Image acquisition method based on three-dimensional modeling

Abstract

The invention discloses an image acquisition method based on three-dimensional modeling, and the method comprises the steps: A, improving the modeling robustness of a complex scene through the integration of a multi-modal sensor, including a visible light camera, a laser radar, an infrared sensor and a six-axis IMU, through the fusion of a data level, a feature level and a decision level; b, evaluating scene complexity in real time by using a convolutional neural network, dynamically adjusting parameters of resolution, frame rate, exposure time and equipment path, and realizing AI adaptive acquisition; c, geometric point cloud modeling and multi-view texture mapping are combined, global model precision is adjusted and optimized through a bundle method, and a hybrid modeling framework is formed. By combining the multi-modal data fusion technology and the artificial intelligence self-adaptive acquisition mechanism, the precision and efficiency in the three-dimensional modeling process are greatly improved, meanwhile, the adaptability of the model to different scenes is remarkably enhanced, the dependence on complex environmental conditions is effectively reduced, and the three-dimensional modeling technology is more flexible and reliable.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling image acquisition technology, specifically to an image acquisition method based on 3D modeling. Background Technology

[0002] 3D modeling refers to the process of creating a virtual 3D model using computer software or hardware technology. It can transform real objects or abstract concepts into digital representations with length, width, and height. In the process of 3D modeling, photogrammetry, laser scanning, or optical imaging technology is required to capture the surface features of the object from different angles and restore the 3D structure through the geometric relationships or depth information between images.

[0003] Current 3D modeling image acquisition mainly relies on photogrammetry, laser scanning, and hybrid acquisition methods. Photogrammetry involves taking pictures from multiple angles (such as drone surround shots) and using the parallax between images to calculate 3D coordinates, which is suitable for large-scene modeling. Laser scanning uses LiDAR or ground-based 3D laser scanners to emit laser beams and generate millimeter-level precision point cloud data through reflection time difference. Hybrid acquisition methods combine photogrammetry and laser scanning, for example, using a drone to obtain the overall outline and then using a handheld scanner to supplement details.

[0004] However, these methods still have certain limitations in practical applications. Although photogrammetry can quickly capture large scenes, the parallax calculation between images may not be accurate enough when there is insufficient lighting or the surface features of objects are not obvious, resulting in deviations in three-dimensional coordinates. Although laser scanning is highly accurate, the equipment is expensive, the operation is complex, and there are certain requirements for the scanning environment, such as avoiding occlusion and reflection interference. Although hybrid acquisition combines the advantages of the first two, the operation process is more cumbersome, and the data fusion processing also requires a high level of technical expertise and computing power.

[0005] Therefore, an image acquisition method based on 3D modeling is provided. Summary of the Invention

[0006] The purpose of this invention is to provide an image acquisition method based on 3D modeling, which has the advantages of multimodal hybrid modeling, improved modeling efficiency, and solves the problems of low image acquisition efficiency and insufficient model accuracy in current 3D modeling.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an image acquisition method based on three-dimensional modeling, the image acquisition method comprising:

[0008] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0009] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0010] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0011] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0012] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0013] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0014] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0015] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0016] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0017] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0018] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0019] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0020] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0021] Preferably, the implementation of the multimodal sensor in step A is as follows:

[0022] 1) Data-level fusion: Directly integrates multi-sensor data at the raw data level, retaining complete information for subsequent processing;

[0023] 2) Feature-level fusion: Extract edge, texture, and shape features from data from various sensors, and perform correlation and complementation at the feature level;

[0024] 3) Decision-level fusion: Based on the independent decision results of each sensor, a comprehensive judgment is made through Bayesian networks or weighted voting.

[0025] Preferably, the AI ​​adaptive acquisition process in step B is as follows:

[0026] 1) Scene complexity assessment: CNN is used to analyze image entropy, texture density, and object motion vectors to generate acquisition strategies in real time;

[0027] 2) Dynamic parameter adjustment: Automatically adjusts camera resolution, frame rate, and exposure time based on scene complexity;

[0028] 3) Path planning optimization: Plan the acquisition trajectory based on the geometric features of the scene to avoid redundant data acquisition.

[0029] Preferably, the process of the hybrid modeling architecture in step C is as follows:

[0030] 1) Geometric modeling: Generate a basic 3D model from LiDAR point cloud data with sub-millimeter accuracy;

[0031] 2) Texture mapping: Feature point matching and texture mapping are performed using multi-view visible light images;

[0032] 3) By optimizing the bundle adjustment method, the camera parameters and 3D point positions are globally optimized to minimize the reconstruction error to below 0.3mm.

[0033] Preferably, the flow supported by the dynamic scene in step D is as follows:

[0034] 1) Motion compensation algorithm: The optical flow method is used to calculate the inter-frame motion vector to compensate for dynamic blur and deformation;

[0035] 2) Temporal data alignment: Multi-frame model alignment is achieved through feature point matching and the ICP algorithm;

[0036] 3) 4D model generation: Integrates temporal point cloud and texture data to generate a 4D model containing motion sequences, supporting model backtracking at any time point.

[0037] Preferably, the multimodal sensor data fusion processing in step A includes:

[0038] Point cloud registration: Improve the ICP algorithm by adding a normal vector constraint term, the formula of which is:

[0039] E=∑|p i -T·q j | 2 +λ∑|n i -R·n j | 2

[0040] Texture mapping: seams are eliminated based on the Poisson fusion algorithm, supporting super-resolution reconstruction.

[0041] Preferably, step A further includes a data processor and a storage module. The data processor is equipped with the NVIDIA-Jetson-AGX-Orin development kit, which integrates a GPU and an AI acceleration core, and is responsible for real-time data fusion, scene analysis, modeling calculation, and storage management. In step A, the data collected by the multimodal sensors is transmitted to the data processing unit via GigE, USB 3.0, or CAN bus. The storage module adopts an NVMe-SSD array, which provides high-speed data writing and reading capabilities and supports local storage of massive amounts of 3D model data.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This invention, by combining multimodal data fusion technology and an adaptive acquisition mechanism based on artificial intelligence, greatly improves the accuracy and efficiency of the 3D modeling process. It also significantly enhances the model's adaptability to different scenarios, effectively reducing reliance on complex environmental conditions and making 3D modeling technology more flexible and reliable. By integrating a high-performance data processing unit and an AI acceleration core, it enables real-time data fusion and scene analysis, thereby quickly and accurately constructing detailed and highly accurate 3D models. The application of multimodal sensors allows the sensors to capture environmental information from different angles and dimensions, ensuring that the collected data is more comprehensive and accurate. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the image acquisition process of the present invention;

[0045] Figure 2 This is a schematic diagram of the AI ​​adaptive acquisition process of the present invention;

[0046] Figure 3 This is a schematic diagram of the hybrid modeling process of the present invention. Detailed Implementation

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] This invention provides a technical solution: an image acquisition method based on three-dimensional modeling, the image acquisition method comprising:

[0049] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0050] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0051] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0052] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0053] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0054] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0055] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0056] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0057] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0058] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0059] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0060] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0061] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0062] Example 2:

[0063] In Example 1, the following process is added:

[0064] The implementation method of the multimodal sensor in step A is as follows:

[0065] 1) Data-level fusion: Directly integrates multi-sensor data at the raw data level, retaining complete information for subsequent processing;

[0066] 2) Feature-level fusion: Extract edge, texture, and shape features from data from various sensors, and perform correlation and complementation at the feature level;

[0067] 3) Decision-level fusion: Based on the independent decision results of each sensor, a comprehensive judgment is made through Bayesian networks or weighted voting.

[0068] Its image acquisition method includes the following steps:

[0069] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0070] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0071] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0072] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0073] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0074] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0075] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0076] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0077] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0078] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0079] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0080] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0081] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0082] Example 3:

[0083] In Example 2, the following process is added:

[0084] The AI ​​adaptive data acquisition process in step B is as follows:

[0085] 1) Scene complexity assessment: CNN is used to analyze image entropy, texture density, and object motion vectors to generate acquisition strategies in real time;

[0086] 2) Dynamic parameter adjustment: Automatically adjusts camera resolution, frame rate, and exposure time based on scene complexity;

[0087] 3) Path planning optimization: Plan the acquisition trajectory based on the geometric features of the scene to avoid redundant data acquisition.

[0088] Its image acquisition method includes the following steps:

[0089] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0090] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0091] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0092] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0093] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0094] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0095] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0096] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0097] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0098] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0099] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0100] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0101] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0102] Example 4:

[0103] In Example 3, the following process is added:

[0104] The process of the hybrid modeling architecture in step C is as follows:

[0105] 1) Geometric modeling: Generate a basic 3D model from LiDAR point cloud data with sub-millimeter accuracy;

[0106] 2) Texture mapping: Feature point matching and texture mapping are performed using multi-view visible light images;

[0107] 3) By optimizing the bundle adjustment method, the camera parameters and 3D point positions are globally optimized to minimize the reconstruction error to below 0.3mm.

[0108] Its image acquisition method includes the following steps:

[0109] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0110] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0111] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0112] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0113] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0114] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0115] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0116] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0117] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0118] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0119] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0120] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0121] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0122] Example 5:

[0123] In Example 4, the following process is added:

[0124] The flow supported by the dynamic scene in step D is as follows:

[0125] 1) Motion compensation algorithm: The optical flow method is used to calculate the inter-frame motion vector to compensate for dynamic blur and deformation;

[0126] 2) Temporal data alignment: Multi-frame model alignment is achieved through feature point matching and the ICP algorithm;

[0127] 3) 4D model generation: Integrates temporal point cloud and texture data to generate a 4D model containing motion sequences, supporting model backtracking at any time point.

[0128] Its image acquisition method includes the following steps:

[0129] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0130] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0131] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0132] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0133] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0134] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0135] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0136] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0137] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0138] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0139] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0140] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0141] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0142] Example 6:

[0143] In Example 5, the following process is added:

[0144] Step A: Multimodal sensor data fusion processing:

[0145] Point cloud registration: Improve the ICP algorithm by adding a normal vector constraint term, the formula of which is:

[0146] E=∑|p i -T·q j | 2 +λ∑|n i -R·n j | 2

[0147] Texture mapping: seams are eliminated based on the Poisson fusion algorithm, supporting super-resolution reconstruction.

[0148] Its image acquisition method includes the following steps:

[0149] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0150] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0151] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0152] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0153] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0154] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0155] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0156] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0157] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0158] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0159] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0160] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0161] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0162] Example 7:

[0163] In Example 6, the following process is added:

[0164] Step A also includes a data processor and a storage module. The data processor is equipped with the NVIDIA-Jetson-AGX-Orin development kit, which integrates a GPU and AI acceleration core, and is responsible for real-time data fusion, scene analysis, modeling calculation and storage management. In Step A, the data collected by the multimodal sensors is transmitted to the data processing unit via GigE, USB 3.0 or CAN bus. The storage module adopts an NVMe-SSD array, which provides high-speed data writing and reading capabilities, and supports local storage of massive 3D model data.

[0165] Its image acquisition method includes the following steps:

[0166] A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion.

[0167] A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm;

[0168] A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference;

[0169] A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration;

[0170] A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm;

[0171] B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition.

[0172] B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution;

[0173] B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry;

[0174] C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture;

[0175] C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds;

[0176] C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images;

[0177] C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm.

[0178] D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

[0179] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0180] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An image acquisition method based on three-dimensional modeling, characterized in that: Its image acquisition methods include: A. By integrating multimodal sensors, including visible light cameras, lidar, infrared sensors, and six-axis IMUs, the robustness of complex scene modeling is improved through data-level, feature-level, and decision-level fusion. A1) A multimodal sensor array is set up, and the lidar scans the geometric contour with an accuracy of 0.1mm; A2) The visible light camera uses ring lighting and is equipped with a polarizing filter to eliminate reflection interference; A3) A six-axis IMU records the device's posture during the acquisition process for subsequent data registration; A4) An infrared sensor array captures motion trajectories with a resolution of 0.5 mm; B. Utilize convolutional neural networks to evaluate scene complexity in real time and dynamically adjust parameters such as resolution, frame rate, exposure time, and device path to achieve AI adaptive acquisition. B1) Utilize convolutional neural networks to detect the surface complexity of objects and automatically adjust the camera resolution; B2) LiDAR uses multi-echo technology to penetrate transparent materials and obtain internal geometry; C. Combining geometric point cloud modeling with multi-view texture mapping, global model accuracy is optimized through bundle adjustment to form a hybrid modeling architecture; C1) Geometric modeling generates a basic model through Poisson reconstruction of LiDAR point clouds; C2) Texture mapping achieves seamless texture mapping by using SIFT feature matching and perspective correction on multi-view images; C3) Accuracy optimization was achieved by iteratively optimizing the model through bundle adjustment 100 times, bringing the model error to 0.2mm. D. For dynamic scenes, a temporal alignment algorithm is introduced to fuse multi-frame data on the object surface and generate a 4D model that includes the time dimension.

2. The image acquisition method based on three-dimensional modeling according to claim 1, characterized in that: The implementation method of the multimodal sensor in step A is as follows: 1) Data-level fusion: Directly integrates multi-sensor data at the raw data level, retaining complete information for subsequent processing; 2) Feature-level fusion: Extract edge, texture, and shape features from data from various sensors, and perform correlation and complementation at the feature level; 3) Decision-level fusion: Based on the independent decision results of each sensor, a comprehensive judgment is made through Bayesian networks or weighted voting.

3. The image acquisition method based on three-dimensional modeling according to claim 1, characterized in that: The AI ​​adaptive acquisition process in step B is as follows: 1) Scene complexity assessment: CNN is used to analyze image entropy, texture density, and object motion vectors to generate acquisition strategies in real time; 2) Dynamic parameter adjustment: Automatically adjusts camera resolution, frame rate, and exposure time based on scene complexity; 3) Path planning optimization: Plan the acquisition trajectory based on the geometric features of the scene to avoid redundant data acquisition.

4. The image acquisition method based on three-dimensional modeling according to claim 1, characterized in that: The process of the hybrid modeling architecture in step C is as follows: 1) Geometric modeling: Generate a basic 3D model from LiDAR point cloud data with sub-millimeter accuracy; 2) Texture mapping: Feature point matching and texture mapping are performed using multi-view visible light images; 3) By optimizing the bundle adjustment method, the camera parameters and 3D point positions are globally optimized to minimize the reconstruction error to below 0.3mm.

5. The image acquisition method based on three-dimensional modeling according to claim 1, characterized in that: The flow supported by the dynamic scene in step D is as follows: 1) Motion compensation algorithm: The optical flow method is used to calculate the inter-frame motion vector to compensate for dynamic blur and deformation; 2) Temporal data alignment: Multi-frame model alignment is achieved through feature point matching and the ICP algorithm; 3) 4D model generation: Integrates temporal point cloud and texture data to generate a 4D model containing motion sequences, supporting model backtracking at any time point.

6. The image acquisition method based on three-dimensional modeling according to claim 1, characterized in that: The multimodal sensor data fusion processing in step A: Point cloud registration: Improve the ICP algorithm by adding a normal vector constraint term, the formula of which is: E=∑||p i -T·q j || 2 +λ∑||n i -R·n j || 2 Texture mapping: seams are eliminated based on the Poisson fusion algorithm, supporting super-resolution reconstruction.

7. The image acquisition method based on three-dimensional modeling according to claim 1, characterized in that: Step A also includes a data processor and a storage module. The data processor is equipped with the NVIDIA-Jetson-AGX-Orin development kit, which integrates a GPU and AI acceleration core, and is responsible for real-time data fusion, scene analysis, modeling calculation and storage management. In step A, the data collected by the multimodal sensor is transmitted to the data processing unit via GigE, USB 3.0 or CAN bus. The storage module adopts an NVMe-SSD array, which provides high-speed data writing and reading capabilities and supports local storage of massive 3D model data.

Images