A Method and System for 3D Reconstruction of Indoor Environments Based on Reflectivity Edge Constraints

Abstract

This paper presents a method and system for 3D reconstruction of indoor environments based on reflectivity edge constraints, belonging to the field of image modeling. To address the problem of reducing odometry errors and improving 3D reconstruction accuracy in degraded indoor environments, the method performs point-line matching of geometric edge points and reflectivity edge points in the current frame with their corresponding local maps, and point-plane matching of geometric plane points with their corresponding local maps. Residuals of geometric edge points and reflectivity edge points are obtained based on the point-line matching; residuals of geometric plane points are obtained based on the point-plane matching; the residuals are iteratively optimized to obtain the optimal pose transformation; and based on the optimal pose transformation, the geometric edge feature points, geometric plane feature points, and reflectivity edge feature points of the current frame are updated to the global coordinate system through rotation and translation transformations, completing the 3D reconstruction of the current frame. The effect is that a new constraint is added during the pose optimization process, which can reduce odometry errors and improve the accuracy of 3D reconstruction to a certain extent.

Description

Technical Field

[0001] This invention belongs to the field of image modeling and relates to a method and system for high-precision three-dimensional reconstruction of indoor environment using a three-dimensional lidar detection system. Background Technology

[0002] With the development of mobile robots, perception and autonomous localization technologies for unknown environments are becoming increasingly mature. LiDAR, due to its advantages such as high resolution, strong resistance to active interference, unaffected by lighting conditions, small size, and light weight, is widely used in problem scenarios involving perception, identification, and localization.

[0003] Currently, the main approach to this problem is laser SLAM (Simultaneous Localization and Mapping) algorithms, with LOAM as the classic framework. At the forefront of research are laser SLAM algorithms coupled with IMUs, which are mainly divided into loosely coupled and tightly coupled methods. Loosely coupled methods process the two sensors separately to infer their motion constraints and then fuse them (the IMU does not participate in the optimization process). Examples include LOAM and LeGO-LOAM, which are classic loosely coupled methods. Tightly coupled methods, on the other hand, directly fuse LiDAR and inertial measurement through joint optimization, such as LIO-SAM. Generally, tightly coupled algorithms achieve higher localization and mapping accuracy than loosely coupled algorithms.

[0004] Current classic laser SLAM algorithms largely follow the point cloud registration method of LOAM, which focuses on geometric edges and planar features. This involves extracting geometric edge and planar feature points based on the smoothness (curvature) of the local point cloud and finding the optimal pose transformation through iterative optimization to achieve high-precision odometry. While this significantly improves real-time performance compared to matching methods based on the original point cloud (e.g., ICP, Iterative Closest Points), laser SLAM algorithms based solely on geometric features do not offer high accuracy for reconstructing degraded environments such as indoor corridors. Although some methods utilize intensity information to improve the accuracy of 3D reconstruction of degraded environments, intensity data is affected by numerous parameters such as surface reflectivity, incident angle, and distance, generally requiring complex calibration before use. LiDAR reflectivity data indicates the reflectivity of the target object; different materials have different reflectivities on their surfaces. It does not require complex calibration before use and can be used to extract boundaries between different materials. This parameter is often ignored in many current laser SLAM algorithms. Summary of the Invention

[0005] To address the problem of reducing odometer errors and improving 3D reconstruction accuracy in degraded indoor environments, this application provides a 3D reconstruction method for indoor environments based on reflectivity edge constraints, including...

[0006] The data acquisition device scans the scene to obtain point cloud data;

[0007] Features are extracted from points in each frame of point cloud data. These features include coordinate geometric edge points, geometric plane points, and reflectivity edge points.

[0008] A local map is constructed based on the coordinate geometric edge points, geometric plane points, and reflectivity edge points in historical frames within a certain spatiotemporal range.

[0009] Perform point-line matching between geometric edge points and reflectivity edge points in the current frame and the corresponding local map; perform point-plane matching between geometric plane points and the corresponding local map.

[0010] The residuals of geometric edge points and reflectivity edge points are obtained based on the point-line matching; the residuals of geometric plane points are obtained based on the point-plane matching.

[0011] The residual is iteratively optimized to obtain the optimal pose transformation;

[0012] Based on the optimal pose transformation, the geometric edge feature points, geometric plane feature points, and reflectivity edge feature points of the current frame are updated to the global coordinate system through rotation and translation transformation, thus completing the three-dimensional reconstruction of the current frame.

[0013] According to some embodiments of the present application, the method for 3D reconstruction of indoor environments based on reflectivity edge constraints includes the following extracted features:

[0014] S2.1. Obtain the geometric curvature c geo and reflectivity curvature c ref ,in:

[0015]

[0016]

[0017] Among them, c geo Represents geometric curvature, c ref Let L represent the reflectivity curvature, L represent the radar system, i represent a point in frame k, j represent a point in set S, and S represent a set of points adjacent to i on the same laser scanning line. This represents the distance from point i to the origin of the lidar. This represents the distance from point j to the origin of the lidar. Represents the reflectivity of point i. This represents the reflectivity of point j;

[0018] S2.2. Based on the geometric curvature c geo and reflectivity curvature c ref Obtain the feature type of the point, where: if the geometric curvature c geo If the curvature c is greater than the threshold, it indicates that the point has geometric edge features; if the curvature c geo If the reflectance curvature is less than the threshold, it indicates that the point has geometric planar characteristics; if the reflectance curvature c ref When the value is greater than the threshold, it indicates that the point has reflectivity edge characteristics.

[0019] The indoor environment 3D reconstruction method based on reflectivity edge constraints according to some embodiments of this application further includes distortion removal of each frame of the point cloud data. Distortion removal of the point cloud data is required before extracting features from the points in each frame of the point cloud data.

[0020] According to some embodiments of this application, the indoor environment three-dimensional reconstruction method based on reflectivity edge constraints, the point cloud data is obtained by scanning the environment with a three-dimensional mechanical lidar, and the inertial navigation system obtains real-time pose information, including angular velocity and linear acceleration.

[0021] According to some embodiments of this application, the method for 3D reconstruction of indoor environments based on reflectivity edge constraints, local map

[0022]

[0023] Where i represents the i-th frame, M i This represents a local map composed of selected frames from the i-th frame and the preceding keyframes. These represent local maps of geometric edge features, local maps of geometric plane features, and local maps of reflectivity edge features, respectively.

[0024] According to some embodiments of the present application, the indoor environment 3D reconstruction method based on reflectivity edge constraints uses lines and surfaces for point-line matching and point-surface matching, which are obtained by KDtree nearest neighbor search method and searched in the corresponding local map. The validity of the extracted lines is determined by the covariance matrix of the nearest points, and the validity of the fitted surfaces is determined by whether the distance from the nearest points to the fitted plane is less than a threshold.

[0025] According to some embodiments of the present application, the indoor environment 3D reconstruction method based on reflectivity edge constraints obtains the residuals of geometric edge points and reflectivity edge points based on the point-line matching:

[0026]

[0027] Based on point-plane matching, obtain the residuals of geometric plane points:

[0028]

[0029] Where i represents the frame index, d ε d represents the residual at the geometric edge point. γ d represents the residual at the edge of reflectivity. H The residuals of geometric plane points are represented by the geometric edge points, geometric plane points, and reflectivity edge points in frame i+1 under radar system L. These points are initially transformed to the global coordinate system through robot pose estimation using inertial navigation. The transformed points are represented by X. (i+1,k) Let k represent the index of the corresponding point in the three feature point sets, and let M represent the local map corresponding to the geometric edge point and the reflectivity edge point. i e and M i r The straight line is represented by two points X along the direction of the largest eigenvector corresponding to the covariance matrix. (i,j) and X (i,l) This indicates that j and l are the identifiers of the two points representing the lines, and the geometric plane points correspond to the local map. The plane is formed by non-collinear X on the fitted plane. (i,u) X (i,v) and X (i,w) Three points are determined: u, v, and w are the identifiers of the three non-collinear points that determine the fitted plane.

[0030] According to some embodiments of the present application, the method for 3D reconstruction of indoor environments based on reflectivity edge constraints iteratively optimizes the residual, including:

[0031] Find the Jacobian matrix J of the residual pair transformation matrix, and then perform Gaussian-Newton iterative optimization:

[0032] in

[0033] J T ·J·ΔX=-J T ·f

[0034] Among them, T i+1 Represents the transformation matrix. Let f represent the three feature points located in the i+1 frame under radar system L, ΔX represent the pose correction amount, and f represent the residual.

[0035] Minimize the residuals to obtain the optimal pose transformation:

[0036]

[0037] Here, ΔT is the final pose transformation that needs to be solved.

[0038] The present invention also provides a three-dimensional reconstruction system for an indoor environment based on reflectivity edge constraints. The system includes: one or more processors, a memory, and one or more programs; wherein the one or more programs are stored in the memory, and the one or more programs include instructions that, when executed by the system, cause the system to perform any of the methods described above.

[0039] The beneficial effects of this invention: Existing 3D reconstruction methods using geometric coordinates can generally only extract geometric edge features such as corners, but cannot extract the boundaries of different materials on the same plane. In degraded environments such as indoor corridors, relying solely on geometric features may lead to significant deviations between the reconstructed results and the real-world scene. This invention adds reflectivity edge constraints to existing 3D reconstruction methods that only rely on geometric feature constraints. This new constraint, added during pose optimization, can reduce odometry errors and improve the accuracy of 3D reconstruction to some extent. Attached Figure Description

[0040] Figure 1 The method flowchart of the present invention.

[0041] Figure 2 A physical image of a 3D lidar detection system.

[0042] Figure 3 Results of geometric edge and reflectivity edge extraction.

[0043] Figure 4 Figure 3 The corresponding real-world scenarios.

[0044] Figure 5 Mapping results.

[0045] Figure 6 Comparison of odometers on the xyz axes with and without reflectivity edge constraints. Detailed Implementation

[0046] The embodiments of this application are described in detail below with reference to the accompanying drawings, examples of which are shown in the drawings. Throughout the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The methods and systems are based on the same technical concept. Since the principles by which the methods and systems solve problems are similar, the implementations of the systems and methods can refer to each other, and repeated details will not be repeated.

[0047] 3D point clouds can describe 3D models. Point clouds offer an intuitive representation and are characterized by their randomness and irregularity. A point cloud is a massive collection of points representing the surface characteristics of a target, obtained by acquiring the spatial coordinates of each sampled point on the target surface. The position of each point in the point cloud is described by a set of Cartesian coordinates (X, Y, Z).

[0048] In one embodiment, this invention proposes a 3D reconstruction method for indoor environments based on reflectivity edge constraints. In this method, a 3D mechanical lidar is used to scan the surrounding environment 360° to generate point cloud data. An inertial navigation system (IMU) provides real-time pose information, such as angular velocity and linear acceleration. The information provided by these two sensors is used to reconstruct the 3D indoor environment. The specific steps of the reconstruction method are as follows:

[0049] Step 1: Laser point cloud preprocessing

[0050] (1) Distortion Reduction: The points in a frame of point cloud are collected at different times, and the origin of the coordinate system corresponding to each point is the radar pose at the time of collection. When such point clouds are stitched together, they will not match the actual shape of the object. Motion compensation is performed using the IMU pose information at the start and end times of the frame to transform all points in the frame to the coordinate system corresponding to the first point.

[0051] (2) Feature Extraction: Coordinate geometric feature extraction and reflectance edge extraction. Curvature is used to distinguish geometric edge points from geometric plane points, and reflectance edge features are extracted. The specific calculation method is as follows:

[0052]

[0053]

[0054] (3) Where, c geo and c ref Let L represent the geometric curvature and reflectivity curvature respectively, L represent the radar system, i represent a point in frame k, j represent a point in set S, and S represent a set of several points adjacent to i on the same laser scanning line. and These represent the distances from point i and point j to the origin of the lidar, respectively. and Let c represent the reflectance of points i and j, respectively. When the points are geometrically planar points such as smooth walls, c geo The value is relatively small when the point is a geometric edge point such as a corner. geo Larger. For c ref It is used to identify objects with a different reflectivity than the wall surface, such as display boards hanging on the wall. The boundary between the object and the wall is precisely where c is located. ref Larger areas. For three types of feature points, three thresholds are set: when the geometric curvature and reflectivity curvature are greater than the corresponding thresholds, they are identified as edge features; when the geometric curvature is less than the corresponding thresholds, they are identified as geometric planar features.

[0055] Step 2: Point Cloud Registration

[0056] (1) Generating a local map: Since the point cloud registration method of the frame image is adopted, a local map corresponding to the current frame needs to be generated before registration. This local map is composed of the three feature points in step (2) of step one. These feature points come from key frames within a certain spatiotemporal range of the i-th frame, that is...

[0057]

[0058] Where i represents the i-th frame, M i This represents a local map composed of selected frames from the i-th frame and the preceding keyframes. These represent the local maps of geometric edge features, geometric plane features, and reflectivity edge features, respectively. It can be understood that this step constructs a local map based on the coordinates of geometric edge points, geometric plane points, and reflectivity edge points in historical frames within a certain spatiotemporal range. Here, the certain spatiotemporal range refers to: for the most recent keyframe, searching for adjacent historical keyframes within a certain radius (searching for historical keyframes within a radius of x meters centered on the current location), and historical keyframes whose timestamps are less than a certain threshold from the current frame (selecting frames whose timestamp difference from the current frame is less than t seconds).

[0059] (2) Matching: This step mainly involves matching the current frame with the local map obtained in the previous step. Specifically, geometric edge points and reflectivity edge points in the current frame are matched with the geometric edge feature local map and reflectivity edge feature local map, respectively, using point-to-line matching. Geometric plane points are matched with the geometric plane feature local map using point-to-plane matching. KD-tree is used for nearest neighbor filtering, typically searching for 5 nearby points. In the point-to-line matching stage, the covariance matrix is ​​used to determine if the 5 points form a line. If the largest eigenvalue of the covariance matrix is ​​significantly larger than the second largest eigenvalue, then the 5 points are considered to form a line, and the direction of the eigenvector corresponding to the largest eigenvalue is the direction of the formed line. In the geometric plane point matching stage, we simply fit a plane using the found 5 points and determine whether the fitted plane is usable by checking if the distances from all 5 points to the fitted plane are less than a threshold.

[0060] After finding the line and the surface using the nearest neighbor algorithm of the KD-tree, the next step is to find the distances from the point to the line and from the point to the surface, which are used as residuals for the next iterative optimization. The distances (residuals) between geometric edge points, reflectivity edge points, and geometric plane points and their corresponding feature local maps are calculated as follows:

[0061]

[0062]

[0063] Where i represents the frame index, d ε d γ and d H These correspond to the residuals of geometric edge points, reflectivity edge points, and geometric plane points, respectively. The geometric edge points, geometric plane points, and reflectivity edge points (all in the radar coordinate system) in frame i+1 are initially transformed to the global coordinate system through robot pose transformation estimated by inertial navigation. The transformed points are represented by X... (i+1,k) Let k represent the index of the corresponding point in the three feature point sets. Geometric edge points and reflectivity edge points correspond to local maps. and The straight line is represented by two points X along the direction of the largest eigenvector corresponding to the covariance matrix. (i,j) and X (i,l) This indicates that j and l are the identifiers of the two points representing the lines, respectively. Geometric plane points correspond to local maps. The plane is formed by non-collinear X on the fitted plane. (i,u) X (i,v) and X (i,w) Three points are determined: u, v, and w are the identifiers of the three non-collinear points that determine the fitted plane.

[0064] (3) Iterative optimization: This step mainly involves minimizing the residuals to obtain the optimal pose transformation, i.e.:

[0065]

[0066] Here, ΔT is the final pose transformation that needs to be solved.

[0067] The iterative optimization process roughly involves finding the Jacobian matrix J of the residual pair transformation matrix, and then performing Gaussian-Newton iterative optimization.

[0068] in

[0069] J T ·J·ΔX=-J T ·f

[0070] Among them, T i+1 Represents the transformation matrix. Let represent the three feature points in frame i+1 (located in radar system L), ΔX represent the pose correction amount, and f represent the residual. The optimal pose transformation is obtained after iteration convergence.

[0071] Finally, using the optimized pose, the geometric edge feature points, geometric plane feature points, and reflectivity edge feature points of the current frame are updated to the global coordinate system through rotation and translation transformations to complete the 3D reconstruction.

[0072] This invention targets indoor environments and adds reflectivity edge constraints to existing 3D reconstruction methods that rely solely on geometric feature constraints. Geometric coordinates can generally only extract geometric edge features such as corners, but cannot extract boundaries between different materials on the same plane. In degraded environments such as indoor corridors, relying solely on geometric features may lead to significant deviations between the reconstruction results and the real-world scene. This invention introduces reflectivity edge extraction into the original feature extraction method, adding a new constraint during pose optimization, thus improving the accuracy of 3D reconstruction to a certain extent.

[0073] In a more specific instance, such as Figure 1 As shown, this invention proposes a method for three-dimensional reconstruction of indoor corridor environments. It improves the feature extraction module and iterative optimization module of the traditional laser SLAM algorithm. In addition to traditional geometric edges and planar features, it proposes to perform smoothness judgment processing on the reflectivity data of the laser radar, extract reflectivity edge information, and add the reflectivity edge features to the iterative optimization module to complete the high-precision three-dimensional reconstruction of the indoor environment.

[0074] The method includes the following steps:

[0075] Step 1: Collect data on the indoor corridor environment and preprocess the point cloud.

[0076] Data acquisition devices such as Figure 2 As shown. The main sensors are a 32-line mechanical LiDAR and an Inertial Measurement Unit (IMU). The LiDAR scans the current scene in a 360° rotation to collect point cloud data. For the raw point cloud, since the points in each frame are collected at different times and the corresponding coordinate system origins are also different, motion compensation, i.e., distortion removal, is first required using the IMU pose information at the start and end times of each frame.

[0077] Next, feature extraction is performed on the distortion-corrected point cloud, including traditional coordinate geometric feature extraction and reflectance edge extraction. Curvature is used to distinguish geometric edge points from geometric plane points, and reflectance edge features are extracted. The specific calculation method is as follows:

[0078]

[0079]

[0080] Among them, c geo and c ref Let L represent the geometric curvature and reflectivity curvature respectively, L represent the radar system, i represent a point in frame k, j represent a point in set S, and S represent a set of several points adjacent to i on the same laser scanning line. and These represent the distances from point i and point j to the origin of the lidar, respectively. and Let i and j represent the reflectance of points i and j, respectively. Three thresholds are set for the three types of feature points: points with a geometric curvature greater than 1.0 are identified as geometric edge points, and points with a curvature less than 0.1 are identified as geometric plane points. The original reflectance data range is 0-255. After normalization, points with a reflectance curvature greater than 0.4 are set as reflectance edge points.

[0081] Figure 3 This is the result of extracting the geometric edges and reflectivity edges of an indoor corridor scene. Spheres represent geometric edge points, and cubes represent reflectivity edge points. Figure 4 It corresponds Figure 3 The images show actual scene photos. It can be seen that the geometric edge extraction method performs relatively well in areas with abrupt geometric changes, such as door frames and wall corners, but it cannot identify the boundaries of different materials. For example... Figure 4 The boundary between the signboard and the plaster wall is beyond the capabilities of geometric edge extraction methods, but reflectivity data can accurately extract this type of boundary.

[0082] Step 2: Point Cloud Registration

[0083] In the previous step, three different feature points were identified and stored separately. Before registration, a local map needs to be constructed using these three feature points from historical frames within a certain spatiotemporal range for later point cloud registration.

[0084] Then, geometric edge points and reflectivity edge points in the current frame are matched with the corresponding local maps using point-line matching, and geometric plane points are matched with the corresponding geometric plane feature local maps using point-plane matching. The lines and planes used for matching are searched in the corresponding local maps using the KDtree nearest neighbor search method. The validity of the extracted lines is determined by the covariance matrix of the nearest points, and the validity of the fitted planes is determined by whether the distance from the nearest points to the fitted plane is less than a threshold.

[0085] The final optimization focuses on the residuals from points to lines and from points to surfaces, which are essentially the distances. The calculation formula is as follows:

[0086]

[0087]

[0088] Where, d ε d γ and d H These correspond to the residuals of geometric edge points, reflectivity edge points, and geometric plane points, respectively. The geometric edge points, geometric plane points, and reflectivity edge points (all in the radar coordinate system) in frame i+1 are initially transformed to the global coordinate system through robot pose transformation estimated by inertial navigation. The transformed points are represented by X... (i+1,k)Let k represent the index of the corresponding point in the three feature point sets. Geometric edge points and reflectivity edge points correspond to local maps. and The straight line is represented by two points X along the direction of the largest eigenvector corresponding to the covariance matrix. (i,j) and X (i,l) This indicates that j and l are the identifiers of the two points representing the lines, respectively. Geometric plane points correspond to local maps. The plane is formed by non-collinear X on the fitted plane. (i,u) X (i,v) and X (i,w) Three points are determined: u, v, and w are the identifiers of the three non-collinear points that determine the fitted plane.

[0089] The final step is to perform Gaussian-Newton iterative optimization on the residuals, that is, to minimize the geometric residuals and reflectivity residuals to estimate the final pose:

[0090]

[0091] Here, ΔT is the final pose transformation that needs to be solved.

[0092] The iterative optimization process roughly involves finding the Jacobian matrix J of the residual pair transformation matrix, and then performing Gaussian-Newton iterative optimization.

[0093] in

[0094] J T ·J·ΔX=-J T ·f

[0095] Among them, T i+1 Represents the transformation matrix. Let represent the three feature points in frame i+1 (located in radar system L), ΔX represent the pose correction amount, and f represent the residual. The optimal pose transformation is obtained after iteration convergence.

[0096] Using the optimized pose, the geometric edge feature points, geometric plane feature points, and reflectivity edge feature points of the current frame are updated to the global coordinate system through rotation and translation transformations to complete the 3D reconstruction. Figure 5 The results of the 3D reconstruction of the indoor corridor circular mirror are shown. Figure 6 This demonstrates a comparison of odometry on the x, y, and z axes with and without reflectivity edge constraints. The actual data acquisition method is handheld. Figure 2 The three-dimensional lidar detection system shown was used to collect data while walking at head height. The results showed that the odometer with reflectivity edge constraints had a smaller z-axis error compared to the one without reflectivity edge constraints.

[0097] This invention adds processing of laser reflectivity data in the feature extraction stage to extract reflectivity edge features, and adds constraints on reflectivity edge residuals in the iterative optimization stage. This is equivalent to adding a new constraint on the basis of geometric constraints. In degraded environments such as indoor corridors, compared with methods that rely solely on geometric features, it can reduce odometer errors and thus improve the accuracy of 3D reconstruction.

[0098] Based on the above, this invention addresses the problem of 3D reconstruction of indoor environments by proposing a laser-inertial navigation tightly coupled 3D reconstruction method incorporating reflectivity edge constraints. LiDAR reflectivity data indicates the reflectivity of a target object; different materials generally have different surface reflectivities, and the boundaries between materials with different reflectivities cannot be extracted using traditional coordinate geometry methods. In the feature extraction stage of laser point clouds, in addition to traditional methods using coordinate geometry information, this invention proposes to perform edge feature extraction on the reflectivity information in the laser point clouds. Furthermore, during frame matching, the loss residuals corresponding to the reflectivity edge feature points are incorporated into the iterative optimization process, thereby improving the accuracy of 3D reconstruction in indoor scenes.

[0099] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0100] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints, characterized in that, include The data acquisition device scans the scene to obtain point cloud data; Features are extracted from points in each frame of point cloud data. These features include coordinate geometric edge points, geometric plane points, and reflectivity edge points. A local map is constructed based on the coordinate geometric edge points, geometric plane points, and reflectivity edge points in historical frames within a certain spatiotemporal range. Perform point-line matching between geometric edge points and reflectivity edge points in the current frame and the corresponding local map; perform point-plane matching between geometric plane points and the corresponding local map. The residuals of geometric edge points and reflectivity edge points are obtained based on the point-line matching; the residuals of geometric plane points are obtained based on the point-plane matching. The residual is iteratively optimized to obtain the optimal pose transformation; Based on the optimal pose transformation, the geometric edge feature points, geometric plane feature points, and reflectivity edge feature points of the current frame are updated to the global coordinate system through rotation and translation transformation, thus completing the three-dimensional reconstruction of the current frame.

2. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 1, characterized in that, The extracted features include S2.

1. Obtain the geometric curvature c geo and reflectivity curvature c ref ,in: Among them, c geo Represents geometric curvature, c ref Let L represent the reflectivity curvature, L represent the radar system, i represent a point in frame k, j represent a point in set S, and S represent a set of points adjacent to i on the same laser scanning line. This represents the distance from point i to the origin of the lidar. This represents the distance from point j to the origin of the lidar. Represents the reflectivity of point i. This represents the reflectivity of point j; S2.

2. Based on the geometric curvature c geo and reflectivity curvature c ref Obtain the feature type of the point, where: if the geometric curvature c geo If the curvature c is greater than the threshold, it indicates that the point has geometric edge features; if the curvature c geo If the reflectance curvature is less than the threshold, it indicates that the point has geometric planar characteristics; if the reflectance curvature c ref When the value is greater than the threshold, it indicates that the point has reflectivity edge characteristics.

3. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 2, characterized in that, It also includes distortion removal for each frame of the point cloud data. Distortion removal of the point cloud data is required before extracting features from the points in each frame of the point cloud data.

4. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 2, characterized in that, The point cloud data is obtained by scanning the environment with a three-dimensional mechanical lidar, and the inertial navigation system acquires real-time pose information, including angular velocity and linear acceleration.

5. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 2, characterized in that, Local map Where i represents the i-th frame, M i This represents a local map composed of selected frames from the i-th frame and the preceding keyframes. These represent local maps of geometric edge features, local maps of geometric plane features, and local maps of reflectivity edge features, respectively.

6. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 2, characterized in that, The lines and surfaces used for point-line matching and point-surface matching are obtained through the KDtree nearest neighbor search method, which searches in the corresponding local map. The validity of the extracted lines is determined by the covariance matrix of the nearest points, and the validity of the fitted surfaces is determined by whether the distance from the nearest points to the fitted plane is less than a threshold.

7. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 6, characterized in that, Based on the point-line matching, obtain the residuals of the geometric edge points and reflectivity edge points: Based on point-plane matching, obtain the residuals of geometric plane points: Where i represents the frame index, d ε d represents the residual at the geometric edge point. γ d represents the residual at the edge of reflectivity. H The residuals of geometric plane points are represented by the geometric edge points, geometric plane points, and reflectivity edge points in frame i+1 under radar system L. These points are initially transformed to the global coordinate system through robot pose estimation using inertial navigation. The transformed points are represented by X. (i+1,k) In this representation, k represents the index of the corresponding point in the three feature point sets, and the geometric edge points and reflectivity edge points correspond to the local map. and The straight line is represented by two points X along the direction of the largest eigenvector corresponding to the covariance matrix. (i,j) and X (i,l) This indicates that j and l are the identifiers of the two points representing the lines, and the geometric plane points correspond to the local map. The plane is formed by non-collinear X on the fitted plane. (i,u) X (i,v) and X (i,w) Three points are determined: u, v, and w are the identifiers of the three non-collinear points that determine the fitted plane.

8. The method for three-dimensional reconstruction of indoor environments based on reflectivity edge constraints according to claim 7, characterized in that, Iterative optimization of the residual includes: Find the Jacobian matrix J of the residual pair transformation matrix, and then perform Gaussian-Newton iterative optimization: in J T ·J·ΔX=-J T ·f Among them, T i+1 Represents the transformation matrix. Let f represent the three feature points located in the i+1 frame under radar system L, ΔX represent the pose correction amount, and f represent the residual. Minimize the residuals to obtain the optimal pose transformation: Here, ΔT is the final pose transformation that needs to be solved.

9. A three-dimensional reconstruction system for indoor environments based on reflectivity edge constraints, the system comprising: One or more processors, memory, and one or more programs; wherein the one or more programs are stored in the memory, and the one or more programs include instructions that, when executed by the system, cause the system to perform the method of any one of claims 1-8.

Images