Monocular three-dimensional reconstruction method and device based on intelligent anchor point selection and medium

Abstract

The invention discloses a monocular three-dimensional reconstruction method and device based on intelligent anchor point selection and a medium, and relates to the technical field of computer vision and three-dimensional perception, and the method comprises the steps: obtaining a single RGB image, inputting the single RGB image into a monocular depth model, and generating a relative depth map and an auxiliary feature map; constructing a uniform grid as candidate anchor points, extracting feature values of the candidate anchor points, and calculating a comprehensive quality score; an initial anchor point set is obtained and divided into initial anchor point subsets, and comprehensive optimization scores of the initial anchor point subsets are calculated; determining an optimal anchor point subset, and calculating a global scale factor; and calibrating the relative depth map, generating an absolute depth map, and generating three-dimensional point cloud data of a physical scene in the RGB image based on an internal reference matrix of the monocular camera so as to perform three-dimensional reconstruction on the RGB image. Scale recovery and global calibration are carried out through intelligently selected anchor points, a more accurate absolute depth map can be obtained, and then a three-dimensional point cloud with a real physical size is generated through geometric back projection.

Description

Monocular 3D Reconstruction Methods, Equipment, and Media Based on Intelligent Anchor Point Selection Technical Field

[0001] This application relates to the fields of computer vision and 3D perception technology, specifically to a monocular 3D reconstruction method, device, and medium based on intelligent anchor point selection. Background Technology

[0002] With the widespread application of computer vision and 3D perception technologies in fields such as industrial inspection, robot navigation, and AR, monocular 3D reconstruction technology has gradually become a research hotspot. This technology primarily uses a pre-trained monocular depth estimation model to infer a dense relative depth map from a single input image, thereby initially constructing the 3D structure of the scene. However, relying solely on a single image inherently suffers from scale ambiguity; the resulting depth map only reflects the relative distances between pixels, lacking absolute depth information in real physical units. This makes it difficult to meet the needs of scenarios requiring strict real-scale data, such as industrial measurement and precise grasping.

[0003] To achieve scale calibration, existing methods require the introduction of known absolute depth anchor points to calibrate the depth map. However, this method requires operators to manually select points and input the actual depth values, which relies on the operator's experience and has drawbacks such as strong subjectivity, low efficiency, and insufficient reliability. This can easily lead to large scale fitting errors during the depth calibration process. Furthermore, if the selected anchor points are collinear, have similar depths, or are located on the surface of dynamic objects, it can also cause scale factor estimation to fail and 3D point cloud distortion, seriously affecting the accuracy of 3D reconstruction. Summary of the Invention

[0004] To address the aforementioned issues, this application proposes a monocular 3D reconstruction method based on intelligent anchor point selection, comprising: acquiring a single RGB image captured by a monocular camera, inputting it into a pre-trained monocular depth model to generate a relative depth map of the RGB image, and extracting image auxiliary features of the RGB image to generate an auxiliary feature map; constructing a uniform grid in the relative depth map, using the grid center of each grid as candidate anchor points, extracting the feature values ​​corresponding to each candidate anchor point in the auxiliary feature map, and calculating a comprehensive quality score for the candidate anchor points; filtering the candidate anchor points based on the comprehensive quality score to obtain an initial anchor point set, dividing the initial anchor point set into multiple initial anchor point subsets, and based on the... The initial pixel coordinates of each initial anchor point in the RGB image and its relative depth value in the relative depth map are used to calculate a comprehensive optimization score for the initial anchor point subset using a predefined combinatorial optimization function. Based on the comprehensive optimization score, an optimal anchor point subset is determined. A global scale factor is calculated based on the true absolute depth value and relative depth value of each optimal anchor point in the optimal anchor point subset. The relative depth map is calibrated using the global scale factor to generate an absolute depth map. Based on the intrinsic parameter matrix of the monocular camera, inverse perspective projection is performed on the absolute depth map to generate 3D point cloud data of the physical scene in the RGB image for 3D reconstruction of the RGB image.

[0005] On the other hand, this application also proposes a monocular 3D reconstruction device based on intelligent anchor point selection, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a monocular 3D reconstruction method based on intelligent anchor point selection as described in the above example.

[0006] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as: a monocular 3D reconstruction method based on intelligent anchor point selection as described in the above example.

[0007] This application proposes a monocular 3D reconstruction method based on intelligent anchor point selection, which brings the following benefits: by automatically evaluating the quality of candidate points through multi-feature fusion, and using a combined optimization function to select the optimal subset of anchor points in terms of depth difference and spatial distribution, the automation level and robustness of the scale calibration process are significantly improved without human intervention. This effectively solves the problems of strong subjectivity, low efficiency and insufficient reliability caused by relying on manual selection of depth reference points in traditional monocular 3D reconstruction methods.

[0008] Furthermore, by performing scale restoration and global calibration based on intelligently selected anchor points, a more accurate absolute depth map can be obtained. Then, a 3D point cloud with real physical dimensions can be generated through geometric back projection, which enhances the accuracy and practicality of metric-level 3D reconstruction of a single image and provides a reliable data foundation for subsequent applications such as measurement, navigation, and interaction. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 is a schematic flowchart of a monocular 3D reconstruction method based on intelligent anchor point selection according to an embodiment of this application; Figure 2 is a schematic diagram of a monocular 3D reconstruction device based on intelligent anchor point selection according to an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0012] As shown in Figure 1, this application embodiment provides a monocular 3D reconstruction method based on intelligent anchor point selection, including: S101, acquiring a single RGB image captured by a monocular camera, inputting it into a pre-trained monocular depth model, generating a relative depth map of the RGB image, and extracting image auxiliary features of the RGB image to generate an auxiliary feature map.

[0013] Acquire a single RGB image captured by a monocular camera, and pre-acquire or calibrate the intrinsic parameter matrix K of the monocular camera. ,in, These represent the focal lengths of the monocular camera along the x and y axes, respectively. The coordinates of the main point.

[0014] A single RGB image is input into a pre-trained monocular depth estimation model, such as the MiDaS model or the DepthAnything model. The encoder of the model performs multi-layer convolution and downsampling on the input RGB image to extract multi-scale visual features containing scene geometric information. These multi-scale visual features are then input into the decoder of the model. The decoder gradually recovers the spatial resolution of the feature map through a series of feature fusion and upsampling operations, and finally regresses a scalar value representing the relative distance of each pixel to the camera. The output is a relative depth map with the same spatial size as the input RGB image, and simultaneously outputs the estimation uncertainty data of the depth value of each pixel during the depth estimation process.

[0015] It should be noted that the relative depth map only preserves the relative depth relationships within the scene, and its numerical range is normalized and does not have a true physical scale.

[0016] Furthermore, auxiliary image features of the RGB image are extracted to generate an auxiliary feature map, which includes a semantic segmentation map, an edge response map, a texture energy map, and a depth confidence map.

[0017] Specifically, a pre-trained semantic segmentation model is invoked, and the RGB image is input into the semantic segmentation model to perform pixel-level classification of the image and output a semantic segmentation map. Each pixel in the semantic segmentation map is assigned a semantic category label.

[0018] Edge detection is performed on RGB images. Specifically, the image is first converted to grayscale, and then a pre-defined edge detection operator such as Sobel, Canny, or Prewitt is used to perform convolution operations on the grayscale image to calculate the gradient magnitude at each pixel location, thereby generating an edge response map to characterize the intensity of object contours and texture boundaries in the image.

[0019] Using preset texture analysis algorithms, such as calculating the Local Binary Pattern (LBP) or calculating the gray-level variance within an image patch, local texture statistical analysis is performed on the RGB image to extract texture description feature values ​​in the neighborhood of each pixel and generate a texture energy map to characterize the texture richness and complexity of each region of the image.

[0020] Uncertainty data from the depth estimation process, such as the variance or logarithmic variance of the predicted depth, is acquired from the synchronous output of the monocular depth model. This uncertainty data is then transformed into a confidence score characterizing the reliability of the depth prediction through processing methods such as taking the reciprocal, negative exponential transformation, or direct normalization. This score is then used to generate a depth confidence map with dimensions consistent with the relative depth map. A higher value indicates a more reliable depth estimate at that location.

[0021] S102. Construct a uniform grid in the relative depth map, take the grid center of each grid as a candidate anchor point, extract the feature value of each candidate anchor point in the auxiliary feature map, and calculate the comprehensive quality score of the candidate anchor point.

[0022] A uniformly distributed two-dimensional grid is constructed on the pixel coordinate system of the relative depth map. The width and height of each grid cell are calculated. All grid cells are traversed, and the pixel coordinates of the grid center point in the image are calculated. Each grid center is used as a candidate anchor point to obtain a set of candidate anchor points.

[0023] For each candidate anchor point The corresponding feature values ​​are extracted from the generated multi-class auxiliary feature maps. Specifically, the semantic segmentation map is queried based on the coordinates of the candidate anchor point to obtain the predefined semantic category identifier to which the pixel at that location belongs. ; Query the edge response map based on the coordinates of the candidate anchor point to obtain the edge intensity value of the pixel at that location. ; Query the texture energy map based on the coordinates of the candidate anchor point to obtain the texture energy value of the pixel at that location. ; Query the depth confidence map based on the coordinates of the candidate anchor point to obtain the depth confidence value of the pixel at that location. .

[0024] For each candidate anchor point, the formula is used: The calculation yields a comprehensive quality score for each candidate anchor point, where... This is an indicator function.

[0025] S103. Based on the comprehensive quality score, the candidate anchor points are screened to obtain an initial anchor point set. The initial anchor point set is divided into multiple initial anchor point subsets. Based on the initial pixel coordinates of each initial anchor point in the RGB image and the relative depth value in the relative depth map, the comprehensive optimization score of the initial anchor point subset is calculated through a predefined combination optimization function.

[0026] All candidate anchor points are sorted in descending order of their overall quality scores. Anchor points are selected from the sorted list according to a preset filtering rule. The final selected anchor points and all their associated information constitute the initial anchor point set C. The filtering rule can be either to directly select the top N candidate anchor points or to select all candidate anchor points with an overall quality score higher than a preset threshold.

[0027] From the initial anchor set C, a strategic approach is used to generate multiple subsets of initial anchor points to be evaluated. Each subset contains K anchor points. The generation methods may include: random sampling, repeatedly selecting K non-repeating anchor points from the initial anchor point set to form multiple subsets; and heuristic search, using greedy algorithms, simulated annealing, or other optimized search methods to explore different combinations of anchor points.

[0028] For each generated initial anchor point subset Through a predefined combinatorial optimization function: Perform calculations. Among them... Initial anchor subset, For depth difference items, For spatial distribution terms, This is a penalty for close-range contact.

[0029] Specifically, for the depth difference term, the relative depth values ​​of all anchor points in the initial anchor point subset are extracted. Through the formula: Calculate the average relative depth value using the formula: Calculate the variance, which indicates the degree of difference between relative depth values. The larger the variance, the wider the depth coverage within the subset and the stronger the constraint on scale calibration.

[0030] For the spatial distribution term, extract the image pixel coordinates of all anchor points in the initial anchor point subset, and then use the formula based on the image pixel coordinates: Calculate the convex hull area of ​​all anchor points, where m is the number of vertices of the convex hull. The larger the area, the more dispersed and uniform the anchor points are on the image plane, which is beneficial for correcting the perspective geometry of the entire image.

[0031] For the proximity penalty, iterate through all distinct anchor pairs in the initial anchor subset. Calculate the Euclidean distance in the pixel coordinate system: This yields the minimum distance among all anchor point pairs. .

[0032] S104. Based on the comprehensive optimization score, determine the optimal anchor point subset, and calculate the global scale factor based on the true absolute depth value and relative depth value of each optimal anchor point in the optimal anchor point subset.

[0033] The algorithm iterates through and compares the comprehensive optimization scores of all evaluated initial anchor point subsets, and determines the candidate subset with the highest comprehensive optimization score as the final optimal anchor point subset. It should be noted that the optimal anchor point subset is the best combination obtained after comprehensively considering multiple optimization objectives such as maximizing depth difference, widest spatial distribution, and avoiding point clustering.

[0034] Using external ranging devices, such as laser rangefinders or Time-of-Flight (ToF) sensors, the distance to each selected anchor point is measured at its corresponding physical location in the real scene, and the measured physical distance value is used as the true absolute depth value of the anchor point.

[0035] Alternatively, known dimensional information can be used for calculation. For example, if a selected anchor point is located on a standard object of known size or on a surface with known equations, its true absolute depth value can be calculated by combining the geometric relationship of the anchor point in the image with the camera model.

[0036] The relative depth value of each anchor point Its corresponding true absolute depth value Pair them up to form a set of data points. Using the set of relative depth values ​​of all optimal anchor points in the aforementioned optimal anchor point subset as the independent variable x and the set of true absolute depth values ​​of all optimal anchor points as the dependent variable y, a linear regression model is constructed: Where s is the global scale factor to be solved, and its physical meaning is the conversion coefficient between the relative depth value and the true absolute depth value.

[0037] The least squares method is used to fit the above linear regression model, and the global scale factor s is calculated to minimize the sum of squares error between the true absolute depth values ​​and the fitted values ​​of all anchor points. Specifically, for the formula: The optimal global scale factor s with the minimum value is obtained by solving the problem. Here, A is the optimal subset of anchor points. This represents the relative depth value of anchor point p in the output of the monocular depth model. This represents the true absolute depth of anchor point p in the real world.

[0038] S105. The relative depth map is calibrated using the global scale factor to generate an absolute depth map. Based on the intrinsic parameter matrix of the monocular camera, inverse perspective projection calculation is performed on the absolute depth map to generate three-dimensional point cloud data of the physical scene in the RGB image for three-dimensional reconstruction of the RGB image.

[0039] Iterate through each valid pixel position in the relative depth map. For the pixel position currently being processed, retrieve the relative depth value stored at that position from the relative depth map. The absolute depth value of the pixel is obtained by multiplying its relative depth value by the global scale factor. .

[0040] After calculating the absolute depth for each pixel, all the calculated absolute depth values ​​are arranged according to their corresponding original pixel positions to generate a new absolute depth map of the same size as the input image. Each pixel value in this map represents the actual physical distance from the camera's optical center to the corresponding point in the scene.

[0041] Iterate through each valid pixel position in the absolute depth map, and calculate the 3D coordinates of that pixel in the camera coordinate system based on the geometric relationship of the pinhole camera model. The calculation formula is as follows: , , The point on the image plane and its measured depth are used to calculate the specific endpoint coordinates of a spatial ray using camera intrinsic parameters.

[0042] The 3D coordinates of each valid pixel are collected, and the color value of the corresponding pixel in the original RGB image is associated with these 3D coordinates to color the point cloud. All collected 3D coordinates that may be associated with color information are organized and output according to a common 3D point cloud data format to form the final 3D point cloud data.

[0043] It should be noted that all pixels in a single RGB image are converted into its three-dimensional point cloud data in the camera coordinate system. The collection of point cloud data accurately represents the surface geometry of the physical scene at the time of shooting and can be directly used for volume measurement, dimension annotation, comparison with CAD models, or as a high-precision three-dimensional environment input for applications such as robot navigation and augmented reality.

[0044] This application automatically evaluates the quality of candidate points through multi-feature fusion and uses a combined optimization function to select the optimal subset of anchor points in terms of depth difference and spatial distribution. Thus, without the need for manual intervention, it significantly improves the automation level and robustness of the scale calibration process, effectively solving the problems of strong subjectivity, low efficiency and insufficient reliability caused by the reliance on manual selection of depth reference points in traditional monocular 3D reconstruction methods.

[0045] Furthermore, by performing scale restoration and global calibration based on intelligently selected anchor points, a more accurate absolute depth map can be obtained. Then, a 3D point cloud with real physical dimensions can be generated through geometric back projection, which enhances the accuracy and practicality of metric-level 3D reconstruction of a single image and provides a reliable data foundation for subsequent applications such as measurement, navigation, and interaction.

[0046] As shown in Figure 2, this application also proposes a monocular 3D reconstruction device based on intelligent anchor point selection, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform a monocular 3D reconstruction method based on intelligent anchor point selection as described in any of the above embodiments.

[0047] This application also provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as: a monocular 3D reconstruction method based on intelligent anchor point selection as described in any of the above embodiments.

[0048] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0049] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0050] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0054] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0055] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0056] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0057] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0058] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A monocular 3D reconstruction method based on intelligent anchor point selection, characterized in that, include: A single RGB image captured by a monocular camera is acquired and input into a pre-trained monocular depth model to generate a relative depth map of the RGB image. The auxiliary features of the RGB image are then extracted to generate an auxiliary feature map. A uniform grid is constructed in the relative depth map, and the center of each grid is used as a candidate anchor point. The feature values ​​corresponding to each candidate anchor point in the auxiliary feature map are extracted, and the comprehensive quality score of the candidate anchor points is calculated. Based on the comprehensive quality score, the candidate anchor points are filtered to obtain an initial anchor point set. The initial anchor point set is divided into multiple initial anchor point subsets. Based on the initial pixel coordinates of each initial anchor point in the RGB image and the relative depth value in the relative depth map, a comprehensive optimization score of the initial anchor point subset is calculated using a predefined combinatorial optimization function. According to the comprehensive optimization score, the optimal anchor point subset is determined. Based on the true absolute depth value and relative depth value of each optimal anchor point in the optimal anchor point subset, the global scale factor is calculated. The relative depth map is calibrated using the global scale factor to generate an absolute depth map. Based on the intrinsic parameter matrix of the monocular camera, inverse perspective projection calculation is performed on the absolute depth map to generate 3D point cloud data of the physical scene in the RGB image for 3D reconstruction of the RGB image.

2. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 1, characterized in that, The step of acquiring a single RGB image captured by a monocular camera and inputting it into a pre-trained monocular depth model to generate a relative depth map of the RGB image specifically includes: acquiring a single RGB image captured by a monocular camera and inputting the RGB image into the pre-trained monocular depth model; extracting multi-scale visual features of the RGB image through the encoder in the monocular depth model, and fusing and upsampling the multi-scale visual features through the decoder in the monocular depth model, and generating a relative depth map and uncertainty data in the depth estimation process by pixel-by-pixel regression.

3. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 2, characterized in that, The auxiliary feature map includes a semantic segmentation map, an edge response map, a texture energy map, and a depth confidence map. The step of extracting auxiliary image features from the RGB image to generate the auxiliary feature map specifically includes: calling a pre-trained semantic segmentation model to extract semantic features and classify pixels in the RGB image, extracting pixel semantic category features, and generating a semantic segmentation map; converting the RGB image to grayscale, performing edge detection on the grayscale RGB image using a preset edge detection algorithm, extracting gradient intensity features, and generating an edge response map; performing local texture statistics on the RGB image using a preset texture analysis algorithm, extracting texture description features, and generating a texture energy map; obtaining the uncertainty data generated by the decoder of the monocular depth model, extracting the confidence features from the uncertainty data, and generating a depth confidence map.

4. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 3, characterized in that, The step of extracting the feature values ​​corresponding to each candidate anchor point in the auxiliary feature map and calculating the comprehensive quality score of the candidate anchor points specifically includes: obtaining the candidate pixel coordinates corresponding to each candidate anchor point in the RGB image; based on the candidate pixel coordinates, extracting the feature values ​​of the corresponding candidate anchor points from the auxiliary feature map; the feature values ​​include semantic category identifier, edge strength value, texture energy value, and depth confidence value; and performing weighted calculation on the feature values ​​based on the preset weight coefficients corresponding to each image auxiliary feature to obtain the comprehensive quality score corresponding to the candidate anchor point.

5. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 1, characterized in that, The step of calculating the comprehensive optimization score of the initial anchor point subset based on the initial pixel coordinates of each initial anchor point in the RGB image and the relative depth value in the relative depth map, using a predefined combined optimization function, specifically includes: for each initial anchor point subset, obtaining the initial pixel coordinates of all initial anchor points in the RGB image and the relative depth value in the relative depth map; calculating the depth difference of the initial anchor point subset based on the relative depth value to obtain a depth difference evaluation value; calculating the spatial distribution uniformity of the initial anchor point subset based on the initial pixel coordinates to obtain a spatial distribution evaluation value; and performing a weighted summation of the depth difference evaluation value and the spatial distribution evaluation value to obtain the comprehensive optimization score of the initial anchor point subset.

6. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 1, characterized in that, The step of calculating the global scale factor based on the true absolute depth and relative depth values ​​of each optimal anchor point in the optimal anchor point subset specifically includes: obtaining the relative depth value of each optimal anchor point in the optimal anchor point subset corresponding to the relative depth map, and collecting the true absolute depth value of each optimal anchor point in the real-world coordinate system; constructing a linear regression model with the set of relative depth values ​​of all optimal anchor points in the optimal anchor point subset as the independent variable and the set of true absolute depth values ​​of all optimal anchor points as the dependent variable; fitting the linear regression model using the least squares method to obtain the optimal proportionality coefficient between the independent variable and the dependent variable, and using the optimal proportionality coefficient as the global scale factor.

7. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 6, characterized in that, The step of calibrating the relative depth map using the global scale factor to generate an absolute depth map specifically includes: traversing the relative depth map, calibrating the relative depth value of each pixel position in the relative depth map based on the global scale factor to obtain the corresponding absolute depth value; and combining the absolute depth values ​​of all pixel positions according to their corresponding pixel positions to generate an absolute depth map.

8. The monocular 3D reconstruction method based on intelligent anchor point selection according to claim 7, characterized in that, The step of performing inverse perspective projection calculation on the absolute depth map based on the intrinsic parameter matrix of the monocular camera to generate 3D point cloud data of the physical scene in the RGB image specifically includes: traversing the absolute depth map to obtain the absolute depth value of each pixel in the absolute depth map and its pixel coordinates in the RGB image; obtaining the intrinsic parameter matrix of the monocular camera; calculating the 3D spatial coordinates of each pixel in the absolute depth map in the camera coordinate system based on the intrinsic parameter matrix and the absolute depth value and pixel coordinates of each pixel in the absolute depth map through inverse perspective projection relationship; and collecting the 3D spatial coordinates based on the pixel position relationship between the pixels to form 3D point cloud data of the physical scene in the RGB image.

9. A monocular 3D reconstruction device based on intelligent anchor point selection, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a monocular 3D reconstruction method based on intelligent anchor point selection as described in any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to execute a monocular 3D reconstruction method based on intelligent anchor point selection as described in any one of claims 1 to 8.

Images