A building model reconstruction method and system based on a quadruped robot and machine vision

Abstract

The application discloses a kind of based on quadruped robot and machine vision's building model reconstruction method and system, based on the quadruped robot of laser scanner, utilize simultaneous localization and map construction technology to generate the global map of the region to be reconstructed.Combining three-dimensional ray tracing and A* path search algorithm determines target scanning point and carries out global path planning.Based on the high-precision sensor of quadruped robot, through adaptive monte carlo positioning and improved dynamic window method algorithm, to realize the accurate positioning and obstacle avoidance of quadruped robot in complex environment.After moving to target scanning point, robot stops and starts laser scanner, carries out high-precision point cloud data acquisition, until completing all sampling.After the preprocessing of the collected point cloud data, through point cloud registration and poisson surface reconstruction technology, generate three-dimensional building model.The application can realize comprehensive, efficient and accurate three-dimensional reconstruction in complex building environment.

Description

Technical Field

[0001] This invention relates to a method and system for reconstructing building models based on quadruped robots and machine vision, belonging to the field of civil engineering and automation interaction technology. Background Technology

[0002] Traditional methods for 3D reconstruction of existing buildings based on laser point clouds mainly rely on manual scanning, which has improved data acquisition efficiency and accuracy. However, these methods are complex, time-consuming, and susceptible to human factors. They also face practical problems such as insufficient scanning coverage in complex environments, making it difficult to meet the needs of large-scale and high-precision operations.

[0003] In recent years, quadruped robots have been increasingly used in complex terrains due to their superior obstacle-crossing ability and flexibility. Quadruped robots can move autonomously in rugged, narrow, and complex building environments, significantly improving data collection capabilities in hard-to-reach areas.

[0004] However, the application of combining quadruped robots with laser point cloud technology for the 3D reconstruction of existing buildings is not yet widespread, with limited research and practical applications. This makes it difficult to achieve efficient, comprehensive, and accurate 3D modeling in practical applications. Existing technologies face the following challenges: First, the selection of scanning points based on human subjective factors leads to low scanning efficiency and poor global coverage. Second, the inability to precisely control the robot's trajectory and posture during movement results in poor stability, safety, and accuracy during the scanning process, making it impossible to acquire high-quality point cloud data.

[0005] Therefore, developing an efficient and accurate 3D reconstruction method and system for existing building models based on quadruped robots and machine vision has become an urgent technical challenge. Summary of the Invention

[0006] Objective: To overcome the shortcomings of existing technologies, this invention provides a method and system for reconstructing building models based on quadruped robots and machine vision. This invention can fully utilize the mobility of quadruped robots and the high precision of laser point clouds to achieve comprehensive, efficient and accurate three-dimensional reconstruction in complex building environments.

[0007] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0008] Firstly, a method for reconstructing architectural models based on quadruped robots and machine vision specifically includes:

[0009] Step S1: Obtain the 3D point cloud data of the reconstructed area and the IMU pose data of the quadruped robot. Use the simultaneous localization and mapping algorithm to perform environmental perception and mapping of the reconstructed area to obtain a global map.

[0010] Step S2: Based on the global map, combine the 3D ray tracing algorithm and the A* path search algorithm to determine the target scan point and its corresponding scan path, and generate a global path plan.

[0011] Step S3: Based on the generated global path plan, the quadruped robot begins to walk. During the walking process, the quadruped robot achieves real-time global localization based on the adaptive Monte Carlo localization algorithm combined with sensor data. An improved DWA algorithm is used for local path planning.

[0012] Step S4: After the quadruped robot reaches the target scanning point, it stops and starts the laser scanner to collect point cloud data and RGB data. After the data collection is completed, the quadruped robot moves to the next target scanning point according to the global path planning in step S2 and repeats step S3 until all raw point cloud data is collected.

[0013] Step S5: Preprocess all the collected raw point cloud data to obtain preprocessed point cloud data.

[0014] Step S6: Apply a multimodal data fusion and registration method to the preprocessed point cloud data to obtain the reconstructed 3D building model.

[0015] As a preferred embodiment, step S1 specifically includes:

[0016] A tightly coupled graph optimization framework is adopted to fuse 3D point cloud data and IMU pose data, and the fused data is used to generate optimized pose estimates by pre-integration technology and dynamic weight allocation mechanism.

[0017] Loop closure detection is performed by combining LiDAR and visual information, and a globally consistent pose-map graph structure is generated through sparse pose graph and LM algorithm.

[0018] Incremental 3D occupancy grid maps are generated based on optimized pose estimation and a globally consistent pose-map graph structure.

[0019] The "pcd2pgm" tool built into the quadruped robot is used to convert the 3D occupied grid map into a 2D grid global map for localization initialization and navigation.

[0020] As a preferred embodiment, step S2 specifically includes:

[0021] The geometric midline of the navigable space is extracted from the global map of a 2D raster to generate a topological skeleton map, and the discretized geometric midline nodes V are transformed using an improved midline transformation algorithm. j Transformed into a three-dimensional candidate scan point set P j .

[0022] Initialize the target point set T as an empty set, and then scan the candidate point set P. j Select the point P with the maximum coverage score. max Add T.

[0023] Set the minimum distance d between two adjacent target scan points min This is to avoid redundancy in the scanned data. And simultaneously, data with the prefix P is removed. max Centered on a radius of d min The neighboring nodes. Continuously update the target point set T, optimize the global coverage C and calculate the increment ΔC, until ΔC is less than 1%, then stop iterating.

[0024] The target scan point set T is obtained.

[0025] After determining the target scan point set T, the A* algorithm is used for path planning to calculate the shortest path and avoid collisions with obstacles, thus generating a preliminary path.

[0026] After generating the initial path, the path is optimized by considering the robot's kinematic constraints to obtain the global path plan.

[0027] As a preferred embodiment, the local path planning specifically includes:

[0028] When executing the global path planning scan path, the quadruped robot uses an adaptive Monte Carlo localization algorithm to achieve real-time localization.

[0029] Based on the current speed and acceleration limitations of quadruped robots, an improved dynamic window algorithm is used to define a dynamic speed window. Within the speed window, each possible speed vector is evaluated, and the distance between the quadruped robot and obstacles and the probability of approaching the target point when the quadruped robot moves with that speed vector are calculated. After comprehensive consideration, the optimal speed vector is selected to guide the quadruped robot's movement.

[0030] As a preferred embodiment, the method for acquiring the preprocessed point cloud data specifically includes:

[0031] An octree-based downsampling method is used to downsample the original point cloud data Q to generate downsampled point cloud data Q′.

[0032] The conditional filtering algorithm is used to remove noise from the processed point cloud data Q′, filtering out outliers and missing points to obtain the processed point cloud data Q″.

[0033] Based on the boundary conditions of the target area, point cloud data irrelevant to the target area are removed to obtain the point cloud data Q″′ of the reconstructed area. The point cloud data Q″′ of the reconstructed area is then used as the preprocessed point cloud data Q″′.

[0034] As a preferred embodiment, the method for obtaining the reconstructed three-dimensional building model specifically includes:

[0035] The preprocessed point cloud data Q″′ is aligned using an iterative nearest point registration algorithm.

[0036] A Gaussian mixture model registration model with visual feature assistance is constructed to further align the point cloud data.

[0037] The PointNet++ architecture is used to complete missing regions of point cloud data, generating a registered and completed point cloud set Q. aligned .

[0038] Set the point cloud Q aligned Input the Poisson surface reconstruction algorithm to perform surface reconstruction and generate a reconstructed 3D building model.

[0039] As a preferred embodiment, the Poisson surface reconstruction algorithm specifically includes:

[0040] A method based on triangular mesh structure fitting is used to estimate the normal vector for each point q in the registered point cloud. i Estimate the normal vector n i , is used to define the local geometric properties of the point cloud surface.

[0041] Poisson surface reconstruction constructs a global optimization problem based on the Poisson equation, using the normal vector in the point cloud as the source term, and solves for a continuous scalar field φ.

[0042] Using a volume rendering algorithm, isosurfaces are extracted from the scalar field φ, which is the reconstructed 3D building model.

[0043] Secondly, a building model reconstruction system based on quadruped robots and machine vision specifically includes:

[0044] The navigation module is used for the autonomous movement and localization of the quadruped robot, as well as the selection of scanning points and path planning. It integrates data from the inertial measurement unit and lidar sensors, and optimizes the target scanning points and scanning path based on algorithms.

[0045] The data acquisition module is used to acquire point cloud data and simultaneously record the robot's current position and posture.

[0046] The data processing module is used to preprocess, register, and reconstruct the three-dimensional surface of the acquired point cloud data.

[0047] Beneficial Effects: This invention provides a method and system for reconstructing building models based on a quadruped robot and machine vision. Using a quadruped robot equipped with a LiDAR scanner, a stop-and-go scanning strategy is employed to achieve high-precision 3D reconstruction of existing buildings in complex environments. The steps include: generating a global map of the area to be reconstructed using Simultaneous Localization and Mapping (SLAM) technology; determining target scanning points and planning a global path using 3D ray tracing and A* path search algorithms; and achieving precise positioning and obstacle avoidance in complex environments using high-precision sensors on the quadruped robot and adaptive Monte Carlo localization (AMCL) and improved dynamic windowing (DWA) algorithms. After moving to the target scanning point, the robot stops and activates the LiDAR scanner to acquire high-precision point cloud data. After acquisition, it moves along the planned path to the next scanning point until all sampling is completed. The acquired point cloud data undergoes preprocessing such as downsampling, filtering, and cropping. Then, a 3D building model is generated using point cloud registration and Poisson surface reconstruction techniques.

[0048] This invention effectively reduces manual intervention, improves reconstruction accuracy and efficiency, and has broad application prospects. Attached Figure Description

[0049] Figure 1 This is a flowchart of a building model reconstruction method based on quadruped robots and machine vision.

[0050] Figure 2 This is a schematic diagram of the Unitree GO2-EDU quadruped robot and its laser scanner used in one embodiment.

[0051] Figure 3 It is a global map constructed in one embodiment.

[0052] Figure 4 This is a schematic diagram of local path planning and obstacle avoidance constructed in one embodiment.

[0053] Figure 5 This is a schematic diagram of raw point cloud data collected by a laser scanner mounted on a quadruped robot in one embodiment.

[0054] Figure 6 This is a schematic diagram of point cloud data of an existing building area to be rebuilt after preprocessing and point cloud registration, as shown in one embodiment.

[0055] Figure 7 This is a schematic diagram of a three-dimensional surface reconstruction performed using the Poisson surface reconstruction algorithm in one embodiment. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0057] The present invention will be further described below with reference to specific embodiments.

[0058] Example 1:

[0059] This embodiment introduces a method for reconstructing architectural models based on quadruped robots and machine vision, such as... Figure 1 As shown, the specific steps include:

[0060] Step S1: The three-dimensional point cloud data of the reconstructed area and the three-axis angular velocity and linear acceleration data of the quadruped robot are acquired by the lidar and inertial measurement unit (IMU) on the quadruped robot, respectively. The Simultaneous Localization and Mapping (SLAM) algorithm is used to perform environmental perception and mapping of the reconstructed area to obtain a global map.

[0061] Step S2: Based on the generated global map, combine the 3D ray tracing algorithm and the A* path search algorithm to determine the target scan points and their corresponding scan paths, and generate a global path plan that satisfies the dual constraints of scan coverage and path efficiency.

[0062] Step S3: Based on the generated global path plan, the quadruped robot begins to walk. During the walking process, the quadruped robot achieves real-time global localization based on the Adaptive Monte Carlo Localization (AMCL) algorithm combined with sensor data. Local path planning adopts an improved DWA algorithm, which achieves stability and safe obstacle avoidance in complex environments by constructing a dynamic velocity constraint function.

[0063] Step S4: After the quadruped robot reaches the target scanning point, it stops and starts the laser scanner to collect point cloud data and RGB data. After the data collection is completed, the quadruped robot moves to the next target scanning point according to the global path planning in step S2 and repeats step S3 until all raw point cloud data is collected.

[0064] Step S5: For all the collected raw point cloud data, perform downsampling to reduce the data volume, then perform filtering to remove noise, and finally perform cropping to remove irrelevant areas to obtain preprocessed point cloud data.

[0065] Step S6: The preprocessed point cloud data is fused and registered using a multimodal data fusion method. By fusion and registration of point cloud data, neural network to fill in missing regions, and fusion of RGB texture information to optimize Poisson reconstruction, a reconstructed 3D building model is obtained, so as to achieve high-precision and complete 3D building model reconstruction.

[0066] Furthermore, in step S1, the SLAM (Simultaneous Localization and Mapping) algorithm is used to perform environmental perception and mapping of the reconstructed area to obtain a global map, as detailed below:

[0067] A tightly coupled graph optimization framework is adopted to fuse LiDAR point cloud data and IMU pose data (three-axis angular velocity and linear acceleration data of a quadruped robot), and the fused data is used to generate optimized pose estimation by pre-integration technology and dynamic weight allocation mechanism.

[0068] Loop closure detection is performed by combining LiDAR and visual information. Robustness is improved by multimodal feature fusion. A globally consistent pose-map graph structure is generated by using sparse pose graphs and the LM algorithm (Levonburg-Marquardt optimization algorithm).

[0069] Incremental 3D occupancy grid maps are generated based on optimized pose estimation and a globally consistent pose-map graph structure.

[0070] The "pcd2pgm" tool built into the quadruped robot (which converts pcd point cloud files into a grid map for navigation) is used to transform the 3D occupied grid map into a global map G of 2D grids for localization initialization and navigation.

[0071] Furthermore, step S2, which combines the 3D ray tracing algorithm and the A* path search algorithm to determine the target scan point and its corresponding scan path, and generates a global path plan, specifically includes the following steps:

[0072] Step S2-1: Determine the target scanning point

[0073] The geometric midline of the navigable space is extracted from the global map G of a 2D raster to generate a topological skeleton map S, and the discretized geometric midline nodes V are transformed using an improved midline transformation algorithm. j Transformed into a three-dimensional candidate scan point set P j Each three-dimensional point P j The coordinates are (x j ,y j ,z j ), where z j Ground height is dynamically compensated using real-time ranging data from lidar and attitude feedback from IMU.

[0074] A three-dimensional coverage evaluation model was constructed, and a ray tracing algorithm with a distance attenuation factor B was used to calculate the coverage score for each point. Coverage score (Score) j The calculation formula is as follows:

[0075]

[0076] Where, α k Here, B represents the area's coverage rate, and d is the attenuation factor. k Let N be the distance between this point and other scan points, and N be the total number of points in the candidate scan point set.

[0077] Implement incremental optimization iterations. Initialize the target point set T as an empty set, and then scan the candidate point set P. j Select the point P with the maximum coverage score. max Add T.

[0078] Taking into account both scanning efficiency and the performance parameters of the lidar, the minimum distance d between two adjacent target scanning points is set. min This is to avoid redundancy in the scanned data. And simultaneously, data with the prefix P is removed. max Centered on a radius of d min The neighboring nodes. Continuously update the target point set T, optimize the global coverage C and calculate the increment ΔC, until ΔC is less than 1%, then stop iterating.

[0079] The target scan point set T is obtained.

[0080] Step S2-2: Scan Path Planning

[0081] After determining the target scan point set T, the A* algorithm is used for path planning to calculate the shortest path and avoid collisions with obstacles, thus generating a preliminary path.

[0082] After generating the initial path, the path is optimized by considering the robot's kinematic constraints (such as maximum acceleration and turning radius) to obtain a global path plan, which ensures that the robot can move smoothly while balancing path length, safety and scanning efficiency.

[0083] Furthermore, the local path planning and obstacle avoidance described in step S3 specifically includes the following steps:

[0084] When executing the global path planning scan path, the quadruped robot uses the Adaptive Monte Carlo Localization (AMCL) algorithm to achieve real-time high-precision positioning.

[0085] The adaptive Monte Carlo localization algorithm randomly distributes particles in the environment and dynamically updates the particle weights based on sensor observation data, ultimately generating a robot position estimate through a resampling process. Based on the quadruped robot's current speed and acceleration limitations, an improved dynamic window algorithm defines a dynamic speed window. Within this window, each possible speed vector is evaluated, calculating the distance to obstacles and the probability of approaching the target point when the quadruped robot moves with that speed vector. After comprehensive consideration, the optimal speed vector is selected, thereby guiding the quadruped robot to move safely and efficiently.

[0086] The improved Dynamic Window (DWA) algorithm is as follows:

[0087] G(v,ω)=α·heading(v,ω)+β·dist(v,ω)+γ·velocity(v,ω)+δ·smoothness(v,ω)

[0088] Here, `heading(v,ω)` describes the deviation of the robot's current direction of motion from the target direction, `dist(v,ω)` describes the minimum distance between the robot and surrounding obstacles, and `velocity(v,ω)` describes the robot's current velocity. `v` and `ω` represent the robot's linear velocity and angular velocity, respectively. `α`, `β`, `γ`, and `δ` are weighting coefficients. An acceleration smoothness term `smoothness(v,ω)` is introduced to penalize drastic changes in velocity and angular velocity, encouraging the quadruped robot to choose trajectories with smaller accelerations and angular accelerations, thereby achieving smooth control and ensuring the stability and safety of the laser scanner mounted on the quadruped robot. Figure 4 The yellow curve represents the local path generated based on the improved Dynamic Window (DWA) algorithm.

[0089] The acceleration smoothness term smoothness(v,ω) is defined as follows:

[0090] Smoothness(v,ω)=-[|(vv current )| / Δt+|(ω-ω current )| / Δt], where Δt is the time step.

[0091] Furthermore, the point cloud data and RGB data will be stored in the local memory of the laser scanner. Furthermore, the method for acquiring the preprocessed point cloud data in step S5 specifically includes the following steps:

[0092] S5-1: Downsampling

[0093] An octree-based downsampling method is used to downsample the original point cloud data Q to a lower resolution, generating downsampled point cloud data Q′, thereby reducing the amount of point cloud data and improving the efficiency of subsequent processing.

[0094] S5-2: Filtering Process

[0095] The conditional filtering algorithm is used to remove noise from the processed point cloud data Q′, filtering out outliers and missing points to obtain the processed point cloud data Q″.

[0096] S5-3: Point Cloud Cropping

[0097] Based on the boundary conditions of the target region, point cloud data irrelevant to the target region is removed to obtain the point cloud data Q″′ of the reconstructed region. The preprocessed point cloud data Q″′ serves as the input for subsequent registration and reconstruction.

[0098] Furthermore, the method for obtaining the reconstructed 3D building model described in step S6 specifically includes the following steps:

[0099] S6-1: Coarse Registration of Point Clouds

[0100] The Iterative Closest Point (ICP) registration algorithm is used to precisely align the preprocessed point cloud data Q″′, thus initially ensuring the consistency of point cloud data collected from different perspectives in three-dimensional space.

[0101] S6-2: Secondary Fine Registration

[0102] Building upon coarse registration, a visual feature-assisted Gaussian Mixture Model (GMM) registration model is constructed to further refine the alignment of point cloud data. By combining visual features (such as edges and corners in RGB images) with point cloud data, registration accuracy is improved, addressing the registration challenges of different modalities.

[0103] S6-3: Missing Region Completion

[0104] Neural network technology, specifically the PointNet++ architecture, is employed to complete missing regions in point cloud data. A deep learning model infers from known point cloud regions to generate appropriate point clouds for the missing areas, ensuring the integrity of the 3D model. The final result is a registered and completed point cloud set Q. aligned .

[0105] S6-4: Poisson Reconstruction Optimization

[0106] Point cloud data Q alignedInput the Poisson surface reconstruction algorithm to perform surface reconstruction and generate a high-precision 3D building model M: M = PoissonReconstruction(Q aligned The Poisson surface reconstruction algorithm described herein combines point cloud data acquired by a laser scanner with RGB information to achieve multimodal data fusion, utilizing color information to assist in normal vector estimation and surface detail reconstruction. This method can efficiently generate smooth and continuous 3D surfaces from dense and noisy point cloud data and improve the detail and accuracy of the reconstructed 3D model. Specifically, it includes the following steps:

[0107] S6-4-1: Normal Vector Estimation

[0108] A method based on triangular mesh structure fitting is used to estimate the normal vector for each point q in the registered point cloud. i Estimate the normal vector n i , is used to define the local geometric properties of the point cloud surface.

[0109] S6-4-2: Constructing the Poisson equation and solving the optimization problem

[0110] Poisson surface reconstruction constructs a global optimization problem based on the Poisson equation, using the normal vector in the point cloud as the source term, to solve for a continuous scalar field φ, whose gradient field is closest to the direction of the point cloud normal vector. Here, φ is the scalar field to be solved, and n is the normal vector field. The solution process of the Poisson equation aims to find a smooth scalar field whose gradient is as consistent as possible with the normal vector field of the point cloud.

[0111] S6-4-3: Extracting Isosurfaces

[0112] The Marching Cubes algorithm is used to extract isosurfaces from the scalar field φ, which are the final three-dimensional surface reconstruction model M.

[0113] Example 2:

[0114] This embodiment introduces a building model reconstruction system based on a quadruped robot and machine vision, specifically including:

[0115] The navigation module is responsible for the quadruped robot's autonomous movement and precise positioning, as well as the selection of scanning points and path planning. By fusing data from the inertial measurement unit and LiDAR sensors, and optimizing the target scanning points and scanning path based on algorithms, it improves the completeness and efficiency of point cloud data acquisition. The navigation module ensures that the robot can safely and accurately reach the predetermined scanning location.

[0116] The data acquisition module is responsible for acquiring high-quality point cloud data and simultaneously recording the robot's current position and posture. This module uses a laser scanner mounted on the quadruped robot to ensure accurate acquisition and timely storage of point cloud and RGB data, providing reliable raw data for subsequent data processing.

[0117] The data processing module is responsible for preprocessing, registering, and reconstructing 3D surfaces from the acquired point cloud data. Preprocessing includes steps such as downsampling, filtering, and cropping. Then, the point cloud data from different perspectives are aligned using a registration algorithm. Finally, a high-precision 3D building model is generated using a Poisson surface reconstruction algorithm.

[0118] Example 3:

[0119] This embodiment describes the working principle of a building model reconstruction method based on a quadruped robot and machine vision, specifically including: the apparatus for implementing the method of the present invention specifically includes: as follows Figure 2 As shown, the quadruped robot Unitree GO2-EDU (including inertial measurement unit (IMU), lidar and other sensors) and a laser scanner mounted on the quadruped robot platform are shown. The specific steps for constructing the 3D building model are as follows:

[0120] Step S1: As Figure 5 As shown, the quadruped robot (Unitree Go2-Edu), equipped with an inertial measurement unit and LiDAR sensors, utilizes Simultaneous Localization and Mapping (SLAM) technology to perform comprehensive environmental perception and mapping of the reconstructed area. Specifically, it employs a factor map-based radar inertial odometry framework (LIO-SAM) algorithm to fuse multi-sensor data in real time. By introducing landmark recognition and loop closure detection mechanisms, combined with backend optimization techniques, high-precision, real-time trajectory estimation and map building for the mobile robot are achieved. The WebRTC module is used to issue control commands, guiding the quadruped robot to complete a full loop around the target scene. During control, the quadruped robot is required to rotate a full circle at turning points to improve the integrity of the scanned data. Subsequently, the pcd2pgm tool is used to convert the 3D point cloud data into a 2D grid global map for localization initialization and navigation.

[0121] Step S2: Based on the generated 2D raster global map, a combination of 3D ray tracing and A* path search algorithms is used to determine the target scan points and their corresponding scan paths, thereby ensuring scan coverage and optimizing scan efficiency. The map parameters are set as follows: resolution is 0.05m, meaning each pixel represents 0.05 meters in actual space; the occupied threshold is set to 0.5, meaning a cell with an occupation probability exceeding this threshold is considered occupied; the free threshold is set to 0.196, meaning a cell with an occupation probability below this threshold is considered free. Specifically, the 3D ray tracing algorithm is used to evaluate the coverage of each candidate scan point. By simulating the propagation path of rays in 3D space, the area that each candidate scan point can cover is accurately calculated. Subsequently, the scan path is globally optimized using the A* algorithm to plan the optimal path, maximizing scan coverage and improving overall scan efficiency. Simultaneously, the scan path planning must satisfy the minimum distance condition d between scan points. min To avoid data redundancy and ensure the comprehensiveness and accuracy of the scan. In this example, considering the coverage area of ​​the laser scanner and the on-site scanning environment, d is defined. min = 2.80m. The determined target scanning points and scanning path are as follows: Figure 3 As shown, Figure 3 The asterisk (*) represents a defined target scanning point, and the blue curve represents the actual movement trajectory (global path) of the quadruped robot. In this example, a total of 5 target scanning points were determined, achieving a theoretical scanning coverage of C = 97.68%.

[0122] Step S3: During the quadruped robot's movement, based on the constructed global map, the Adaptive Monte Carlo Localization (AMCL) algorithm is used to achieve real-time, high-precision localization. This algorithm randomly distributes particles in the environment and dynamically updates the particle weights based on sensor observation data, ultimately generating a position estimate for the robot through a resampling process. In this example, the key parameters are set as follows: kld_err = 0.01, kld_z = 0.99, update_min_d = 0.2m, resample_interval = 2s, initial_pose_a = 0.0m. These parameter settings optimize computational efficiency while ensuring localization accuracy. Furthermore, an improved Dynamic Window Approach (DWA) algorithm is used for local path planning and obstacle avoidance. The weight coefficients α, β, γ, and δ are determined based on the relative importance of each evaluation index in path planning. In this example, α = 1.0, β = 1.5, γ = 0.8, and δ = 1.2. Specifically, α = 1.0 ensures the consistency of the quadruped robot's movement direction with the target direction; β = 1.5 emphasizes the importance of maintaining a safe distance from obstacles; γ = 0.8 balances movement speed and stability; and δ = 1.2 enhances the smoothness of the movement trajectory, improving the stability of the laser scanner and the accuracy of data acquisition. The combination of these algorithms and parameters ensures the robot's stability and navigation efficiency in complex environments, meeting the requirements for high-precision 3D reconstruction.

[0123] Step S4: After the quadruped robot reaches the target scanning point, it stops and starts the laser scanner to collect point cloud data and RGB data. After the data collection is completed, the quadruped robot moves to the next target scanning position according to the global path planned in step S2 and repeats step S3 until all point cloud data of the 5 target scanning points are collected.

[0124] Step S5: Preprocess the raw point cloud data collected by the quadruped robot to ensure the efficiency and accuracy of subsequent registration and reconstruction processes. First, an octree-based downsampling method is used to downsample the raw point cloud data Q, significantly reducing the data volume and improving the efficiency of subsequent processing. Specifically, the Octree filter in the Open3D library is used, with the octree level set to 11, to spatially partition the point cloud data. The octree method recursively divides the space into subspaces and refines them layer by layer, ultimately sampling and simplifying the point cloud at the specified level. With Octree Level = 11, the raw point cloud data is reduced from 102,041,211 points to 6,851,764 points, thus generating the downsampled point cloud data Q'. Subsequently, a statistical filtering algorithm is used to remove noise and outliers from the point cloud data Q'. In practice, the StatisticalOutlierRemoval filter from the Open 3D library is used. The mean K-neighborhood is set to 100 points, and the standard deviation threshold is 1.0. By statistically analyzing the distance distribution of each point's surrounding neighbors, outliers with a mean distance exceeding the set threshold are removed, thus generating the processed point cloud data Q". Finally, based on the boundary conditions of the target reconstruction region, the processed point cloud Q" is cropped, removing areas irrelevant to the target and retaining the point cloud Q"' of the region to be reconstructed. Figure 6 As shown, the preprocessed point cloud data has high quality, providing reliable and efficient input data for subsequent point cloud registration and 3D surface reconstruction.

[0125] Step S6: The Iterative Closest Point (ICP) algorithm is used to accurately align the preprocessed point cloud data Q”'. SCENE software is used to perform point-to-point registration of point cloud data acquired from multiple viewpoints. To improve registration accuracy, the maximum allowable distance threshold between registration point pairs is set to 1.5mm. After registration, the system uses a triangular mesh structure fitting method to estimate the normal vectors of the aligned point cloud data. Specifically, the "Compute Normals" function in CloudCompare software is used to estimate the normal vectors for each point cloud q”'. i Estimate the normal vector n i Subsequently, using the "Poisson Surface Reconstruction" plugin, isosurfaces were extracted from the scalar field φ based on the Marching Cubes algorithm to generate a continuous and smooth 3D surface model. Finally, the generated 3D surface reconstruction model M is as follows: Figure 7 As shown, it possesses high-precision and detailed architectural structural features.

[0126] This invention significantly improves the coverage and efficiency of point cloud data acquisition by optimizing scanning path planning and robot movement strategies, while ensuring stability and security during the data acquisition process. Secondly, based on multi-sensor data fusion and advanced positioning algorithms on a quadruped robot, precise positioning of target scanning points is achieved, ensuring high-quality point cloud data stitching and accurate 3D models. Furthermore, the quadruped robot's excellent obstacle-crossing capabilities enable it to cover areas inaccessible by traditional equipment, ensuring the integrity and comprehensiveness of the 3D model. Simultaneously, the laser scanner on the quadruped robot enables simultaneous acquisition and fusion of point cloud and RGB data, allowing for the 3D reconstruction of objects such as pipelines, equipment, and furniture. The methods and systems described exhibit good adaptability and stability in complex building environments, ensuring the accuracy and reliability of the data acquisition process. The entire 3D reconstruction process is highly automated, reducing manual intervention and improving work efficiency and 3D reconstruction results. Through the above technical solutions, this invention provides an efficient, accurate, and comprehensive method and system for 3D reconstruction of existing buildings, possessing significant technical advantages and broad application prospects.

[0127] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for reconstructing architectural models based on quadruped robots and machine vision, characterized in that: Specifically, it includes: Step S1: Acquire the 3D point cloud data of the reconstructed area and the IMU pose data of the quadruped robot. Use the simultaneous localization and mapping algorithm to perform environmental perception and mapping of the reconstructed area to obtain a global map. Step S2: Based on the global map, combined with the 3D ray tracing algorithm and the A* path search algorithm, determine the target scanning point and its corresponding scanning path, and generate a global path plan; Step S3: The quadruped robot begins to walk according to the generated global path plan. During the walking process, the quadruped robot achieves real-time global localization based on the adaptive Monte Carlo localization algorithm and combined with sensor data; and the improved DWA algorithm is used for local path planning. Step S4: After the quadruped robot reaches the target scanning point, it stops and starts the laser scanner to collect point cloud data and RGB data. After the collection is completed, the quadruped robot moves to the next target scanning point according to the global path planning in step S2 and repeats step S3 until all the original point cloud data is collected. Step S5: Preprocess all the collected raw point cloud data to obtain preprocessed point cloud data; Step S6: Apply a multimodal data fusion and registration method to the preprocessed point cloud data to obtain the reconstructed 3D building model; Step S2 specifically includes: The geometric midline of the navigable space is extracted from the global map of a 2D raster to generate a topological skeleton map, and the discretized geometric midline nodes V are transformed using an improved midline transformation algorithm. j Transformed into a three-dimensional candidate scan point set P j ; Initialize the target point set T as an empty set, and then scan the candidate point set P. j Select the point P with the maximum coverage score. max Add T; Set the minimum distance d between two adjacent target scan points min To avoid redundancy in scanned data; and to simultaneously remove data with P max Centered on a radius of d min The neighboring nodes; continuously update the target point set T, optimize the global coverage C and calculate the increment Δt, until Δt is less than 1% and stop iterating; Obtain the target scan point set T; After determining the target scan point set T, the A* algorithm is used for path planning to calculate the shortest path and avoid collisions with obstacles, thus generating a preliminary path. After generating the initial path, the path is optimized by considering the robot's kinematic constraints to obtain the global path plan.

2. The method for reconstructing architectural models based on quadruped robots and machine vision according to claim 1, characterized in that: Step S1 specifically includes: A tightly coupled graph optimization framework is adopted to fuse 3D point cloud data and IMU pose data, and the fused data is used to generate optimized pose estimates by pre-integration technology and dynamic weight allocation mechanism. Loop closure detection is performed by combining LiDAR and visual information, and a globally consistent pose-map graph structure is generated through sparse pose graph and LM algorithm. Incremental 3D occupied grid map is generated based on an optimized pose estimation and a globally consistent pose-map graph structure. The 3D occupied grid map is converted into a global map of 2D grids for positioning initialization and navigation.

3. The method for reconstructing architectural models based on quadruped robots and machine vision according to claim 1, characterized in that: The local path planning specifically includes: When executing the scanning path of global path planning, the quadruped robot adopts an adaptive Monte Carlo localization algorithm to achieve real-time localization; Based on the current speed and acceleration limitations of quadruped robots, an improved dynamic window algorithm is used to define a dynamic speed window. Within the speed window, each possible speed vector is evaluated, and the distance between the quadruped robot and the obstacle and the probability of approaching the target point when the quadruped robot moves with the speed vector are calculated. The optimal speed vector is selected to guide the movement of the quadruped robot.

4. The method for reconstructing architectural models based on quadruped robots and machine vision according to claim 1, characterized in that: The method for acquiring the preprocessed point cloud data specifically includes: An octree-based downsampling method is used to downsample the original point cloud data Q to generate downsampled point cloud data Q′. The conditional filtering algorithm is used to remove noise from the processed point cloud data Q′, filtering out outliers and missing points to obtain the processed point cloud data Q′′. Based on the boundary conditions of the target area, point cloud data irrelevant to the target area are removed to obtain the point cloud data Q′′′ of the reconstructed area. The point cloud data Q′′′ of the reconstructed area is then used as the preprocessed point cloud data Q′′′.

5. The method for reconstructing architectural models based on quadruped robots and machine vision according to claim 1, characterized in that: The method for obtaining the reconstructed 3D building model specifically includes: The preprocessed point cloud data Q′′′ is aligned using an iterative nearest point registration algorithm. Construct a visual feature-assisted Gaussian mixture model registration model to further align the point cloud data; Perform missing region completion on the point cloud data to generate a registered and completed point cloud set Q. aligned ; Set the point cloud Q aligned Input the Poisson surface reconstruction algorithm to perform surface reconstruction and generate a reconstructed 3D building model.

6. The method for reconstructing architectural models based on quadruped robots and machine vision according to claim 5, characterized in that: The Poisson surface reconstruction algorithm specifically includes: A method based on triangular mesh structure fitting is used to estimate the normal vector for each point q in the registered point cloud. i Estimate the normal vector n i , used to define the local geometric properties of the point cloud surface; Poisson surface reconstruction constructs a global optimization problem based on the Poisson equation, using the normal vector in the point cloud as the source term, and solves for a continuous scalar field ϕ. Using a volume rendering algorithm, isosurfaces are extracted from the scalar field ϕ, which is the reconstructed 3D building model.

Images