicates the dynamic confidence level of the previous frame; t Indicates the image frame number; When the dynamic confidence level of the current frame is not lower than the dynamic decision threshold, the dynamic suspicious point is identified as a dynamic suppression point and suppressed. All dynamic suppression points are obtained and suppressed; the remaining feature points are the static feature points.

[0012] Optionally, the step of optimizing the pose of static feature points to obtain stable static features and performing 3D reconstruction to obtain a high-quality map includes: Based on all static feature points, pose optimization is performed to obtain stable static features; The stable static features are reconstructed in three dimensions to obtain a high-quality map.

[0013] The second aspect: A visual SLAM system for dynamic environments, used to execute the visual SLAM method for dynamic environments as described in the first aspect, the system comprising: The image acquisition module is used to acquire visual images; The feature extraction module is connected to the image acquisition module and is used to acquire feature points and feature point sets in visual images; The semantic detection module is connected to the feature extraction module to obtain potential dynamic regions; The 3D reconstruction module is connected to the semantic detection module to acquire 3D spatial points; The dynamic feature discrimination module is connected to the 3D reconstruction module and is used to acquire dynamic suspicious points; The dynamic confidence update module, connected to the dynamic feature discrimination module, is used to update the dynamic confidence and obtain static feature points; The pose estimation module is connected to the dynamic confidence update module to obtain stable static features; The map building module, connected to the pose estimation module, is used to create high-quality maps.

[0014] The beneficial effects of this application include: 1. Improved localization stability. This application introduces "depth consistency discrimination" within the semantic candidate region, which further compares the consistency of the depth of feature points within the bounding box with the reference depth of the region, rather than relying solely on 2D geometric consistency. Depth consistency discrimination provides constraints at the 3D structural level: when dynamic objects move along the epipolar direction or the camera baseline is small, causing a decrease in the 2D geometric constraint discrimination capability, dynamic objects and the background are still more likely to show inconsistency in depth. Therefore, dynamic points are more likely to be identified as "dynamic suspicious points," and dynamic suspicious points are excluded from pose optimization and map updates more promptly and accurately. This allows the optimization residuals to be contributed more by the static background structure, reducing the interference of non-rigid motion on optimization. Short-term jitter and drift in camera pose estimation are reduced, the trajectory is smoother, and the system tracking stability is improved, especially under geometric degradation conditions, making it less prone to losing track.

[0015] 2. Improve map integrity. This application uses semantic detection results only as "candidate screening" and distinguishes between "dynamic object points" and "static background points within the bounding box" through depth consistency verification within the candidate region. For feature points that fall within the detection box but belong to the background structure, as long as their depth is consistent with the region's reference depth, they will be judged as "static background points" and retained, rather than being directly eliminated because they are located within the detection box or due to fluctuations in single-frame geometric errors. With more static structure points retained, a more stable feature association can be maintained between local mapping and keyframes, making map point triangulation and subsequent observation accumulation more sufficient, increasing the number and proportion of effective map points, making the map structure more complete and reducing sparsity, providing more stable constraints for backend optimization, and improving the overall mapping quality.

[0016] 3. Improve system stability. This application establishes a dynamic confidence score S_t for map points and integrates historical and current observations through exponential sliding accumulation to achieve "gradual suppression." Occasional misjudgments in a single frame (caused by noise, occlusion, or matching anomalies) will only cause short-term fluctuations in S_t and will not immediately trigger removal. Only when dynamic evidence appears continuously in multiple frames will S_t gradually increase and exceed the threshold η, thereby triggering suppression. Since the dynamic point processing is changed from "single-frame binary decision" to "multi-frame statistical decision," the point retention / removal behavior is smoother, reducing the randomness of accidentally deleting static points and intermittently retaining dynamic points. The system's tracking results are more stable in complex dynamic scenes, pose estimation fluctuations are smaller, map updates are more continuous, and the overall operational stability and robustness are improved.

[0017] 4. Improving robustness while maintaining real-time performance. This application adopts a "three-layer structure": first, semantic candidates are used to narrow down the scope; then, depth consistency is performed within the candidate scope; and finally, multi-frame statistics are used to determine whether to suppress. Depth consistency verification only applies to points within the detection box, not all points in the map, reducing the amount of additional discrimination computation. Multi-frame confidence updates only maintain a scalar score for existing map points, resulting in low update overhead. Without significantly increasing the complexity of front-end processing, the system obtains more reliable dynamic point discrimination capabilities and a smoother suppression strategy. Compared with solutions that require pixel-level segmentation or complex reconstruction, this application is easier to implement in engineering, has the potential for real-time applications, and can significantly improve localization and mapping performance in dynamic environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the embodiments of this application or the prior art will be briefly introduced below. Obviously, the following description is only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings are used to provide a further understanding of this disclosure and constitute a part of the specification. They are used together with the following detailed description to explain this disclosure, but do not constitute a limitation of this disclosure. In the drawings: Figure 1 A flowchart of a visual SLAM method in a dynamic environment is provided as an embodiment of this application; Figure 2 This is a schematic diagram illustrating the dynamic confidence multi-frame accumulation and suppression triggering in a visual SLAM method under dynamic conditions, provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, in the embodiments of this application, "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0020] Existing visual SLAM systems are prone to misusing dynamic features for pose estimation and mapping when moving objects such as pedestrians and vehicles are present, leading to unstable localization and degraded map quality. Current methods either completely eliminate semantic target regions and inadvertently delete static features, or rely solely on geometric constraints, making it difficult to identify dynamic points under specific motion degradation conditions. To address this, this application proposes a dynamic feature suppression scheme combining depth-semantic consistency discrimination and multi-frame temporal statistics: first, semantic detection is used to identify potential dynamic regions; then, RGB-D depth information is used to perform point-level consistency verification of feature points within the region; and finally, progressive suppression of dynamic features is achieved through multi-frame dynamic confidence accumulation. This method can reduce dynamic interference and retain more stable static features in dynamic scenes, thereby improving the localization robustness and map construction quality of the SLAM system.

[0021] like Figure 1 The above is a flowchart of a visual SLAM method in a dynamic environment provided by an embodiment of this application. Specifically: S110. Acquire images captured during camera movement to obtain visual images.

[0022] S120. Perform feature extraction on the visual image to obtain a set of feature points.

[0023] S130. Perform target detection on the visual image to obtain a set of potential dynamic regions.

[0024] S140. Perform three-dimensional reconstruction on the feature points to obtain three-dimensional spatial points.

[0025] S150. Obtain the comparison results of feature points and three-dimensional spatial points with potential dynamic regions to obtain dynamic suspicious points.

[0026] S160. Update the dynamic confidence level, perform optimization requirement verification on the dynamic suspicious points, determine and suppress the dynamic suppression points, and obtain static feature points.

[0027] S170. Optimize the pose of static feature points to obtain stable static features, and perform 3D reconstruction to obtain a high-quality map.

[0028] The purpose of the above steps is to remove dynamic information from the scene and then build a map to obtain a high-quality map.

[0029] The following will provide a detailed explanation of all the steps above: As described in step S110, the purpose of this step is to acquire and preprocess all the obtained images to ensure that the final images carry the feature data that needs to be obtained during map construction, thereby ensuring the accuracy of the source information. Specifically: Acquire all images captured during the camera's movement and arrange them according to time frames; The images are quality checked and images with insufficient quality are removed to obtain a visual image.

[0030] The process of acquiring visual images is based on the image acquisition module, which is used to acquire RGB images and corresponding depth images, providing a three-dimensional information basis for subsequent feature reconstruction.

[0031] The image quality inspection includes various detection measures such as defects and blurring. If a defect that cannot pass the inspection is found, the image is removed.

[0032] In this process, the internal parameters of the camera need to be determined before subsequent image information processing can be performed based on these parameters.

[0033] As described in step S120, the purpose of this step is to process feature points in the image, thereby enabling subsequent map construction based on these feature points. Specifically: Obtain key image representation information within the visual image to obtain feature points; Obtain all feature points within each visual image to obtain a feature point set.

[0034] In the acquisition of feature points and feature point bases, a feature extraction module is used. This module extracts key points and calculates descriptors in the image to obtain a set of feature points.

[0035] In the acquisition of feature points, key points with obvious texture or structural information are extracted from the image and used for matching between different image frames to estimate camera motion.

[0036] Specifically, for the obtained feature point set, a dataset is constructed according to the feature points in each visual image, thus obtaining the feature point set.

[0037] As described in step S130, the purpose of this step is to decompose and analyze the acquired image to identify potential dynamic regions, thereby determining areas of dynamic change and suppressing them. Specifically: The visual image is segmented to obtain a segmented image; Information within the segmented image is obtained and compared with the established specific category meaning library; When the information in the segmented image matches any specific category meaning in the specific category meaning library, the segmented image is a potential dynamic region; Acquire potential dynamic regions within all visual images to construct a set of potential dynamic regions.

[0038] Specifically, the object detection network identifies regions in the image that may belong to moving objects, such as pedestrians and vehicles, and outputs a set of corresponding detection bounding boxes. The output of this module is used as a candidate range for subsequent dynamic feature filtering.

[0039] In determining the potential dynamic region set, the semantic detection module performs object detection on the image and outputs the potential dynamic region set. This module only provides the "suspicious region range" and does not directly delete feature points.

[0040] In the operation of the semantic detection module, the delineation of potential dynamic regions can be achieved by using methods such as pixel-level semantic regions obtained based on semantic segmentation networks, potential dynamic regions obtained based on motion segmentation or foreground detection methods, candidate dynamic regions obtained based on optical flow field anomaly region detection, and candidate regions obtained based on historical trajectory unstable region statistics.

[0041] As described in step S140, the purpose of this step is to perform three-dimensional reconstruction of the obtained feature points, and based on the obtained three-dimensional reconstruction results, to perform subsequent accurate identification of dynamic points. Specifically: Obtain depth maps and fixed parameters from the camera to obtain 3D reconstruction parameters; Based on the 3D reconstruction parameters, feature points are back-projected from the 2D image coordinates to the 3D space to obtain 3D spatial points.

[0042] In this process, feature points are back-projected from two-dimensional image coordinates to three-dimensional space using depth maps and camera intrinsic parameters to obtain three-dimensional spatial points.

[0043] As described in step S150, the purpose of this step is to process the obtained feature points and three-dimensional spatial points separately to identify potential dynamic points in the image, thereby laying the foundation for subsequent dynamic point removal. Specifically: The feature point is compared with the potential dynamic region in terms of position. If the feature point is in the potential dynamic region, the feature point is determined to be a static point. When a feature point is located in a potential dynamic region, its 3D spatial point after back-projection is obtained, and the reference depth of the 3D spatial point and the potential dynamic region is compared. If the judgment criteria are not met, the feature point is judged as a dynamic suspicious point. The judgment criteria are as follows: , in, z i Indicates the first i The depth value of a three-dimensional spatial point; z c Indicates the reference depth of the potential dynamic region; i Represents a sequence index of points in three-dimensional space; This indicates the threshold for the comparison result.

[0044] If a feature point is not in the semantic candidate region (i.e., the potential dynamic region), it is directly judged as a static point. In this case, the feature points can be directly compared in the two-dimensional image. As long as they do not overlap, the feature point can be considered as a static point.

[0045] When a feature point is found to be within a semantic candidate region, a depth consistency verification is required. For feature points within the semantic region, their depth is compared with the region reference depth.

[0046] In this process, deep consistency verification is performed on all feature points within the semantic region.

[0047] Points that meet the criteria are considered static background points; otherwise, they are considered dynamic suspicious points.

[0048] Among them, depth consistency judgment can also be achieved by using the distance between the 3D point and the local planar model of the region as a consistency criterion, using whether the change amplitude of the 3D point position in multiple frames satisfies the rigid body assumption as a criterion, using the consistency of the point cloud neighborhood normal vector as a 3D structural consistency constraint, and using the depth gradient or depth change rate to determine the structural continuity.

[0049] As described in step S160, the purpose of this step is to suppress the existing dynamic feature points, which is equivalent to removing dynamic objects from the map scene to be built. Therefore, it is necessary to determine the static feature points. Obviously, this process requires determining the confidence level of suspected dynamic points, then judging whether the suspected point truly belongs to a dynamic point, and analyzing whether it needs to be removed. Specifically: The dynamic confidence level is updated using the following equation: , in, Indicates the dynamic confidence level of the current frame; Indicates the historical weighting coefficient; S （t-1） Indicates the dynamic confidence level of the previous frame; t Indicates the image frame number; When the dynamic confidence level of the current frame is not lower than the dynamic decision threshold, the dynamic suspicious point is identified as a dynamic suppression point and suppressed. All dynamic suppression points are obtained and suppressed; the remaining feature points are the static feature points.

[0050] Among them, dynamic confidence is maintained for each map point. S t This is used to characterize the probability that a point belongs to a dynamic point. The dynamic confidence score is updated using an exponential sliding cumulative method.

[0051] in The current frame is marked as a dynamic observation (1 indicates a dynamic observation, 0 indicates a static observation), and λ is the historical weight coefficient. S t Exceeding the dynamic decision threshold η At this time, the map point is marked as a dynamic suppression point and does not participate in the optimization during subsequent pose estimation and map construction.

[0052] Among them, dynamic confidence modeling can also be achieved through methods such as dynamic proportional statistics based on sliding time windows, dynamic probability estimation based on Bayesian updates, dynamic state estimation based on Kalman filtering, and threshold determination based on cumulative occurrence counts.

[0053] After discovering dynamic suspicious points, these suspicious points are suppressed, i.e. removed, and the remaining feature points are static points.

[0054] The process of suppressing dynamic points can be achieved by reducing the optimization weight of feature points with high dynamic confidence instead of directly eliminating them, prohibiting them from generating map points during the mapping stage, ignoring high dynamic confidence features in loop closure detection, and excluding their influence only during the key frame selection stage.

[0055] As described in step S170, the purpose of this step is to construct a specific map. Specifically: Based on all static feature points, pose optimization is performed to obtain stable static features; The stable static features are reconstructed in three dimensions to obtain a high-quality map.

[0056] In the pose estimation process, specific coordinates of each static point are obtained, and then specific pose information is acquired and optimized to obtain stable static features.

[0057] In the process of map construction, only stable static features are reconstructed into map points in three dimensions to form a high-quality map.

[0058] The methods described in all the above steps also require system support, such as... Figure 2 The diagram shown is a schematic of a visual SLAM system in a dynamic environment provided in an embodiment of this application. Specifically, the system includes: The image acquisition module is used to acquire visual images; The feature extraction module is connected to the image acquisition module and is used to acquire feature points and feature point sets in visual images; The semantic detection module is connected to the feature extraction module to obtain potential dynamic regions; The 3D reconstruction module is connected to the semantic detection module to acquire 3D spatial points; The dynamic feature discrimination module is connected to the 3D reconstruction module and is used to acquire dynamic suspicious points; The dynamic confidence update module, connected to the dynamic feature discrimination module, is used to update the dynamic confidence and obtain static feature points; The pose estimation module is connected to the dynamic confidence update module to obtain stable static features; The map building module, connected to the pose estimation module, is used to create high-quality maps.

[0059] The beneficial effects of this application include: 1. Improved localization stability. This application introduces "depth consistency discrimination" within the semantic candidate region, which further compares the consistency of the depth of feature points within the bounding box with the reference depth of the region, rather than relying solely on 2D geometric consistency. Depth consistency discrimination provides constraints at the 3D structural level: when dynamic objects move along the epipolar direction or the camera baseline is small, causing a decrease in the 2D geometric constraint discrimination capability, dynamic objects and the background are still more likely to show inconsistency in depth. Therefore, dynamic points are more likely to be identified as "dynamic suspicious points," and dynamic suspicious points are excluded from pose optimization and map updates more promptly and accurately. This allows the optimization residuals to be contributed more by the static background structure, reducing the interference of non-rigid motion on optimization. Short-term jitter and drift in camera pose estimation are reduced, the trajectory is smoother, and the system tracking stability is improved, especially under geometric degradation conditions, making it less prone to losing track.

[0060] 2. Improve map integrity. This application uses semantic detection results only as "candidate screening" and distinguishes between "dynamic object points" and "static background points within the bounding box" through depth consistency verification within the candidate region. For feature points that fall within the detection box but belong to the background structure, as long as their depth is consistent with the region's reference depth, they will be judged as "static background points" and retained, rather than being directly eliminated because they are located within the detection box or due to fluctuations in single-frame geometric errors. With more static structure points retained, a more stable feature association can be maintained between local mapping and keyframes, making map point triangulation and subsequent observation accumulation more sufficient, increasing the number and proportion of effective map points, making the map structure more complete and reducing sparsity, providing more stable constraints for backend optimization, and improving the overall mapping quality.

[0061] 3. Improve system stability. This application establishes a dynamic confidence score S_t for map points and integrates historical and current observations through exponential sliding accumulation to achieve "gradual suppression." Occasional misjudgments in a single frame (caused by noise, occlusion, or matching anomalies) will only cause short-term fluctuations in S_t and will not immediately trigger removal. Only when dynamic evidence appears continuously in multiple frames will S_t gradually increase and exceed the threshold η, thereby triggering suppression. Since the dynamic point processing is changed from "single-frame binary decision" to "multi-frame statistical decision," the point retention / removal behavior is smoother, reducing the randomness of accidentally deleting static points and intermittently retaining dynamic points. The system's tracking results are more stable in complex dynamic scenes, pose estimation fluctuations are smaller, map updates are more continuous, and the overall operational stability and robustness are improved.

[0062] 4. Improving robustness while maintaining real-time performance. This application adopts a "three-layer structure": first, semantic candidates are used to narrow down the scope; then, depth consistency is performed within the candidate scope; and finally, multi-frame statistics are used to determine whether to suppress. Depth consistency verification only applies to points within the detection box, not all points in the map, reducing the amount of additional discrimination computation. Multi-frame confidence updates only maintain a scalar score for existing map points, resulting in low update overhead. Without significantly increasing the complexity of front-end processing, the system obtains more reliable dynamic point discrimination capabilities and a smoother suppression strategy. Compared with solutions that require pixel-level segmentation or complex reconstruction, this application is easier to implement in engineering, has the potential for real-time applications, and can significantly improve localization and mapping performance in dynamic environments.

[0063] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to computer program instructions. The aforementioned computer program can be stored in a non-volatile storage medium, and when executed, it performs the steps of the above method embodiments. Alternatively, if the integrated units of this application are implemented as software functional modules and sold or used as independent products, they can also be stored in a non-volatile storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a non-volatile storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, network device, etc.) to execute all or part of the methods described in the various embodiments of this application.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A visual SLAM method for dynamic environments, characterized in that, The method includes: Acquire image and depth information captured during camera movement to obtain visual images and distance values ​​for each pixel; Feature extraction is performed on the visual image to obtain a set of feature points; Target detection is performed on the visual image to obtain a set of potential dynamic regions; The feature points are reconstructed in three dimensions to obtain three-dimensional spatial points; The comparison results of feature points and 3D spatial points with potential dynamic regions are obtained to identify dynamic suspicious points; Update the dynamic confidence level, perform optimization requirement verification on the dynamic suspicious points, determine and suppress dynamic suppression points, and obtain static feature points; Static feature points are pose optimized to obtain stable static features, and then 3D reconstruction is performed to obtain a high-quality map.

2. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The acquisition of image and depth information obtained during camera movement to obtain the visual image and distance value of each pixel includes: Acquire all images captured during the camera's movement and arrange them according to time frames; The image quality is checked, and images with insufficient quality are removed to obtain a visual image. The distance value of each pixel in the visual image is then obtained. Acquire depth information within the visual image during camera movement.

3. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The step of extracting features from the visual image to obtain a set of feature points includes: Obtain key image representation information within the visual image to obtain feature points; Obtain all feature points within each visual image to obtain a feature point set.

4. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The step of performing target detection on the visual image to obtain a set of potential dynamic regions includes: The visual image is segmented to obtain a segmented image; Information within the segmented image is obtained and compared with the established specific category meaning library; When the information in the segmented image matches any specific category meaning in the specific category meaning library, the segmented image is a potential dynamic region; Acquire potential dynamic regions within all visual images to construct a set of potential dynamic regions.

5. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The step of performing three-dimensional reconstruction of feature points to obtain three-dimensional spatial points includes: Obtain depth maps and fixed parameters from the camera to obtain 3D reconstruction parameters; Based on the 3D reconstruction parameters, feature points are back-projected from the 2D image coordinates to the 3D space to obtain 3D spatial points.

6. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The comparison results of acquiring feature points and three-dimensional spatial points with potential dynamic regions to obtain dynamic suspicious points include: The feature point is compared with the potential dynamic region in terms of position. If the feature point is in the potential dynamic region, the feature point is determined to be a static point. When a feature point is located in a potential dynamic region, its 3D spatial point after back-projection is obtained, and the reference depth of the 3D spatial point and the potential dynamic region is compared. If the judgment criteria are not met, the feature point is judged as a dynamic suspicious point. The judgment criteria are as follows: ， Among them, z i z represents the depth value of the i-th point in three-dimensional space; c The reference depth of the potential dynamic region is represented; i represents the sequence index of the three-dimensional spatial point; This indicates the threshold for the comparison result.

7. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The process of updating the dynamic confidence level involves verifying the optimization requirements for the dynamically suspicious points, determining and suppressing dynamic suppression points, and obtaining static feature points, including: Binarize point k at time t. Represents 1 or 0: ， in, This represents the binarization result of point k at time t; k represents the point label; t represents the time. The dynamic confidence level is updated using the following equation: ， in, This represents the dynamic confidence level of point k at time t; Indicates the historical weighting coefficient; The value represents the dynamic confidence level of the previous frame; t represents a specific moment in time. When the dynamic confidence level of the current frame is not lower than the dynamic decision threshold, the dynamic suspicious point is identified as a dynamic suppression point and suppressed. All dynamic suppression points are obtained and suppressed; the remaining feature points are the static feature points.

8. The visual SLAM method in a dynamic environment according to claim 1, characterized in that, The process of optimizing the pose of static feature points to obtain stable static features and performing 3D reconstruction to obtain a high-quality map includes: Based on all static feature points, pose optimization is performed to obtain stable static features; The stable static features are reconstructed in three dimensions to obtain a high-quality map.

9. A visual SLAM system for dynamic environments, used to execute a visual SLAM method for dynamic environments as described in any one of claims 1 to 8, characterized in that, The system includes: The image acquisition module is used to acquire visual images; The feature extraction module is connected to the image acquisition module and is used to acquire feature points and feature point sets in visual images; The semantic detection module is connected to the feature extraction module to obtain potential dynamic regions; The 3D reconstruction module is connected to the semantic detection module to acquire 3D spatial points; The dynamic feature discrimination module is connected to the 3D reconstruction module and is used to acquire dynamic suspicious points; The dynamic confidence update module, connected to the dynamic feature discrimination module, is used to update the dynamic confidence and obtain static feature points; The pose estimation module is connected to the dynamic confidence update module to obtain stable static features; The map building module, connected to the pose estimation module, is used to create high-quality maps.

Images