Binocular three-dimensional scanning system and method with texture mapping function

Abstract

The invention relates to the technical field of three-dimensional scanning, and discloses a binocular three-dimensional scanning system and method with a texture mapping function. The system comprises a depth data acquisition module, a depth data correction module, a point cloud topology construction module, a texture mapping engine module and a mapping optimization module. The depth data acquisition module synchronously acquires a left view and a right view of a target object through a binocular camera array, and generates an original depth image data set; the depth data correction module performs abnormal value detection on the original depth data, extracts discrete distribution through aggregation boundary analysis and replaces abnormal values; the point cloud topology construction module generates a three-dimensional point cloud spatial topology structure, and associates point cloud coordinates with left view pixel positions; the texture mapping engine module projects left view RGB data to the point cloud to generate an initial texture map; and the mapping optimization module verifies the local color consistency through the generative adversarial network, and outputs an optimized three-dimensional texture model. The system improves the texture precision and authenticity of the three-dimensional model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional scanning, in particular to a binocular three-dimensional scanning system and method with texture mapping function. BACKGROUND

[0002] With the in-depth application of three-dimensional modeling technology in various industries, the demand for three-dimensional models with precise shape and real texture is increasing. Binocular three-dimensional scanning system has become a common tool for obtaining three-dimensional information of objects due to its non-contact measurement, controllable cost and other characteristics, and is widely used in reverse engineering, virtual display, medical imaging and other scenes. However, there are still many technical bottlenecks in the texture mapping link of traditional binocular three-dimensional scanning system. In the depth data acquisition stage, the binocular camera of the traditional system often causes matching errors between the left view and the right view due to the deviation of the shooting angle, the fluctuation of the environmental light intensity or the difference of the object surface material. A large number of abnormal data points are mixed in the generated original depth image. If these abnormal values are directly used for subsequent processing, the three-dimensional point cloud will appear local bulge, depression or break in the reconstruction process, which will destroy the original geometric shape of the object and cause hidden troubles for subsequent texture mapping. In the point cloud topology construction link, the traditional method often relies on simple coordinate conversion algorithm to associate the spatial position of the point cloud with the view pixel. This method is easily affected by factors such as lens distortion and object edge blur, causing the matching dislocation of the point cloud coordinates and the corresponding pixels. When performing texture projection, the RGB color data cannot be accurately attached to the surface of the point cloud, and then texture stretching, overlapping or missing occurs, which seriously affects the visual effect of the model. As for the texture mapping optimization, the existing technology generally uses traditional algorithms such as mean filtering and median filtering. These methods can only perform overall smoothing on the texture, and are difficult to deal with local color differences in complex scenes. For example, in the area where there are shadows, highlights or texture mutations on the surface of the object, the traditional optimization algorithm cannot accurately identify the color boundary, which easily leads to hard texture transition, and even covers up the original details of the object. Finally, the generated three-dimensional texture model cannot meet the requirements of high-precision application scenarios. SUMMARY

[0003] The purpose of the present application is to provide a binocular three-dimensional scanning system and method with texture mapping function to solve the problems raised in the background technology.

[0004] To achieve the above purpose, the present application provides a binocular three-dimensional scanning system with texture mapping function, which comprises: a depth data acquisition module for synchronously acquiring the left view and the right view of a target object through a binocular camera array to generate an original depth image data set; a depth data correction module configured to perform outlier detection on the original depth image dataset, extract discrete distribution of depth data by aggregated boundary analysis, and replace abnormal depth values in the discrete distribution based on a confidence interval threshold; a point cloud topology construction module configured to generate a three-dimensional point cloud space topology based on the depth data output by the depth data correction module, and associate each point cloud space coordinate with a pixel position mapping relationship of the left view; a texture mapping engine module configured to project RGB color data of the left view to the three-dimensional point cloud space topology based on the pixel position mapping relationship, and generate an initial texture map; a texture optimization module configured to perform local color consistency verification on the initial texture map by a generative adversarial network, and output an optimized three-dimensional texture model.

[0005] Preferably, the specific steps of the depth data correction module performing abnormal depth value detection include: arranging the original depth image dataset in ascending order of depth values, extracting the 25th percentile position value of the depth sequence as the lower quartile point of the depth distribution, and extracting the 75th percentile position value as the upper quartile point of the depth distribution; calculating the numerical difference between the upper quartile point and the lower quartile point, defining it as the depth aggregation interval, and setting the depth confidence interval boundary based on the multiple relationship of the depth aggregation interval; traversing the original depth image dataset, marking the depth values outside the depth confidence interval boundary as invalid data, and filling the invalid data with the weighted average of adjacent depth values.

[0006] Preferably, the operation logic of the point cloud topology construction module includes: establishing a three-dimensional space grid division rule, dividing the surface of the target object into high-curvature area grid elements and low-curvature area grid elements, and assigning high-density topology nodes to the high-curvature area grid elements; generating a set of spatial coordinates of the topology nodes based on the corrected depth data, and constructing a spatial connection relationship matrix of adjacent topology nodes; mapping the pixel coordinates of the left view to the set of spatial coordinates of the topology nodes to form a pixel position mapping relationship index table.

[0007] Preferably, the texture mapping engine module includes: a color projection unit configured to project RGB color data of the left view to the three-dimensional point cloud space topology according to the pixel position mapping relationship index table based on the topology node coordinates; a texture fusion unit configured to perform bilinear interpolation calculation on the color data of adjacent topology nodes to eliminate color discontinuity phenomenon in the projection process; An illumination compensation unit is configured to analyze an ambient light intensity distribution of the left view and perform normalization compensation processing on the projected color data based on the light intensity distribution.

[0008] Preferably, the texture optimization module comprises: A generator unit is configured to receive three-dimensional color distribution data of an initial texture map and generate adversarial synthetic texture samples. A discriminator unit is configured to compare the synthetic texture samples with real color features of the left view and output a local texture distortion probability matrix. An optimization feedback unit is configured to inversely adjust color interpolation weight parameters of the generator unit according to the local texture distortion probability matrix and iteratively update the three-dimensional texture model.

[0009] Preferably, the system further comprises: A dynamic scanning parameter regulation module is configured to calculate a point cloud stability coefficient based on curvature distribution data of a three-dimensional point cloud spatial topology. A scanning path planning module is configured to index the point cloud stability coefficient to retrieve an optimal binocular camera sampling step in a scanning parameter configuration space. The optimal binocular camera sampling step is fed back to the depth data acquisition module to dynamically adjust exposure time and focal length parameters of the binocular camera array.

[0010] Preferably, the operation logic of the dynamic scanning parameter regulation module comprises: A three-dimensional parameter space coordinate system is constructed, with the horizontal axis representing point cloud curvature distribution variance, the vertical axis representing depth data aggregation interval, and the vertical axis representing ambient light intensity standard deviation. A scanning parameter configuration space is established in the three-dimensional parameter space coordinate system, and each parameter coordinate point is labeled with a corresponding historical camera sampling step. Based on the current point cloud stability coefficient, a target parameter subspace is located in the scanning parameter configuration space, and the optimal camera sampling step in the target parameter subspace is retrieved through a neighborhood density clustering algorithm.

[0011] Preferably, the application further comprises a binocular three-dimensional scanning method with texture mapping function, applied to the binocular three-dimensional scanning system with texture mapping function as described above, and the method comprises the following steps: Synchronously acquire the left view and the right view of the binocular camera, generate an original depth image data set, and perform depth data discrete distribution correction. According to the corrected depth data, a three-dimensional point cloud spatial topology is constructed, and a mapping relationship index between topological node coordinates and left view pixels is established. Based on the mapping relationship index, the left view RGB color data is projected to the three-dimensional point cloud spatial topology to generate an initial three-dimensional texture map. The initial three-dimensional texture map is subjected to local color distortion detection by an adversarial generation network to generate an optimized three-dimensional texture model.

[0012] Preferably, the specific steps of generating the initial three-dimensional texture map comprise: A high-curvature area grid cell in the three-dimensional point cloud spatial topology is identified, and an additional topology node is assigned to the high-curvature area grid cell. The left-view pixel coordinates are subjected to spatial remapping according to the additional topology node coordinates to generate an enhanced pixel position mapping relationship index. RGB color data projection is performed based on the enhanced pixel position mapping relationship index, and Gaussian filtering color smoothing processing is performed between adjacent additional topology nodes.

[0013] Preferably, the operation process of the adversarial generation network comprises: A local color feature vector of the three-dimensional texture model is extracted, and a generator unit is input to generate a synthetic texture data block; The synthetic texture data block and the left-view real color data block are input into a discriminator unit to output a texture distortion probability distribution map; The color interpolation kernel function parameters of the generator unit are corrected in reverse according to the texture distortion probability distribution map, and the mapping output result of the three-dimensional texture model is updated.

[0014] Compared with the prior art, the present application has the following advantages: The left view and the right view are synchronously acquired by the depth data acquisition module using a binocular camera array, and the generated original depth image data set provides comprehensive basic data for subsequent processing. The depth data correction module uses an abnormal value detection and aggregated boundary analysis combination to extract the discrete distribution of the depth data, and replaces the abnormal depth values based on a confidence interval threshold, effectively eliminating the error data caused by environmental interference or device errors, so that the depth data is more consistent with the actual shape of the object. The point cloud topology construction module generates a three-dimensional point cloud spatial topology according to the corrected depth data, and accurately associates each point cloud spatial coordinate with the pixel position mapping relationship of the left view. This accurate association ensures the accuracy of subsequent color projection and avoids texture misplacement caused by position matching deviation. The texture mapping engine module projects the RGB color data of the left view to the three-dimensional point cloud based on the above mapping relationship, and generates an initial texture map that can preliminarily present the color and texture characteristics of the object surface. The map optimization module introduces an adversarial generative network to perform local color consistency verification on the initial texture map. Through the learning and adjustment of the network model, the problem of uneven local color can be effectively improved. Even in complex or slightly occluded parts of the object surface, the color transition is natural, and the overall coordination of the texture map is improved. Through the synergistic effect of a series of modules, the final output three-dimensional texture model is improved in shape accuracy and texture reality, and can better meet the needs of high-precision three-dimensional models in the fields of industrial design, digital twin, digitalization of cultural relics, etc. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A working principle diagram of the binocular three-dimensional scanning system with a texture mapping function according to the present application; Figure 2 A flowchart of the abnormal value detection of the depth data correction module; Figure 3 A flowchart of the working of the texture mapping engine module; Figure 4 A flowchart of the adversarial generation of the map optimization module; Figure 5 A flowchart of the dynamic scanning parameter regulation module. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0017] Please refer to Figure 1 The present application provides a binocular three-dimensional scanning system with a texture mapping function, which comprises a depth data acquisition module, a depth data correction module, a point cloud topology construction module, a texture mapping engine module and a map optimization module.

[0018] The depth data acquisition module synchronously acquires the left view and the right view of the target object through the binocular camera array to generate an original depth image dataset. The depth data correction module performs outlier detection on the original depth image dataset, extracts the discrete distribution of the depth data through aggregated boundary analysis, and replaces the abnormal depth values in the discrete distribution based on a confidence interval threshold. The point cloud topology construction module generates a three-dimensional point cloud spatial topology according to the corrected depth data, and associates the point cloud spatial coordinates with the pixel position mapping relationship of the left view. The texture mapping engine module projects the RGB color data of the left view to the three-dimensional point cloud spatial topology based on the pixel position mapping relationship to generate an initial texture map. The map optimization module performs local color consistency verification on the initial texture map through the generative adversarial network to output an optimized three-dimensional texture model.

[0019] Embodiment 1: refer to Figure 2 The processing procedure of the depth data correction module on the original depth image dataset first needs to sort and arrange the collected depth data. The original depth data is synchronously collected by the binocular camera array and contains the depth information of the target object in the left view and the right view. These data are stored in the form of a matrix, and each pixel point corresponds to a depth value. In order to detect outliers, the depth data needs to be arranged in ascending order according to the numerical value to form an ordered depth sequence. In this sequence, the value at the 25th percentile position is extracted as the lower quartile point of the depth distribution, and the value at the 75th percentile position is extracted as the upper quartile point. These two quartile points are used to describe the central tendency and dispersion degree of the depth data.

[0020] The difference between the upper quartile point and the lower quartile point is defined as the depth aggregation interval, which reflects the overall distribution range of the depth data. Based on the depth aggregation interval, the depth confidence interval boundary is set, and a fixed multiple of the interval is usually used as the threshold value. For example, if the depth aggregation interval is D, the confidence interval boundary can be set as the lower quartile point minus k times D, and the upper quartile point plus k times D, where k is an adjustable coefficient. This coefficient can be adjusted according to the actual application requirements to adapt to the depth data distribution characteristics in different scenarios.

[0021] After setting the confidence interval boundary, the system traverses the original depth image dataset and checks each depth value one by one to see if it is outside the confidence interval range. If a depth value is below the lower boundary or above the upper boundary, it is marked as invalid data. The filling of invalid data uses the weighted average of adjacent valid depth values, and the weight is calculated according to the pixel distance or depth similarity. For example, for a certain invalid depth point, several valid depth points around it can be selected, different weights are given according to the spatial distance, and finally the weighted average value is calculated as the filling value. This process can effectively reduce the noise and outliers in the depth data and improve the accuracy of subsequent three-dimensional reconstruction.

[0022] The point cloud topology construction module generates a three-dimensional point cloud spatial topology structure according to the corrected depth data. The process first establishes a three-dimensional space grid division rule to divide the surface of the target object into a plurality of grid cells. The division of the grid cells is based on the geometric characteristics of the target object, especially the curvature variation. High curvature areas, such as edges, corners and other parts, have large curvature variation, and require more intensive grid division to improve the detail restoration capability. Low curvature areas, such as flat surfaces, have small curvature variation, and can use sparser grid division to reduce computational complexity.

[0023] Each grid cell corresponds to a set of topology nodes, and the density of the nodes is dynamically adjusted according to the type of the grid cell. The high curvature area grid cell is assigned a high density of topology nodes to ensure that the geometric features of the detail part can be accurately expressed. The low curvature area grid cell is assigned a lower density of topology nodes to reduce redundant calculations. The spatial coordinates of the topology nodes are directly generated from the corrected depth data, and each node corresponds to a three-dimensional space point.

[0024] After generating the set of spatial coordinates of the topology nodes, the system further constructs a spatial connection relationship matrix between adjacent nodes. The matrix describes the adjacency relationship between the nodes, which is used for subsequent point cloud surface reconstruction and texture mapping. For example, if the distance between two nodes in three-dimensional space is less than a set threshold, they are marked as adjacent nodes in the matrix. This relationship matrix helps to optimize the structure of the point cloud data, making it more consistent with the surface continuity of the actual object.

[0025] The system maps the pixel coordinates of the left view to the set of spatial coordinates of the topology nodes to form a pixel position mapping relationship index table. The index table records the left view pixel coordinates corresponding to each topology node, so that subsequent texture mapping can accurately project two-dimensional image color information to the three-dimensional point cloud structure. The mapping process is based on camera calibration parameters and geometric transformation relationship to ensure the accurate correspondence between two-dimensional pixels and three-dimensional space points.

[0026] Embodiment 2: refer to Figure 3 The texture mapping engine module is responsible for accurately mapping color information in a two-dimensional image to a three-dimensional point cloud structure, which involves three key steps: color projection, texture fusion and illumination compensation. Through fine algorithm design, the module ensures that the final generated texture map can truly reflect the visual features of the object surface.

[0027] The working of the color projection unit is based on a pixel position mapping relationship index table. The index table is generated by a point cloud topology construction module, which records in detail the coordinate correspondence between each three-dimensional topology node and the corresponding pixel in the left view. During the projection process, the system accurately maps the left view RGB color data according to the spatial coordinates of the topology nodes based on this correspondence. Due to the angle difference between the two cameras during the acquisition process, direct projection may result in missing or misplacement of color information in some areas. To solve this problem, the system uses a multi-view color compensation mechanism. When a topology node is blocked in the left view, the system automatically references the corresponding area of the right view for color compensation.

[0028] The main task of the texture fusion unit is to eliminate the discontinuity that may occur during the color projection process. When the color difference between adjacent topology nodes is large, direct mapping will result in obvious seams and discontinuous areas. The system uses a bilinear interpolation algorithm to smooth these transition areas. This algorithm first identifies combinations of adjacent nodes with large color changes, and then inserts transition color values between these nodes. The calculation of transition values takes into account the color distribution characteristics of multiple surrounding nodes, resulting in a natural gradient effect for the final texture. For high-curvature areas, the system appropriately increases the number of interpolation sampling points to ensure color continuity on complex geometric structures.

[0029] The role of the light compensation unit is to eliminate the impact of uneven environmental lighting on texture quality. Due to local differences in lighting conditions during acquisition, directly projected color data often contains unnecessary changes in brightness. The system analyzes the lighting intensity distribution characteristics of the left view to establish a lighting influence model. This model first decomposes the image into basic color components and lighting components, and then normalizes the lighting components. During the projection process, the system dynamically adjusts the color values of each node based on the spatial orientation information of the three-dimensional point cloud, so that the final texture presents consistent brightness performance under different viewing angles. For high-reflectance or shadow areas, the system will combine multiple frames of acquisition data for comprehensive compensation to avoid the impact of single-frame lighting anomalies on overall texture quality.

[0030] In the specific implementation process of texture mapping, the system uses a hierarchical processing strategy. First, the overall point cloud structure is projected with coarse-grained color to quickly establish a basic texture framework. Then, fine-grained processing is performed on detail areas, especially multiple iterative optimization on high-frequency feature areas such as edges and corners. This hierarchical method ensures both processing efficiency and texture quality in critical areas. The system also maintains a texture quality evaluation mechanism to monitor the mapping effect of each area in real time and automatically triggers the re-projection process for areas that do not meet quality standards.

[0031] For texture mapping of dynamic objects, the system introduces a temporal consistency maintenance mechanism. When processing consecutive frames of three-dimensional scan data, the system records the change patterns of texture features between adjacent frames and makes predictive adjustments during the mapping process of new frames by referring to these patterns. This method effectively reduces inter-frame texture flickering, making the texture of dynamic objects more stable and natural. For fast-moving areas, the system automatically increases the sampling frequency and uses a motion compensation algorithm to correct the texture distortion caused by object displacement.

[0032] In terms of hardware acceleration, the texture mapping engine fully utilizes the parallel computing capabilities of modern GPUs. The data structures of three-dimensional point cloud topology data, color projection matrix, and illumination compensation parameters are organized to facilitate efficient processing by the GPU. In particular, for computationally intensive operations such as bilinear interpolation, the system designs special shader programs for optimization, significantly improving the speed of texture mapping for large-scale point clouds. At the same time, the system supports dynamic load balancing, which can automatically allocate computing resources according to the density distribution of point clouds, ensuring that high-density areas receive more processing power.

[0033] The texture mapping engine also includes an adaptive parameter adjustment subsystem. This system monitors various quality indicators in real time during the texture generation process, including color consistency, edge sharpness, and illumination uniformity. Based on the monitoring results, key parameters such as projection parameters, interpolation coefficients, and compensation strength are dynamically adjusted. This adaptive mechanism enables the system to adapt to texture mapping requirements under different materials and different lighting conditions, maintaining stable output quality. For special materials such as transparent or highly reflective surfaces, the system activates a special material processing mode, using physically-based rendering techniques to enhance the realism of the texture.

[0034] In the final texture output stage, the system performs comprehensive quality checks. The checks include the integrity of color channels, the correctness of texture coordinates, and the matching accuracy with geometric models. Defective areas are put into an automatic repair process, which corrects them through methods such as adjacent data interpolation or multi-view information fusion. After all processing is complete, the system outputs an optimized three-dimensional texture model that retains the visual details of the original image and perfectly fits the three-dimensional geometric structure.

[0035] Embodiment 3: refer to Figure 4 The core function of the texture optimization module is to perform local color consistency verification on the initial texture map through a generative adversarial network (GAN) to eliminate texture distortion or color deviation that may occur during the projection process. This module consists of a generator unit, a discriminator unit, and an optimization feedback unit, gradually improving the texture quality of the three-dimensional texture model through an iterative optimization mechanism.

[0036] The generator unit receives the three-dimensional color distribution data of the initial texture map, and the input is the set of RGB values associated with the topology nodes. The unit uses a convolutional neural network (CNN) with an encoder-decoder structure to extract multi-scale color features. The encoder part is composed of four layers of convolution, with a kernel size of 3x3 and a step size of 2, using LeakyReLU activation function. The encoding process gradually reduces the spatial resolution of the feature map while increasing the number of channels, finally generating a 128-dimensional latent space feature vector. The decoder part gradually recovers the spatial resolution through the transpose convolution layer, and outputs a synthesized texture sample with the same size as the input. The loss function of the generator includes two parts: content loss and adversarial loss. The content loss uses L1 norm to measure the difference between the synthesized sample and the real color data, and the adversarial loss is calculated through the feedback signal of the discriminator.

[0037] The structure of the discriminator unit is a five-layer convolutional network, with an input of 64x64 pixel local texture blocks and an output of a texture distortion probability matrix. The discriminator first downsamples the input block, and after each convolution, it is followed by batch normalization and LeakyReLU activation. The last layer of convolution outputs a single-channel feature map, which is mapped to a probability value between 0 and 1 through the Sigmoid function, representing the possibility of texture distortion in the corresponding region. The training goal of the discriminator is to minimize the misclassification rate of real texture blocks while maximizing the recognition accuracy of synthesized texture blocks.

[0038] The optimization feedback unit dynamically adjusts the parameters of the generator based on the distortion probability matrix output by the discriminator. Each element in the probability matrix represents the distortion probability of the texture block with coordinates , and its calculation method is: ; where represents the output value of the discriminator, and the input texture block. The feedback unit calculates the gradient of the generator parameters through the backpropagation algorithm, and the gradient update amount is determined by the following rules: During the iterative optimization process, the generator and discriminator are updated alternately. After each iteration, the optimization feedback unit resamples the local area of the three-dimensional texture model and calculates a new distortion probability matrix. When the mean value change of the probability matrix of three consecutive iterations is less than the preset tolerance , the model is determined to be converged and the optimization is terminated.

[0039] By using the adaptive texture distortion region positioning mechanism through adversarial training, there is no need to define a mathematical model of color deviation in advance; the multi-scale feature extraction capability of the generator can simultaneously process high-frequency details and low-frequency color transitions; and the dynamic weight adjustment strategy of the optimization feedback unit balances the needs of texture repair and feature preservation. The final output three-dimensional texture model has higher visual reality while maintaining geometric accuracy.

[0040] The core function of the dynamic scanning parameter regulation module is to dynamically adjust the acquisition parameters of the binocular camera according to the spatial characteristics of the three-dimensional point cloud, so as to optimize the quality and scanning efficiency of the depth data. The module analyzes the point cloud stability coefficient to retrieve the optimal camera sampling step in the pre-defined parameter configuration space, and feeds back to the depth data acquisition module to realize adaptive scanning control.

[0041] The calculation of the point cloud stability coefficient is based on the curvature distribution and the depth data aggregation interval. The curvature distribution reflects the geometric complexity of the object surface, and the local curvature value is obtained by calculating the mean of the principal curvature at each topological node, and then the curvature variance of the entire point cloud is calculated. The depth data aggregation interval represents the dispersion degree of the depth data, which is determined by the difference between the upper quartile and the lower quartile. The stability coefficient is defined as the normalized product of the curvature variance and the depth aggregation interval. The larger the value, the more unstable the point cloud data in the current scanning area, and the camera parameters need to be adjusted to improve the data quality. The scanning parameter configuration space stores the historical optimal parameter combination under different scanning conditions, and its structure is shown in the following table: Table 1: Historical optimal parameter combination structure table under different scanning conditions.

[0042]

[0043] During the scanning process, the dynamic scanning parameter regulation module monitors the curvature variance, depth aggregation interval and ambient light intensity standard deviation of the current point cloud in real time, and matches the optimal parameters in the configuration space. The matching process uses a multi-dimensional interval query algorithm. First, the parameter subspace is located according to the curvature variance and the depth aggregation interval, and then the final optimal parameter group is selected according to the light condition. For example, when the curvature variance is 0.35, the depth aggregation interval is 3.2 mm, and the light intensity standard deviation is 80 lux, the module will match the second row of parameters, set the sampling step to 1.5 mm, adjust the exposure time to 15 ms, and increase the focal length by 5%.

[0044] The scanning path planning module generates the motion trajectory of the camera according to the optimal sampling step. For high stability regions (such as flat surfaces), a larger sampling step is used to improve scanning efficiency; for low stability regions (such as high curvature or complex texture parts), a smaller sampling step is used to ensure data accuracy. The trajectory planning algorithm is based on B-spline curve interpolation, ensuring smooth camera movement and covering all key areas.

[0045] Dynamic compensation of ambient light intensity is achieved by analyzing the brightness histogram of the left view. The module calculates the mean and standard deviation of the brightness of the current frame, and if a light mutation (such as a shadow or a reflection) is detected, the exposure time and gain parameters are fine-tuned. The compensation strategy prioritizes the uniformity of exposure in high-curvature areas to avoid depth data jumps caused by changes in light.

[0046] Data-driven adaptive adjustment is achieved through a historical parameter configuration space, avoiding over-sampling or under-sampling problems caused by relying on fixed parameters; a multi-dimensional parameter matching mechanism can simultaneously consider the influence of geometric complexity, data dispersion, and ambient light; dynamic exposure compensation and focal length adjustment work together to ensure data consistency under different scanning conditions.

[0047] Embodiment 5: refer to Figure 5 The dynamic scanning parameter control module achieves fine adjustment of camera sampling parameters by constructing a three-dimensional parameter space coordinate system. This coordinate system has the variance of point cloud curvature distribution as the horizontal axis, the depth data aggregation interval as the vertical axis, and the standard deviation of ambient light intensity as the vertical axis, forming a three-dimensional parameter mapping structure. Each discrete point in the coordinate system is associated with a set of historical scanning parameters and the corresponding point cloud quality score, which is calculated based on geometric consistency and texture clarity.

[0048] The construction process of the parameter configuration space adopts an incremental updating strategy. The initial space is filled with scanning data from standard test scenes, including parameter combinations under different curvature, depth dispersion, and light conditions. As actual scanning tasks are performed, the system continuously records the optimal parameters in new scenes and dynamically expands the configuration space. New data points need to pass the outlier detection to exclude abnormal records caused by device failure or environmental interference. The distribution density of parameter points in the space reflects the common working conditions in actual applications, and high-density areas correspond to frequently occurring scanning conditions.

[0049] The calculation of the point cloud stability coefficient integrates factors such as curvature distribution, depth dispersion, and light fluctuation. The curvature distribution variance is calculated by the moving window method to estimate the local surface variation rate, and the window size is adjusted adaptively according to the object size. The depth aggregation interval uses a dynamic quartile range algorithm to avoid misjudgment of complex geometry with fixed thresholds. The light intensity standard deviation is calculated based on the brightness channel of the left view, and the sampling area is strictly aligned with the depth data acquisition field of view. The stability coefficient is finally normalized to a scalar value between 0 and 1, with a higher value indicating a more complex scanning condition.

[0050] A hierarchical search strategy is used to locate the target parameter subspace. First, the primary search level is determined based on the stability coefficient. A coefficient between 0 and 0.3 corresponds to simple conditions, 0.3 to 0.7 corresponds to moderately complex conditions, and above 0.7 corresponds to highly complex conditions. Within each primary level, secondary subspaces are further divided according to the weighted ratios of curvature, depth, and illumination. The search process utilizes an octree spatial index structure to accelerate queries, and bounding box collision detection is used to quickly eliminate regions with irrelevant parameters.

[0051] The neighborhood density clustering algorithm operates within the target subspace, with clustering characteristics including parameter variation trends and operating condition continuity. The algorithm first calculates the Mahalanobis distance of each parameter point to exclude isolated points with significantly different statistical characteristics. The core clustering process utilizes the adaptive radius DBSCAN method, with the density threshold dynamically adjusted based on the sparseness of the points within the subspace. The center point of each cluster represents a typical parameter configuration for that operating condition, and a weighted voting mechanism ultimately determines the optimal parameters. Voting weights are determined by a combination of historical ratings and recent usage frequency, with parameters that have been successfully applied recently receiving higher weights.

[0052] The parameter feedback mechanism employs a gradual adjustment strategy. When a change in scanning conditions is detected, the system does not immediately switch to a new parameter set. Instead, it adjusts the parameters in stages based on the magnitude of the change. For gradual changes in ambient lighting, the exposure parameters are gradually transitioned using linear interpolation. For sudden changes in geometric complexity, the sampling step size is adjusted using exponential decay. This mechanism avoids discontinuities in acquired data caused by jumps in parameter values ​​while giving the algorithm sufficient time to converge to the optimal state.

[0053] The Environmental Adaptability Enhancement Module continuously monitors parameter performance. After each parameter adjustment, the system collects verification frames and calculates actual point cloud quality metrics. If the metrics fall below a desired threshold, a parameter rollback mechanism is triggered, automatically reverting to the last stable parameter set and marking the current operating condition as requiring special attention.

[0054] 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 any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0055] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A binocular 3D scanning system with texture mapping function, characterized in that: include: The depth data acquisition module is used to synchronously acquire the left and right views of the target object through a binocular camera array to generate an original depth image dataset; The depth data correction module is used to perform outlier detection on the original depth image dataset, extract the discrete distribution of depth data through aggregate boundary analysis, and replace abnormal depth values ​​in the discrete distribution based on the confidence interval threshold; The point cloud topology construction module is used to generate a three-dimensional point cloud spatial topology structure based on the depth data output by the depth data correction module, and associate the mapping relationship between the spatial coordinates of each point cloud and the pixel position of the left view; The texture mapping engine module is used to project the RGB color data of the left view onto the three-dimensional point cloud space topology structure based on the pixel position mapping relationship to generate an initial texture map; The texture optimization module is used to perform local color consistency verification on the initial texture map through the adversarial generative network and output the optimized three-dimensional texture model.

2. The binocular 3D scanning system with texture mapping function according to claim 1, characterized in that: The specific steps of the depth data correction module performing abnormal depth value detection include: Arrange the original depth image dataset in ascending order of depth value, extract the 25th percentile value of the depth sequence as the lower quartile of the depth distribution, and extract the 75th percentile value as the upper quartile of the depth distribution; The numerical difference between the upper quartile and the lower quartile is calculated and defined as the depth aggregation interval. The depth confidence interval boundary is set based on the multiple relationship of the depth aggregation interval. Traverse the original depth image dataset, mark the depth values ​​that exceed the depth confidence interval boundary as invalid data, and use the weighted average of adjacent depth values ​​to fill the invalid data.

3. The binocular 3D scanning system with texture mapping function according to claim 2, characterized in that: The operation logic of the point cloud topology construction module includes: Establish a three-dimensional space grid division rule to divide the surface of the target object into grid cells in the high curvature area and grid cells in the low curvature area, and assign high-density topological nodes to the grid cells in the high curvature area; Generate a set of spatial coordinates of topological nodes based on the corrected depth data, and construct a spatial connection relationship matrix of adjacent topological nodes; The pixel coordinates of the left view are mapped to the spatial coordinate set of the topological nodes to form a pixel position mapping relationship index table.

4. The binocular 3D scanning system with texture mapping function according to claim 3, characterized in that: The texture mapping engine module includes: A color projection unit is used to project the RGB color data of the left view onto the three-dimensional point cloud space topology structure according to the topological node coordinates based on the pixel position mapping relationship index table; The texture fusion unit is used to perform bilinear interpolation calculation on the color data of adjacent topological nodes to eliminate the color fault phenomenon in the projection process; The illumination compensation unit is configured to analyze the ambient illumination intensity distribution of the left view and perform normalization compensation processing on the projection color data based on the illumination intensity distribution.

5. The binocular 3D scanning system with texture mapping function according to claim 4, characterized in that: The mapping optimization module includes: A generator unit, configured to receive three-dimensional color distribution data of an initial texture map and generate adversarial synthetic texture samples; The discriminator unit is used to compare the synthetic texture samples with the real color features of the left view and output the local texture distortion probability matrix; The optimization feedback unit is used to reversely adjust the color interpolation weight parameters of the generator unit according to the local texture distortion probability matrix and iteratively update the three-dimensional texture model.

6. The binocular 3D scanning system with texture mapping function according to claim 5, characterized in that: The system further comprises: Dynamic scanning parameter control module, used to calculate the point cloud stability coefficient based on the curvature distribution data of the three-dimensional point cloud spatial topological structure; The scanning path planning module is used to retrieve the optimal binocular camera sampling step size in the scanning parameter configuration space using the point cloud stability coefficient as an index; The optimal binocular camera sampling step size is fed back to the depth data acquisition module to dynamically adjust the exposure time and focal length parameters of the binocular camera array.

7. The binocular 3D scanning system with texture mapping function according to claim 6, characterized in that: The operating logic of the dynamic scanning parameter control module includes: Construct a three-dimensional parameter space coordinate system, where the horizontal axis represents the variance of the point cloud curvature distribution, the vertical axis represents the depth data aggregation spacing, and the vertical axis represents the standard deviation of the ambient light intensity; Establish a scanning parameter configuration space in the three-dimensional parameter space coordinate system, and mark the corresponding historical camera sampling step for each parameter coordinate point; The target parameter subspace is located in the scanning parameter configuration space based on the current point cloud stability coefficient, and the optimal camera sampling step in the target parameter subspace is retrieved through the neighborhood density clustering algorithm.

8. A binocular 3D scanning method with texture mapping function, applied to a binocular 3D scanning system with texture mapping function as claimed in any one of claims 1 to 7, characterized in that: The following steps are involved: Synchronously acquire the left and right views of the binocular camera, generate the original depth image dataset, and perform discrete distribution correction of the depth data; Construct a 3D point cloud spatial topology structure based on the corrected depth data, and establish a mapping relationship index between the topological node coordinates and the left view pixels; Project the RGB color data of the left view onto the 3D point cloud spatial topology based on the mapping relationship index to generate an initial 3D texture map; The initial 3D texture map is subjected to local color distortion detection through a generative adversarial network to generate an optimized 3D texture model.

9. The binocular 3D scanning method with texture mapping function according to claim 8, characterized in that: The specific steps of generating the initial three-dimensional texture map include: Identify grid cells in high curvature areas in the three-dimensional point cloud spatial topology structure and assign additional topological nodes to the grid cells in the high curvature areas; The left view pixel coordinates are spatially remapped according to the additional topological node coordinates to generate an enhanced pixel position mapping relationship index; RGB color data projection is performed based on the enhanced pixel position mapping relationship index, and Gaussian filtering color smoothing is performed between adjacent additional topological nodes.

10. The binocular 3D scanning method with texture mapping function according to claim 9, characterized in that: The operation process of the adversarial generative network includes: Extracting the local color feature vector of the three-dimensional texture model and inputting it into the generator unit to generate a synthetic texture data block; The synthesized texture data block and the left view real color data block are input into the discriminator unit, and the texture distortion probability distribution map is output; The color interpolation kernel function parameters of the generator unit are reversely corrected according to the texture distortion probability distribution map, and the mapping output result of the three-dimensional texture model is updated.

Images