Positioning method based on monocular camera and building BIM model

Abstract

The invention relates to the technical field of computer vision, in particular to a positioning method based on a monocular camera and a building BIM model. The positioning method comprises the following steps: converting a building BIM model into a four-dimensional point cloud data set containing space coordinates and component type labels; calibrating an internal reference and an external reference of the monocular camera; collecting an image and carrying out denoising, illumination correction and distortion correction; identifying building component types in the image by using a classification neural network, and extracting feature points with component type labels; and based on the component type label consistency constraint, solving the position coordinate and the attitude angle of the camera relative to the BIM model through an iterative optimization algorithm. According to the positioning method based on the monocular camera and the building BIM model, the positioning cost is reduced, and the positioning accuracy and success rate are improved.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a positioning method based on a monocular camera and a building BIM model. Background Technology

[0002] With the development of smart buildings, indoor navigation and Industry 4.0, high-precision indoor positioning technology has become a key infrastructure. Currently, positioning methods in the industry can be roughly divided into hardware-dependent, hardware-free vision, and point cloud matching.

[0003] Among these methods, hardware-dependent approaches, such as those based on UWB, Bluetooth beacons, or WiFi fingerprinting, are costly and complex to implement due to the need to deploy dedicated base stations or signal tags. Hardware-free vision methods, while avoiding hardware costs, suffer from poor recognition performance with a failure rate exceeding 40% in areas with repetitive textures in building structures (such as white walls in corridors or standardized offices) due to the scarcity of feature points. Point cloud matching methods, although utilizing the geometric coordinate information of BIM models, have a mismatch rate exceeding 30% in complex scenarios such as areas with dense doors and windows. Therefore, there is an urgent need in this field for a positioning method that can simultaneously meet the requirements of low hardware deployment costs and high positioning accuracy. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this application is to provide a positioning method based on a monocular camera and a building BIM model, thereby reducing positioning costs and improving positioning accuracy.

[0005] To achieve the above objectives, this application provides a positioning method based on a monocular camera and a building BIM model, comprising:

[0006] Convert the building BIM model into a four-dimensional point cloud dataset containing spatial coordinates and component type labels;

[0007] Calibrate the intrinsic and extrinsic parameters of a monocular camera;

[0008] Acquire images and perform noise reduction, illumination correction, and distortion correction;

[0009] Use a classification neural network to identify the types of building components in images and extract feature points labeled with component types;

[0010] Based on the component type label consistency constraint, the position coordinates and attitude angle of the camera relative to the BIM model are solved by an iterative optimization algorithm.

[0011] Furthermore, the component type label includes at least one of walls, doors, windows, columns, beams, floors, and stairs.

[0012] Furthermore, the specific steps for converting the building BIM model into a four-dimensional point cloud dataset containing spatial coordinates and component type labels include:

[0013] Perform point cloud sampling on the geometric surface of the BIM model according to the preset sampling density;

[0014] Attach a corresponding component type label to each sampling point;

[0015] Establish a mapping relationship between point cloud coordinates and component type labels to generate a four-dimensional point cloud dataset.

[0016] Furthermore, the intrinsic parameter calibration adopts the Zhang Zhengyou calibration method, and the extrinsic parameter calibration adopts the PnP algorithm.

[0017] Furthermore, the specific steps for acquiring images and performing noise reduction, illumination correction, and distortion correction include:

[0018] Image denoising is performed using a Gaussian filter kernel;

[0019] Illumination correction is performed using contrast-limited adaptive histogram equalization.

[0020] Distortion correction is performed by inverse mapping of the radial and tangential distortion coefficients of a monocular camera.

[0021] Furthermore, the classification neural network includes either ResNet-50 or EfficientNet.

[0022] Furthermore, the specific steps for extracting feature points labeled with component type include:

[0023] Simultaneously, SIFT, SURF, and ORB algorithms are used to extract feature points with component type labels, with each algorithm extracting a maximum of 500 feature points;

[0024] 1,000 optimal feature points were selected based on feature response intensity and spatial distribution.

[0025] Furthermore, the iterative optimization algorithm is an improved iterative closest point algorithm, and the specific steps include:

[0026] Image feature points are back-projected into three-dimensional space to form an observation point cloud;

[0027] In the 4D point cloud dataset, only points with the same component type label as the observation point are matched;

[0028] The rotation matrix and translation vector that minimize the sum of squared distances between matching point pairs are obtained by using the least squares method.

[0029] Iterative optimization continues until the pose change between adjacent iterations is less than a threshold or the maximum number of iterations is reached.

[0030] Furthermore, a KD-Tree is used to accelerate the search during the matching process, with the search radius set to 0.5 meters, and weights are assigned to feature points of different component types.

[0031] Furthermore, the method also includes:

[0032] Kalman filtering or particle filtering algorithms are used to smooth the pose estimation results of consecutive frames in the temporal domain, thereby reducing positioning jitter.

[0033] Furthermore, the method also includes:

[0034] A layered positioning strategy is adopted for large architectural spaces, including:

[0035] In the coarse positioning stage, low-resolution BIM point cloud is used to determine the location area;

[0036] In the fine positioning stage, the high-resolution BIM point cloud of the loaded area is used to solve the pose at the centimeter level.

[0037] To achieve the above objectives, this application also provides a positioning system based on a monocular camera and a building BIM model, including: a BIM preprocessing module for converting the building BIM model into a four-dimensional point cloud dataset with component type labels;

[0038] Camera calibration module, used to calibrate the intrinsic and extrinsic parameters of a monocular camera;

[0039] An image processing module, connected to the camera calibration module, is used to receive images acquired by a monocular camera and perform noise reduction, illumination correction, and distortion correction on the images based on intrinsic and extrinsic parameters.

[0040] The feature recognition module, connected to the image processing module, includes a classification neural network and a feature detection unit, used to perform component class recognition and feature extraction on the processed image, and generate a set of feature points with component types;

[0041] The pose estimation module is connected to the feature recognition module and the BIM preprocessing module. It is used to execute the ICP algorithm with component type label constraints based on feature points with component type and a four-dimensional point cloud dataset with component type labels, and output the three-dimensional position coordinates and pose angle of the target.

[0042] The positioning method based on a monocular camera and a building BIM model provided in this application does not require the deployment of dedicated base stations or signal tags, resulting in low deployment costs.

[0043] The positioning method based on a monocular camera and a building BIM model provided in this application achieves a positioning accuracy of 5-10 centimeters by fusing prior structural knowledge and visual feature information from the BIM model. This is 30-60 times more accurate than the traditional WiFi positioning method (3-5 meters) and 2-10 times more accurate than the pure visual SLAM method (20-50 centimeters).

[0044] The localization method based on a monocular camera and a building BIM model provided in this application, based on a feature matching strategy constrained by building component type, effectively solves the localization failure problem of traditional visual methods in environments with repetitive textures, varying lighting, and sparse features. The localization success rate reaches over 95% under different lighting conditions, which is 15-25% higher than that of pure visual methods.

[0045] The positioning method based on a monocular camera and a building BIM model provided in this application adopts an optimized ICP algorithm combined with component type constraints to significantly reduce the search space. The processing time for a single frame image is less than 2 seconds, which improves the computational efficiency by 3-5 times compared with the traditional ICP algorithm, thus meeting the needs of real-time navigation applications.

[0046] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing this application. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the present application and form part of the specification. Together with the embodiments of the present application, they serve to explain the present application but do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a flowchart illustrating the positioning method based on a monocular camera and a building BIM model according to Embodiment 1 of this application.

[0049] Figure 2 This is a flowchart illustrating the positioning method based on a monocular camera and a building BIM model according to Embodiment 2 of this application.

[0050] Figure 3 This is a flowchart illustrating the positioning method based on a monocular camera and a building BIM model according to Embodiment 3 of this application.

[0051] Figure 4 This is a schematic diagram of the positioning system based on a monocular camera and a building BIM model according to Embodiment 7 of this application. Detailed Implementation

[0052] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0053] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0054] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0055] It should be noted that the terms "one" and "multiple" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless explicitly stated otherwise in the context, they should be understood as "one or more". "Multiple" should be understood as two or more.

[0056] Example 1

[0057] One embodiment of this application provides a positioning method based on a monocular camera and a building BIM model, which is applied to an office building. Figure 1 This is a flowchart illustrating the positioning method based on a monocular camera and a building BIM model according to Embodiment 1 of this application. The following will refer to... Figure 1 The localization method based on a monocular camera and a building BIM model described in this application is described in detail, including:

[0058] Step S101: Convert the building BIM model into a four-dimensional point cloud dataset containing spatial coordinates and component type labels.

[0059] In this implementation, the BIM (Building Information Modeling) model file (in IFC or Revit format) of the office building is first obtained. The BIM model file contains three floors with a total building area of ​​2000 square meters. Then, the BIM model is processed into point clouds using the open-source libraries Open3D or PCL, with a sampling density of 200 sampling points per square meter. The interior surfaces of the building (including walls, floors, and ceilings) are uniformly sampled, generating approximately 400,000 three-dimensional coordinate points. Next, based on the component classification information in the BIM model, a component type label is attached to each point cloud. The component type labels include: Wall (Wal, label number 1), Door (Door, label number 2), Window (Window, label number 3), Column (Column, label number 4), Slab (Slab, label number 5), etc. The final constructed four-dimensional point cloud dataset is stored in the following format: each row contains 4 values ​​(x, y, z, label), where x, y, and z are the three-dimensional coordinates of the point cloud point, and label is the component type label of the point cloud point. The file size is approximately 50MB.

[0060] In other implementations, the sampling density can be 100-500 sampling points per square meter. For example, a layered positioning strategy is adopted for large building spaces. In the coarse positioning stage, a sampling density of 50-100 sampling points per square meter is used, and in the fine positioning stage, a sampling density of 300-500 sampling points per square meter is used.

[0061] Step S102: Calibrate the intrinsic and extrinsic parameters of the monocular camera.

[0062] In this embodiment, the monocular camera used is the built-in monocular lens of a smartphone (lens resolution 1920x1080, field of view approximately 70 degrees) for intrinsic parameter calibration. The Zhang Zhengyou calibration method was employed, and a 9x6 checkerboard calibration board was fabricated. Twenty sets of calibration images were captured at different angles and distances. The camera's intrinsic parameters were calculated using the `calibrateCamera` function from the OpenCV library: focal length fx = 1420 pixels, fy = 1425 pixels; principal point coordinates cx = 960 pixels, cy = 540 pixels; radial distortion coefficients k1 = -0.15, k2 = 0.08, k3 = -0.02; tangential distortion coefficients p1 = 0.001, p2 = -0.002. The calibration reprojection error was 0.35 pixels, meeting the accuracy requirements. The extrinsic parameter calibration determines the spatial pose of the camera relative to the world coordinate system, including the 3x3 rotation matrix R and the 3x1 translation vector t. The camera extrinsic parameters are solved by the PnP (Perspective-n-Point) algorithm, providing accurate geometric constraints for subsequent 3D reconstruction and pose estimation.

[0063] Step S103: Acquire the image and perform noise reduction, illumination correction and distortion correction.

[0064] In this embodiment, for each frame of image, a 5x5 Gaussian filter kernel with a standard deviation of o = 1.0 is used for filtering to remove CCD sensor noise.

[0065] In this embodiment, the CLAHE (Contrast Limiting Adaptive Histogram Equalization) algorithm is used, with a clipping limit of 2.0 and a grid size of 8x8, to enhance the local contrast of the image for illumination correction.

[0066] In this embodiment, the undistort function is used to eliminate radial and tangential distortion based on the calibrated distortion coefficients, ensuring the geometric accuracy of the image.

[0067] In some other embodiments, a bilateral filtering algorithm can be used for filtering, with a filter kernel size of 5x5 or 7x7. In other embodiments, a gamma correction algorithm can be used for illumination correction, with a gamma value set between 0.8 and 1.2. Step S104: Use a classification neural network to identify the types of building components in the image and extract feature points labeled with the component type. In this embodiment, a pre-trained ResNet-50 classification neural network is used for building component classification. The classification neural network is trained on a dataset containing 12,000 indoor images, achieving a classification accuracy of 92%. The classification neural network outputs the component type probability for each pixel, and the component segmentation result is obtained through an argmax operation.

[0068] In this implementation, based on component classification, the ORB feature detection algorithm is used to extract image feature points, with an upper limit of 1000 feature points and a feature descriptor dimension of 256. Each feature point is assigned a corresponding component type label, forming a labeled feature point set. Typically, 800-1000 effective feature points can be extracted per image.

[0069] Understandably, a classification neural network includes a feature extraction layer, a spatial attention mechanism, a global average pooling layer, and a classification output layer.

[0070] In other embodiments, a pre-trained EfficientNet framework classification neural network can be used for building component classification. In still other embodiments, image feature points can be extracted using SIFT (Scale-Invariant Feature Transform), SURF (Speeded Up Robust Features), or ORB (Oriented FAST and Rotated BRIEF) feature detection algorithms. Each feature point contains two-dimensional image coordinates (u, v), a 128-dimensional or 256-dimensional feature descriptor, and a corresponding component type label, forming a labeled feature point set {(ui, vi, di, labeli) | i = 1, 2, ..., M}, where M is the number of extracted feature points and di is the dimension.

[0071] Step S105: Based on the component type label consistency constraint, solve the position coordinates and attitude angle of the camera relative to the BIM model through an iterative optimization algorithm.

[0072] In this embodiment, an improved ICP (Iterative Closest Point) algorithm is used for pose estimation. First, image feature points are back-projected onto a 3D space with a depth of 1 meter using camera intrinsic parameters, forming an observation point cloud. Then, matching points are searched in the 4D BIM point cloud dataset. The matching strategy is to consider only point pairs with the same component type label, using KD-Tree to accelerate the nearest neighbor search, with a search radius of 0.5 meters. After establishing the point pair correspondence, the least squares method is used to solve for the rotation matrix R and translation vector t, and the optimization objective function is to minimize the sum of squared distances between point pairs. The maximum number of iterations is set to 50, and the convergence threshold is set to a position change of less than 0.01 meters and an angle change of less than 0.1 degrees. Finally, the observer's 3D position coordinates (xyz) and pose angles (roll, pitch, yaw) are output.

[0073] In some other implementations, the maximum number of iterations is 50-100.

[0074] In this embodiment, performance verification and testing were conducted in five different types of buildings.

[0075] The five types include: office buildings (2000m²) 2 ), Hospital (8000m) 2 Shopping mall (12000m) 2 )Residential(500m 2 ), factory (5000m) 2For each type of building, 100 test points were selected, for a total of 50 test points. The actual positions were measured using a high-precision total station as the reference.

[0076] Table 1 is a comparison table of positioning performance:

[0077] Table 1

[0078] Positioning method Average error (cm) Standard deviation (cm) Maximum error (cm) Success rate (%) WiFi fingerprint 342 156 680 78 Bluetooth Beacon 185 89 420 85 UWB 28 12 65 92 Visual SLAM 35 18 85 88 This embodiment 7.2 3.1 15 96

[0079] As shown in Table 1, compared with several existing positioning methods, the positioning method of this embodiment 1 significantly reduces the average error, standard deviation, and maximum error, and the positioning success rate reaches 96%.

[0080] Table 2 is a comparison table of computational performance: Table 2

[0081] Performance indicators WiFi fingerprint Bluetooth Beacon UWB Visual SLAM This embodiment Processing time (s) 0.5 0.3 0.1 3.2 1.8 Memory usage (MB) 50 30 20 200 120 Deployment cost (ten thousand yuan) 15 25 50 0 0 Maintenance cost (ten thousand yuan / year) 3 5 8 0 0

[0082] As shown in Table 2, the positioning method of this embodiment 1 has a processing time of only 1.8s, a memory usage of only 120MB, and a deployment and maintenance cost of less than 10,000 yuan.

[0083] Table 3 shows the qualitative performance of the boundary table under different lighting conditions:

[0084]

[0085]

[0086] As shown in Table 3, the positioning method of this embodiment 1 has a positioning success rate of up to 98% on sunny days, with an average error of only 6.5cm. The positioning success rate can also be maintained at 92% in dim environments, with an average error of only 9.1cm. That is, the positioning method of this embodiment 1 has strong positioning robustness under different lighting conditions and is less affected by lighting.

[0087] Example 2

[0088] One embodiment of this application provides a positioning method based on a monocular camera and a building BIM model. Figure 2 This is a flowchart illustrating the positioning method based on a monocular camera and a building BIM model according to Embodiment 2 of this application. Figure 2 The positioning method based on a monocular camera and a building BIM model in Embodiment 2 of this application differs from that in Embodiment 1 in that:

[0089] In this implementation, adaptive BIM point cloud sampling is employed, employing an adaptive sampling strategy based on the geometric complexity and component density of the building space. In areas with dramatic geometric changes (such as column-beam connections around doors and windows), the sampling density is increased to 500 points per square meter, while in flat wall areas, the sampling density is reduced to 100 points per square meter. This strategy ensures richness of features in key areas while reducing the overall data volume and improving matching efficiency.

[0090] In this implementation, a multi-scale feature fusion strategy is adopted. During the feature extraction stage, three feature detection algorithms—SIFT, SURF, and ORB—are used simultaneously to fully leverage their complementary advantages. SIFT extracts scale-invariant features, making it suitable for handling distance variations; SURF is computationally efficient, making it suitable for real-time applications; and ORB is robust to rotational changes, making it suitable for mobile scenarios. Each algorithm is set to extract a maximum of 500 feature points, resulting in a total of 1500 candidate feature points. Feature points are then selected based on feature response intensity and spatial distribution uniformity, ultimately retaining 1000 optimal feature points.

[0091] In this embodiment, a weighted ICP optimization algorithm is used. Feature importance weights are introduced into the ICP algorithm of Example 1, assigning different weights based on the response intensity of feature points, the rarity of component types, and spatial distribution characteristics. For example, the weight of structural feature points such as wall corners and door frames is set to 1.0, while the weight of ordinary wall feature points is set to 0.5. The optimization objective function is modified to a weighted least squares form to enhance the influence of key features.

[0092] The method of Embodiment 2 of this application was verified using the same 50 test points. It was found that the average positioning error was reduced to 5.8 cm, the standard deviation was reduced to 2.4 cm, and the maximum error was controlled within 11.2 cm. The calculation time increased slightly, averaging 2.1 seconds, but the positioning accuracy was improved by about 20%. That is, the positioning method of Embodiment 2 of this application has a significant advantage in high-precision scenarios.

[0093] Example 3

[0094] One embodiment of this application provides a positioning method based on a monocular camera and a building BIM model. Figure 3 This is a flowchart illustrating the positioning method based on a monocular camera and a building BIM model according to Embodiment 3 of this application. (Refer to...) Figure 3 The positioning method based on a monocular camera and a building BIM model in Embodiment 3 of this application differs from that in Embodiment 1 in that:

[0095] In this embodiment, the positioning method is applied to large commercial buildings, employing a layered positioning and area switching strategy.

[0096] In this implementation, large buildings are divided into layers based on floors and functional areas, with each area maintaining an independent 4D BIM point cloud dataset. For example, a large shopping mall may have three floors, each with an area of ​​5,000 square meters, and be divided into 16 sub-areas, each with an area of ​​approximately 300 square meters. This partitioning strategy effectively reduces the amount of data required for a single matching operation and improves computational efficiency.

[0097] In this implementation, a two-stage coarse-to-fine positioning mechanism is employed: First, a coarse positioning stage is used to quickly determine the approximate area and floor of the observer using a low-resolution BIM model (sampling density 50 points / square meter) and a limited number of feature points (200). After coarse positioning is completed, a high-resolution BIM model (sampling density 300 points / square meter) is loaded for the corresponding area to perform precise positioning. This two-stage mechanism keeps the positioning time within 1.5 seconds while maintaining high accuracy.

[0098] In this embodiment, area switching and continuity assurance are also implemented: when the observer moves between different areas, overlapping area design and smooth transition algorithm are used to ensure positioning continuity. A 10% overlap area is set between adjacent areas. When the observer is detected to enter the overlapping area, the BIM data of both areas are loaded simultaneously for positioning, and the Kalman filter algorithm is used to smooth the pose estimation results of consecutive frames in the temporal domain to avoid positioning jumps.

[0099] In other implementations, a particle filtering algorithm can be used to smooth the pose estimation results of consecutive frames in the temporal domain, thereby reducing positioning jitter.

[0100] In this embodiment, multi-sensor data fusion is used: combining IMU (Inertial Measurement Unit) data from a smartphone, including information from a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, to construct a vision-inertial fusion positioning system. The IMU data is used to predict the motion trajectory of the monocular camera, providing an initial estimate for the ICP algorithm and significantly improving the algorithm's convergence speed and accuracy.

[0101] The method of Embodiment 3 of this application was verified in a large shopping mall with a building area of ​​15,000 square meters, including three floors and an underground parking garage. The total test path length was 1.2 kilometers, including complex scenarios such as floor switching and area transitions. The experimental results showed that the average positioning error was 8.5 centimeters. The error increased slightly at the area switching points but remained within 15 centimeters, and the overall positioning success rate reached 96%.

[0102] Example 4

[0103] One embodiment of this application provides a positioning method based on a monocular camera and a building BIM model. The positioning method based on a monocular camera and a building BIM model in Embodiment 4 of this application differs from that in Embodiment 1 in that:

[0104] In this embodiment, the localization method is applied to the autonomous navigation system of an indoor service robot. The robot is equipped with a monocular camera and a LiDAR sensor, which, combined with a BIM model, enables precise path planning and obstacle avoidance. Based on the depth information provided by the LiDAR, the vision-BIM fusion localization method provides the robot with an absolute position reference, solving the cumulative error problem inherent in pure SLAM methods. Experimental results show that after 8 hours of continuous operation, the robot's localization error remains within 10 cm, representing an 80% reduction in cumulative error compared to the pure vision-SLAM method.

[0105] Example 5

[0106] One embodiment of this application provides a positioning method based on a monocular camera and a building BIM model. The difference between the positioning method based on a monocular camera and a building BIM model in embodiment 5 of this application and embodiment 1 is as follows:

[0107] In this embodiment, the positioning method is applied in building construction and maintenance scenarios, and combined with AR technology to achieve accurate overlay display of the BIM model and the real scene. Construction workers use AR glasses or tablet devices to obtain precise position and posture information through the positioning method of this invention. The system overlays information such as pipelines and structures from the BIM model onto the real field of view. The positioning accuracy reaches 5 centimeters, meeting the requirements for accurate AR overlay display and improving construction efficiency and quality control.

[0108] Example 6

[0109] One embodiment of this application provides a positioning method based on a monocular camera and a building BIM model. The difference between the positioning method based on a monocular camera and a building BIM model in Embodiment 6 of this application and Embodiment 1 is as follows:

[0110] In this embodiment, the positioning method is applied in emergency rescue scenarios such as fires and earthquakes. Rescuers utilize the technology of this invention for indoor positioning and navigation in harsh environments such as smoke and dust. The system is pre-installed with a building BIM model, and rescuers acquire image information through helmet cameras. Even in environments with extremely low visibility, reliable positioning can still be achieved based on building structural features. Combined with infrared imaging technology, the positioning success rate reaches 85% in environments with visibility less than 2 meters, providing important location references for rescuers.

[0111] Example 7

[0112] One embodiment of this application provides a positioning system based on a monocular camera and a building BIM model. Figure 4 This is a schematic diagram of the positioning system based on a monocular camera and a building BIM model according to Embodiment 7 of this application. Figure 4 As shown, the positioning system includes:

[0113] BIM preprocessing module 100 is used to convert building BIM models into four-dimensional point cloud datasets with component type labels;

[0114] Camera calibration module 200 is used to calibrate the intrinsic and extrinsic parameters of monocular camera 800;

[0115] The image processing module 300 is connected to the camera calibration module 200 and is used to receive images acquired by the monocular camera 800 and perform noise reduction, illumination correction and distortion correction on the images based on intrinsic and extrinsic parameters.

[0116] The feature recognition module 400, connected to the image processing module 300, includes a classification neural network and a feature detection unit, used to perform component class recognition and feature extraction on the processed image, and generate a set of feature points with component types;

[0117] The pose estimation module 500 is connected to the feature recognition module 400 and the BIM preprocessing module. It is used to execute the ICP algorithm with component type label constraints based on feature points with component type and a four-dimensional point cloud dataset with component type labels, and output the three-dimensional position coordinates and pose angle of the target.

[0118] The above description is merely a partial embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

[0119] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in sequential order. Multitasking and parallel processing may be advantageous in certain environments. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0120] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A positioning method based on a monocular camera and a building BIM model, characterized in that, include: Convert the building BIM model into a four-dimensional point cloud dataset containing spatial coordinates and component type labels; Calibrate the intrinsic and extrinsic parameters of a monocular camera; Acquire images and perform noise reduction, illumination correction, and distortion correction; Use a classification neural network to identify the types of building components in images and extract feature points labeled with component types; Based on the component type label consistency constraint, the position coordinates and attitude angle of the camera relative to the BIM model are solved by an iterative optimization algorithm.

2. The positioning method based on a monocular camera and a building BIM model according to claim 1, characterized in that, The component type label includes at least one of the following: wall, door, window, column, beam, floor slab, and staircase.

3. The positioning method based on a monocular camera and a building BIM model according to claim 1, characterized in that, The specific steps for converting a building BIM model into a four-dimensional point cloud dataset containing spatial coordinates and component type labels include: Perform point cloud sampling on the geometric surface of the BIM model according to the preset sampling density; Attach a corresponding component type label to each sampling point; Establish a mapping relationship between point cloud coordinates and component type labels to generate a four-dimensional point cloud dataset.

4. The positioning method based on a monocular camera and a building BIM model according to claim 1, characterized in that, The internal parameter calibration uses the Zhang Zhengyou calibration method, and the external parameter calibration uses the PnP algorithm.

5. The positioning method based on a monocular camera and a building BIM model according to claim 1, characterized in that, The specific steps for acquiring images and performing noise reduction, illumination correction, and distortion correction include: Image denoising is performed using a Gaussian filter kernel; Illumination correction is performed using contrast-limited adaptive histogram equalization. Distortion correction is performed by inverse mapping of the radial and tangential distortion coefficients of a monocular camera.

6. The positioning method based on a monocular camera and a building BIM model according to claim 1, characterized in that, The classification neural network includes either ResNet-50 or EfficientNet.

7. The positioning method based on a monocular camera and a building BIM model according to claim 4, characterized in that, The specific steps for extracting feature points with component type labels include: Simultaneously, SIFT, SURF, and ORB algorithms are used to extract feature points with component type labels, with each algorithm extracting a maximum of 500 feature points; 1,000 optimal feature points were selected based on feature response intensity and spatial distribution.

8. The positioning method based on a monocular camera and a building BIM model according to claim 4, characterized in that, The iterative optimization algorithm is an improved iterative closest point algorithm, and the specific steps include: Image feature points are back-projected into three-dimensional space to form an observation point cloud; In the 4D point cloud dataset, only points with the same component type label as the observation point are matched; The rotation matrix and translation vector that minimize the sum of squared distances between matching point pairs are obtained by using the least squares method. Iterative optimization continues until the pose change between adjacent iterations is less than a threshold or the maximum number of iterations is reached.

9. The positioning method based on a monocular camera and a building BIM model according to claim 8, characterized in that, The KD-Tree was used to accelerate the search during the matching process, with the search radius set to 0.5 meters, and weights were assigned to feature points of different component types.

10. The positioning method based on a monocular camera and a building BIM model according to claim 8, characterized in that, The method further includes: Kalman filtering or particle filtering algorithms are used to smooth the pose estimation results of consecutive frames in the temporal domain, thereby reducing positioning jitter.

11. The positioning method based on a monocular camera and a building BIM model according to claim 8, characterized in that, The method further includes: A layered positioning strategy is adopted for large architectural spaces, including: In the coarse positioning stage, low-resolution BIM point cloud is used to determine the location area; In the fine positioning stage, the high-resolution BIM point cloud of the loaded area is used to solve the pose at the centimeter level.

12. A positioning system based on a monocular camera and a building BIM model, characterized in that, include: The BIM preprocessing module is used to convert building BIM models into four-dimensional point cloud datasets with component type labels. Camera calibration module, used to calibrate the intrinsic and extrinsic parameters of a monocular camera; An image processing module, connected to the camera calibration module, is used to receive images acquired by a monocular camera and perform noise reduction, illumination correction, and distortion correction on the images based on intrinsic and extrinsic parameters. The feature recognition module, connected to the image processing module, includes a classification neural network and a feature detection unit, used to perform component class recognition and feature extraction on the processed image, and generate a set of feature points with component types; The pose estimation module is connected to the feature recognition module and the BIM preprocessing module. It is used to execute the ICP algorithm with component type label constraints based on feature points with component type and a four-dimensional point cloud dataset with component type labels, and output the three-dimensional position coordinates and pose angle of the target.

Images