Dynamic data acquisition method for high-precision three-dimensional reconstruction

Abstract

The invention discloses a high-precision three-dimensional reconstruction-oriented data dynamic acquisition method, which comprises the following steps of: selecting an initial shooting mode from a preset shooting mode pool of an imaging system, acquiring an image sequence for a target in a simulation environment in the initial shooting mode, and performing three-dimensional reconstruction to obtain a reconstructed point cloud model; performing quality evaluation on the reconstruction point cloud model of the target in the simulation environment based on the real point cloud model of the target to obtain a quality evaluation score; and inputting the reconstruction point cloud model of the target in the simulation environment, the quality evaluation score and the environment prior information into a prediction network, taking the quality evaluation score as an award to construct a strategy gradient loss function to carry out iterative optimization of the prediction network, and determining a final optimal shooting mode in a shooting mode pool. The method is used for target image sequence acquisition and three-dimensional reconstruction of an imaging system in an actual environment. According to the method, the three-dimensional structure of the scene can still be accurately reconstructed under the condition of many interferences, and high-quality dynamic acquisition of image data is realized.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a method for dynamically acquiring data for high-precision three-dimensional reconstruction. Background Art

[0002] With the rapid development of aerospace technology, the number and types of space debris (such as satellites, space stations, etc.) are increasing, and the detection, identification and tracking of space debris have become particularly important. 3D reconstruction technology can restore the 3D geometric structure of the target from 2D optical image data, providing a more intuitive and comprehensive perspective for understanding and analyzing space debris, and plays an important role in space situational awareness, spacecraft design and manufacturing, and deep space exploration.

[0003] 3D reconstruction is the process of estimating the camera pose and reconstructing the 3D structure based on a series of optical images taken from different perspectives. The quality of the 3D reconstruction result depends on the quality of the input image, which in turn depends on the shooting mode, that is, the shooting parameters during shooting. For space debris targets, the controllable parameters mainly include turntable angle, exposure time, number of exposures, etc. However, current technology cannot obtain good 3D results for image data under any conditions. When the image data does not cover enough perspectives, the overlap between images is insufficient, or the exposure is extreme, a complete and accurate 3D model is usually not obtained. Although the optimal shooting mode can be obtained by testing all combinations of simulated data in actual applications, this not only takes up a lot of storage resources but is also very inefficient, and the shooting mode set based on experience is not reliable. Summary of the invention

[0004] The purpose of the present invention is to provide a method for dynamic data acquisition for high-precision three-dimensional reconstruction, which can accurately reconstruct the three-dimensional structure of a scene under many interference conditions and realize high-quality dynamic acquisition of image data.

[0005] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0006] A method for dynamically acquiring data for high-precision three-dimensional reconstruction, comprising:

[0007] An initial shooting mode is selected from a shooting mode pool preset by the imaging system, and an image sequence is acquired for a target in a simulated environment under the initial shooting mode and three-dimensional reconstruction is performed to obtain a reconstructed point cloud model;

[0008] Performing a quality assessment on the reconstructed point cloud model of the target in the simulation environment based on the real point cloud model of the target to obtain a quality assessment score;

[0009] The reconstructed point cloud model of the target in the simulated environment, the quality assessment score and the environmental prior information are input into the prediction network. The policy gradient loss function is constructed with the quality assessment score as the reward to iteratively optimize the prediction network. The final optimal shooting mode is determined in the shooting mode pool, which is used for the imaging system to acquire the image sequence and reconstruct the three-dimensional target in the actual environment.

[0010] Furthermore, the shooting mode pool stores a plurality of shooting modes, each of which is used by the imaging system to shoot a target to form a corresponding image sequence; the shooting mode is represented by a triplet: ,in represents the i-th shooting mode, represents the angle of each rotation of the turntable carrying the imaging system in the i-th shooting mode, Indicates exposure time and exposure times respectively; in each shooting mode, the turntable starts from the initial position within its rotation range and rotates each time And follow The target is photographed to form a corresponding image sequence under the photographing mode.

[0011] Furthermore, the process of performing three-dimensional reconstruction is:

[0012] (1) building a simulation environment to simulate the real state of the target; constructing models of the target and the imaging system in the simulation environment and simulating the motion trajectory of the target; the imaging system images the target according to the selected initial shooting mode to obtain the image sequence; and obtaining environmental prior information of the target during shooting by the imaging system;

[0013] (2) For each image in the image sequence, extract the feature points of the image;

[0014] (3) Perform brute force matching on the images in the image sequence, obtain the correspondence between the feature points of the two images, and form a scene graph with the images as nodes and the correspondence between the feature points as edges;

[0015] (4) Select a pair of images with the most matching feature points for initial reconstruction, solve the position and posture of the camera in the imaging system, perform triangulation on the matching feature points to obtain the corresponding 3D point cloud, and finally use bundle adjustment to optimize the camera position and posture and generate a 3D point cloud model;

[0016] (5) Add other images in the image sequence for image registration. During image registration, the camera pose corresponding to the newly added image is solved using the PnP algorithm. The feature points in the newly added image are triangulated to obtain three-dimensional points and added to the three-dimensional point cloud model to continuously improve the three-dimensional point cloud model. At the same time, local bundle adjustment and outlier filtering are performed. Repeat this step until all other images in the image sequence have completed image registration, and obtain a reconstructed point cloud model of the target, which is recorded as the reconstructed point cloud model.

[0017] Furthermore, the environmental prior information includes the corresponding sunlight angle, the distance between the target and the imaging system, and the moving speed and angular velocity of the target.

[0018] Furthermore, the quality of the reconstructed point cloud model of the target in the simulation environment is evaluated based on the real point cloud model of the target, and a quality evaluation score is obtained, including:

[0019] Construct a real point cloud model of the target; calculate the distance from each 3D point in the real point cloud model to the reconstructed point cloud model, and the distance from each 3D point in the reconstructed point cloud model to the real point cloud model;

[0020] Calculate quality assessment scores, including accuracy and completeness; accuracy is defined as the average distance between each 3D point in the reconstructed point cloud model and the corresponding 3D point in the real point cloud model, and completeness is defined as the average distance between each 3D point in the real point cloud model and the corresponding 3D point in the reconstructed point cloud model.

[0021] Furthermore, the prediction network includes a point cloud segmentation network, a first multi-layer perceptron, a second multi-layer perceptron and a shooting mode confidence prediction unit; and the weights of the point cloud segmentation network, the first multi-layer perceptron and the second multi-layer perceptron are randomly initialized.

[0022] Furthermore, the point cloud segmentation network is used to extract global features from the reconstructed point cloud model of the target;

[0023] The first multi-layer perceptron extracts features of the environmental prior information corresponding to the image sequence used to construct the reconstructed point cloud model to obtain environmental features;

[0024] The second multi-layer perceptron performs feature mapping on the quality assessment score of the reconstructed point cloud model to obtain high-dimensional features;

[0025] The global features extracted by the point cloud segmentation network, the environmental features extracted by the first multi-layer perceptron, and the high-dimensional features mapped by the second multi-layer perceptron are combined into a joint feature vector and input into a shooting mode confidence prediction unit, and the shooting mode confidence prediction unit is used to obtain an original score;

[0026] The original scores are normalized and converted into probability distributions to obtain the confidence of each shooting mode. The current optimal shooting mode is determined based on the confidence and the prediction network is trained iteratively.

[0027] Furthermore, the training iteration of the prediction network includes:

[0028] Taking the current optimal shooting mode as the reselected initial shooting mode, after acquiring an image sequence of the target and performing three-dimensional reconstruction, the current optimal shooting mode is iteratively calculated and the prediction network is trained until the accuracy and completeness of the quality assessment score meet the corresponding preset thresholds, or the preset maximum number of iterations is reached, and the current optimal shooting mode is output as the final optimal shooting mode.

[0029] Furthermore, the strategy gradient loss function is constructed with the quality evaluation score as the reward, which is expressed as:

[0030] ;

[0031] in, , Respectively represent the accuracy and completeness of the quality assessment score. is the confidence of the current optimal shooting mode a; during training, the weight of the prediction network is updated through the gradient descent algorithm.

[0032] An imaging system utilizes the data dynamic acquisition method for high-precision three-dimensional reconstruction to determine the final optimal shooting mode, thereby acquiring an image sequence of a target and performing three-dimensional reconstruction.

[0033] A terminal device comprises a memory, a processor and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the method for dynamically acquiring data for high-precision three-dimensional reconstruction is implemented.

[0034] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the method for dynamically acquiring data for high-precision three-dimensional reconstruction is implemented.

[0035] Compared with the prior art, the present invention has the following technical features:

[0036] 1. Active perception optimization mechanism: For the first time, deep learning is combined with 3D reconstruction evaluation indicators to form a closed-loop system of "shooting-reconstruction-evaluation-optimization".

[0037] 2. Multimodal feature fusion: Combine point cloud geometric features, environmental priors and historical scores to enhance the scene adaptability of pattern prediction.

[0038] 3. End-to-end optimization: MLP directly learns the mapping from features to shooting patterns without the need for manual design of rules.

[0039] 4. Dynamic adaptability: The confidence vector reflects the applicability of each shooting mode in the current scene, and the system can dynamically select the optimal mode.

[0040] 5. Simulation-to-reality migration framework: Reduce data collection costs through virtual simulation and use domain adaptation technology to solve the reality gap problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process of the present invention;

[0042] Figure 2 A sequence of images obtained by photographing a target in an initial photographing mode in one embodiment of the present invention;

[0043] Figure 3 In one embodiment of the present invention, a point cloud model of a target is reconstructed using an image sequence acquired in an initial shooting mode;

[0044] Figure 4 A visualization effect diagram of a truth model of a target in one embodiment of the present invention;

[0045] Figure 5 A target three-dimensional reconstruction effect diagram of the final optimal shooting mode in one embodiment of the present invention;

[0046] Figure 6 is a flowchart of three-dimensional reconstruction in one embodiment of the present invention;

[0047] Figure 7 A flowchart of the process of optimizing the shooting mode in the present invention;

[0048] Figure 8 This is the architecture diagram of the point cloud segmentation network PointNet. DETAILED DESCRIPTION

[0049] The present invention provides a method for dynamically acquiring data for high-precision three-dimensional reconstruction, the method comprising:

[0050] Step 1: Select an initial shooting mode from a preset shooting mode pool of the imaging system, acquire an image sequence for a target in a simulated environment under the initial shooting mode, and perform three-dimensional reconstruction to obtain a reconstructed point cloud model.

[0051] Step 1.1, construction of shooting pattern pool.

[0052] The shooting mode pool stores a plurality of shooting modes, each of which is used by the imaging system to shoot a target to form a corresponding image sequence; the shooting mode is represented by a triplet: ,in represents the i-th shooting mode, represents the angle of each rotation of the turntable carrying the imaging system in the i-th shooting mode, Represent the corresponding exposure time and exposure times respectively.

[0053] In each shooting mode, the turntable starts from the initial position within its rotation range and rotates each time And follow The target is photographed to form a corresponding image sequence under the photographing mode.

[0054] When constructing a shooting mode pool, discrete sampling can be performed from the rotation range of the turntable, the exposure time range of the imaging system, and the exposure number range, so as to combine and generate multiple groups of shooting modes.

[0055] Step 1.2, selection of initial shooting mode.

[0056] Select any one shooting mode from the shooting mode pool as the initial shooting mode, or select a shooting mode with better imaging effect as the initial shooting mode based on experience.

[0057] Step 1.3: 3D reconstruction of the target.

[0058] The process of 3D reconstruction using image sequences is as follows:

[0059] (1) Building a simulation environment to simulate the real state of the target; the simulation environment can be built by, for example, software such as Blender, building models of the target and the imaging system in the simulation environment and simulating the motion trajectory of the target; the imaging system images the target according to the selected initial shooting mode to obtain the image sequence; during the shooting process of the imaging system, obtaining the environmental prior information of the target during shooting, including the corresponding sunlight angle, the distance between the target and the imaging system, the motion speed and angular velocity of the target.

[0060] (2) For each image in the image sequence, extract the SIFT feature points of the image.

[0061] (3) The images in the image sequence are violently matched pairwise to obtain the correspondence between the feature points of the two images, thus forming a scene graph with images as nodes and the correspondence between the feature points as edges.

[0062] (4) A pair of images with the most matching feature points are selected for initial reconstruction, the position and posture of the camera in the imaging system are solved, and the matching feature points are triangulated to obtain the corresponding three-dimensional point cloud. Finally, the bundle adjustment is used to optimize the camera position and generate a three-dimensional point cloud model.

[0063] (5) Add other images in the image sequence (except for a pair of matching images) for image registration. During image registration, the camera pose corresponding to the newly added image is solved using the PnP algorithm, and the feature points in the newly added image are triangulated to obtain three-dimensional points and added to the three-dimensional point cloud model to continuously improve the three-dimensional point cloud model. At the same time, local bundle adjustment and outlier filtering are performed. Repeat this step until all other images in the image sequence have completed image registration, and obtain the reconstructed point cloud model of the target, which is recorded as the reconstructed point cloud model R.

[0064] Step 2: Perform a quality assessment on the reconstructed point cloud model of the target in the simulated environment based on the real point cloud model of the target to obtain a quality assessment score.

[0065] Step 2.1: Construct a real point cloud model of the target, record the real point cloud model as D, and calculate the distance from each 3D point d in the real point cloud model D to the reconstructed point cloud model R. for:

[0066] ;

[0067] in, is the position of the three-dimensional point d, is the position of the three-dimensional point r, Calculates the distance between two locations.

[0068] The distance between each 3D point r in the reconstructed point cloud model R and the real point cloud model D for:

[0069] .

[0070] In step 2.2, quality assessment scores are calculated, including accuracy and completeness.

[0071] Accuracy is defined as the average distance between each 3D point in the reconstructed point cloud model R and the corresponding 3D point in the real point cloud model D. ,Right now:

[0072] ;

[0073] in Represents the number of 3D points in the reconstructed point cloud model R.

[0074] Integrity is defined as the average distance between each 3D point in the real point cloud model D and the corresponding 3D point in the reconstructed point cloud model R ,Right now:

[0075] ;

[0076] in Represents the number of 3D points in the real point cloud model D.

[0077] Therefore, the lower the quality assessment score, the better the quality of the 3D reconstruction.

[0078] Step 3: Input the reconstructed point cloud model of the target in the simulated environment, the quality assessment score, and the environmental prior information into the prediction network, and use the quality assessment score as a reward to construct a policy gradient loss function to iteratively optimize the prediction network. The final optimal shooting mode is determined in the shooting mode pool, which is used for the imaging system to acquire the image sequence and reconstruct the three-dimensional target in the actual environment.

[0079] Step 3.1, construct a prediction network; the prediction network includes a point cloud segmentation network PointNet, a first multi-layer perceptron (MLP), a second multi-layer perceptron (MLP) and a shooting mode confidence prediction unit; the weights of the point cloud segmentation network PointNet, the first multi-layer perceptron, and the second multi-layer perceptron are randomly initialized.

[0080] In step 3.2, use the point cloud segmentation network PointNet to perform global feature extraction (256 dimensions) on the reconstructed point cloud model of the target.

[0081] Step 3.3, using the first multi-layer perceptron to extract features from the environmental prior information corresponding to the image sequence used to construct the reconstructed point cloud model, to obtain environmental features.

[0082] The design purpose of this step is to automatically learn the nonlinear relationship between environmental parameters (such as the joint influence of "sunlight angle, distance between target and imaging system" on light attenuation) and to make the output dimension consistent with the point cloud feature dimension extracted by the point cloud segmentation network PointNet, so as to facilitate subsequent fusion.

[0083] Among them, the structure of the first multi-layer perceptron is as follows:

[0084] The input layer is a vector composed of parameters in the environmental prior information. In this embodiment, the dimension of the vector is 4.

[0085] There are two hidden layers. The first layer is a fully connected layer (4 dimensions → 64 dimensions) with the activation function ReLU, which introduces nonlinear expression capabilities. The first layer is also a fully connected layer (64 dimensions → 128 dimensions) with the activation function ReLU, which extracts high-order feature interactions.

[0086] The output layer is a fully connected layer (128-dimensional → 256-dimensional feature vector), which finally outputs the extracted environmental feature vector.

[0087] In step 3.4, the second multi-layer perceptron performs feature mapping on the quality assessment score of the reconstructed point cloud model to obtain high-dimensional features.

[0088] This step is designed to learn the nonlinear relationship between completeness and accuracy (e.g., high completeness but low accuracy may mean more noise) and to make the output dimension consistent with the dimension of the global feature and environment feature vectors for subsequent fusion.

[0089] Among them, the structure of the second multi-layer perceptron is as follows:

[0090] The input layer is a vector composed of accuracy and completeness in the quality assessment score. In this embodiment, the dimension of the vector is 2.

[0091] There are two hidden layers. The first layer is a fully connected layer (2D→64D) with ReLU as the activation function, which introduces nonlinear expression capability. The second layer is also a fully connected layer (64D→128D) with ReLU as the activation function, which extracts high-order feature interactions.

[0092] The output layer is a fully connected layer (128-dimensional → 256-dimensional feature vector), which finally outputs the mapped high-dimensional features.

[0093] Through multiple layers of nonlinear transformations, MLP is able to learn the complex relationships between the three types of features. For example, low completeness may mean that the shooting angle needs to be increased; low accuracy may mean that the exposure time or rotation angle needs to be adjusted.

[0094] In step 3.5, the global features extracted by the point cloud segmentation network, the environmental features extracted by the first multi-layer perceptron, and the high-dimensional features mapped by the second multi-layer perceptron are combined into a joint feature vector and input into the shooting mode confidence prediction unit (MPL), and the original score is obtained using the shooting mode confidence prediction unit.

[0095] The structure of the shooting mode confidence prediction unit is:

[0096] The input layer is a joint feature vector = [global feature, environmental feature, high-dimensional feature], which is a 640-dimensional feature in this embodiment.

[0097] There are two hidden layers. The first layer is a fully connected layer (640 dimensions → 512 dimensions) with ReLU activation function to extract high-order feature interactions. The second layer is also a fully connected layer (512 dimensions → 256 dimensions) with ReLU activation function to further abstract features.

[0098] The output layer is a fully connected layer (256-dimensional → dimension), is the number of shooting modes in the shooting mode pool, and finally we get a The dimensional vector s is used as the original score.

[0099] Step 3.6, normalize the original scores and convert them into probability distribution:

[0100] ;

[0101] in, is the confidence of the i-th shooting mode, ; express The i-th and j-th features in the dimensional vector s.

[0102] The current optimal shooting mode a is expressed as:

[0103] .

[0104] Step 3.7, taking the current optimal shooting mode a as the reselected initial shooting mode, after acquiring the image sequence of the target and performing three-dimensional reconstruction, iterate steps 2 and 3 and train the prediction network until the accuracy and completeness of the quality assessment score meet the corresponding preset thresholds, or stop after reaching the preset maximum number of iterations, and output the current optimal shooting mode as the final optimal shooting mode.

[0105] When the prediction network is trained, the policy gradient loss function L is constructed with the quality evaluation score as the reward, which is expressed as follows:

[0106] ;

[0107] in, , Respectively represent the accuracy and completeness of the quality assessment score. is the confidence of the current optimal shooting mode a; during training, the weight of the prediction network is updated through the gradient descent algorithm.

[0108] The final optimal shooting mode is written into the shooting program of the imaging system, and then the actual shooting can be carried out, and the three-dimensional reconstruction is carried out after the image sequence is obtained. The imaging system of the present invention can be mounted on a platform such as a road-based, space-based or sea-based platform.

[0109] In one embodiment of the present invention, a low-orbit satellite (orbital altitude 500km) is used as a target, and the specific application process of the present invention is as follows:

[0110] Taking into account factors such as the inclination of the low-orbit satellite, the latitude of the ground station, and horizon obstruction, the rotation range of the imaging system turntable in this embodiment is: azimuth angle is 120°-180°, pitch angle is 80°-100°, exposure time is 1-500 milliseconds, and exposure times are 0.1-10 times / second; based on these parameters, multiple groups of shooting modes are discretely formed to construct a shooting mode pool.

[0111] The initial shooting mode in this embodiment is selected as follows: the angle of each rotation of the turntable is 0.5°, the exposure time is 10 milliseconds, and the number of exposures is 1 time / second; the image sequence obtained in this shooting mode is as follows Figure 2 As shown, the reconstructed point cloud model of the target is as follows Figure 3 As shown, it can be seen that the reconstruction result is not accurate.

[0112] Visualization of the real point cloud model of the target Figure 4 As shown; Comparison between the reconstructed point cloud model and the real point cloud model, calculation integrity The accuracy is 0.36 mm. It is 1.77 mm.

[0113] like Figure 6 and Figure 7 , the reconstructed point cloud model of the target in the simulated environment, the quality assessment score and the environmental prior information are input into the prediction network, where the environmental prior information is specifically set as: sunlight angle 30°, distance between the target and the imaging system 500 km, target movement speed 7.5 km / s and angular velocity 0.1° / s. The current optimal shooting mode is output by the prediction network and enters the next round of iteration; until the final iteration is completed, the final optimal shooting mode is output as follows: the turntable rotates at an angle of 1.4° each time, the exposure time is 12 milliseconds, and the number of exposures is 1 time / second.

[0114] The image sequence obtained by the final optimal shooting mode is used for 3D reconstruction, and the effect is as follows Figure 5 As shown, the completeness The accuracy is 0.23 mm. is 0.25 mm.

[0115] By comparison, it can be found that the data obtained by the shooting mode obtained by the method of the present invention can achieve a better reconstruction effect than the experience, and compared with the manual method, the process automation is realized.

[0116] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for dynamic data acquisition for high-precision three-dimensional reconstruction, characterized in that: include: An initial shooting mode is selected from a shooting mode pool preset by the imaging system, and an image sequence is acquired for a target in a simulated environment under the initial shooting mode and three-dimensional reconstruction is performed to obtain a reconstructed point cloud model; Performing a quality assessment on the reconstructed point cloud model of the target in the simulation environment based on the real point cloud model of the target to obtain a quality assessment score; The reconstructed point cloud model of the target in the simulated environment, the quality assessment score and the environmental prior information are input into the prediction network. The policy gradient loss function is constructed with the quality assessment score as the reward to iteratively optimize the prediction network. The final optimal shooting mode is determined in the shooting mode pool, which is used for the imaging system to acquire the image sequence and reconstruct the three-dimensional target in the actual environment.

2. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 1, characterized in that: The shooting mode pool stores a plurality of shooting modes, each of which is used by the imaging system to shoot a target to form a corresponding image sequence; the shooting mode is represented by a triplet ,in represents the i-th shooting mode, represents the angle of each rotation of the turntable carrying the imaging system in the i-th shooting mode, Indicates exposure time and exposure times respectively; in each shooting mode, the turntable starts from the initial position within its rotation range and rotates each time And follow The target is photographed to form a corresponding image sequence under the photographing mode.

3. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 1, characterized in that: The process of performing three-dimensional reconstruction is as follows: (1) building a simulation environment to simulate the real state of the target; constructing models of the target and the imaging system in the simulation environment and simulating the motion trajectory of the target; the imaging system images the target according to the selected initial shooting mode to obtain the image sequence; and obtaining environmental prior information of the target during shooting by the imaging system; (2) For each image in the image sequence, extract the feature points of the image; (3) Perform brute force matching on the images in the image sequence, obtain the correspondence between the feature points of the two images, and form a scene graph with the images as nodes and the correspondence between the feature points as edges; (4) Select a pair of images with the most matching feature points for initial reconstruction, solve the position and posture of the camera in the imaging system, perform triangulation on the matching feature points to obtain the corresponding 3D point cloud, and finally use bundle adjustment to optimize the camera position and posture and generate a 3D point cloud model; (5) Add other images in the image sequence for image registration. During image registration, the camera pose corresponding to the newly added image is solved using the PnP algorithm. The feature points in the newly added image are triangulated to obtain three-dimensional points and added to the three-dimensional point cloud model to continuously improve the three-dimensional point cloud model. At the same time, local bundle adjustment and outlier filtering are performed. Repeat this step until all other images in the image sequence have completed image registration, and obtain a reconstructed point cloud model of the target, which is recorded as the reconstructed point cloud model.

4. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 1, characterized in that: The quality of the reconstructed point cloud model of the target in the simulation environment is evaluated based on the real point cloud model of the target, and the quality evaluation score is obtained, including: Construct a real point cloud model of the target; calculate the distance from each 3D point in the real point cloud model to the reconstructed point cloud model, and the distance from each 3D point in the reconstructed point cloud model to the real point cloud model; Calculate quality assessment scores, including accuracy and completeness; accuracy is defined as the average distance between each 3D point in the reconstructed point cloud model and the corresponding 3D point in the real point cloud model, and completeness is defined as the average distance between each 3D point in the real point cloud model and the corresponding 3D point in the reconstructed point cloud model.

5. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 4, characterized in that: The prediction network includes a point cloud segmentation network, a first multi-layer perceptron, a second multi-layer perceptron and a shooting mode confidence prediction unit; the weights of the point cloud segmentation network, the first multi-layer perceptron and the second multi-layer perceptron are randomly initialized.

6. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 5, characterized in that: The point cloud segmentation network is used to extract global features from the reconstructed point cloud model of the target; The first multi-layer perceptron extracts features of the environmental prior information corresponding to the image sequence used to construct the reconstructed point cloud model to obtain environmental features; The second multi-layer perceptron performs feature mapping on the quality assessment score of the reconstructed point cloud model to obtain high-dimensional features; The global features extracted by the point cloud segmentation network, the environmental features extracted by the first multi-layer perceptron, and the high-dimensional features mapped by the second multi-layer perceptron are combined into a joint feature vector and input into a shooting mode confidence prediction unit, and the shooting mode confidence prediction unit is used to obtain an original score; The original scores are normalized and converted into probability distributions to obtain the confidence of each shooting mode. The current optimal shooting mode is determined based on the confidence and the prediction network is trained iteratively.

7. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 6, characterized in that: The training iteration of the prediction network includes: Taking the current optimal shooting mode as the reselected initial shooting mode, after acquiring an image sequence of the target and performing three-dimensional reconstruction, the current optimal shooting mode is iteratively calculated and the prediction network is trained until the accuracy and completeness of the quality assessment score meet the corresponding preset thresholds, or the preset maximum number of iterations is reached, and the current optimal shooting mode is output as the final optimal shooting mode.

8. The method for dynamic data acquisition for high-precision three-dimensional reconstruction according to claim 6, characterized in that: The policy gradient loss function is constructed with the quality evaluation score as the reward, expressed as: ； in, , Respectively represent the accuracy and completeness of the quality assessment score. is the confidence of the current optimal shooting mode a; during training, the weight of the prediction network is updated through the gradient descent algorithm.

9. An imaging system, characterized in that: The imaging system uses the data dynamic acquisition method for high-precision three-dimensional reconstruction according to any one of claims 1 to 8 to determine the final optimal shooting mode, thereby acquiring image sequences and performing three-dimensional reconstruction of the target.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that: When the processor executes the computer program, the method for dynamic data acquisition for high-precision three-dimensional reconstruction according to any one of claims 1 to 8 is implemented.

Images