Enterprise scene-oriented data visualization method and device, terminal equipment and storage medium

Abstract

The invention discloses an enterprise scene-oriented data visualization method and device, terminal equipment and a storage medium, and the method comprises the steps: carrying out the feature extraction of each face image through a pre-trained ArcFace model, and obtaining a plurality of multi-dimensional feature vectors; matching the multi-dimensional feature vector with an employee feature vector in an employee feature library, and determining a personnel type corresponding to each human body point cloud according to a matching result; determining a data generation library according to the personnel type corresponding to each human body point cloud; according to the point cloud distance between each human body point cloud and the display screen, determining a chart scaling; determining a plurality of metadata in the data generation library according to question and answer prompt words input into the display screen; and converting the plurality of metadata into a visual chart displayed on the display screen according to the chart scaling. According to the method, the identity, the position and the intention of the user are automatically perceived, and personalized and self-adaptive visual display of enterprise data is realized.

Description

Technical Field

[0001] This invention relates to the field of enterprise data processing, and in particular to a data visualization method, apparatus, terminal device, and storage medium for enterprise scenarios. Background Technology

[0002] Currently, enterprise data visualization primarily relies on traditional business intelligence (BI) tools such as Tableau, Power BI, and Qlik. While these tools offer a rich variety of chart types and interactive features, they have significant limitations in practical applications: First, the existing system uses a fixed visualization configuration, which cannot dynamically adjust the content based on the viewer's identity, location, and permissions. In a conference room setting, when people with different permission levels view the large screen simultaneously, the system cannot automatically balance data security and information display needs. This usually requires manually switching views or preparing multiple display options in advance, severely impacting decision-making efficiency.

[0003] Second, existing visualization systems lack the ability to understand physical space. The readability of the same chart varies greatly at different viewing distances: users at close range want to see detailed data, while users at a distance need aggregated views and large fonts. Traditional solutions rely on preset "mobile views" and "desktop views," which cannot achieve real-time, continuous spatial adaptation. Summary of the Invention

[0004] This invention provides a data visualization method, apparatus, terminal device, and storage medium for enterprise scenarios. By integrating 3D spatial perception, biometric recognition, and natural language understanding, it automatically perceives user identity, location, and intent, thereby achieving personalized and adaptive visualization of enterprise data.

[0005] To achieve the above objectives, a first aspect of this application provides a data visualization method for enterprise scenarios, comprising: The system acquires images captured by several cameras, and performs point cloud modeling on each human body based on the captured images to obtain several human body point clouds. When performing point cloud modeling, the origin of the point cloud coordinate system is set at the geometric center of the display screen, and the X-axis direction of the point cloud coordinate system is consistent with the orientation of the display screen. The overlapping range of adjacent cameras is within a preset range, and the camera whose lens orientation is consistent with the orientation of the display screen is the main camera. The MTCNN model is used to perform face detection and alignment on the images captured by the main camera to obtain several face images; each face image corresponds to a human point cloud. The pre-trained ArcFace model is used to extract features from each face image to obtain several multi-dimensional feature vectors; The multidimensional feature vector is matched with the employee feature vector in the employee feature library, and the personnel type corresponding to each human point cloud is determined according to the matching result. A data generation library is determined based on the personnel type corresponding to each human point cloud; the chart scaling ratio is determined based on the point cloud distance between each human point cloud and the display screen. Based on the question-and-answer prompts input to the display screen, several metadata items are determined in the data generation library; the metadata items are then converted into a visual chart displayed on the display screen according to the chart scaling ratio.

[0006] In one possible implementation of the first aspect, the step of performing point cloud modeling on each human body based on the captured images to obtain several human body point clouds specifically includes: The ICP algorithm is used to align images captured by different cameras to obtain aligned point clouds; PointNet was used to separate individual human point clouds from the aligned point cloud.

[0007] In one possible implementation of the first aspect, before extracting features from each face image using a pre-trained ArcFace model to obtain several multi-dimensional feature vectors, the specific steps include: Collect and clean a set of facial images of the company's employees, where each employee has more than three facial images, and more than one of the facial images was taken in front of the display screen; The base ArcFace model was trained using ArcFace loss on a publicly available face dataset; By combining the set of facial images of the company's employees, the basic ArcFace model is adjusted using multi-task training.

[0008] In one possible implementation of the first aspect, the step of combining the set of facial images of the company's employees and adjusting the base ArcFace model using multi-task training specifically includes: Load the set of facial images of the company's employees, and occlude and rotate the faces in the images within the set by a preset ratio to obtain a fine-tuned training set; The weighted sum of the arc surface loss function, the triple loss function, and the center loss function is used as the optimization objective. The base ArcFace model is trained using the fine-tuned training set to obtain a pre-trained ArcFace model. The weight factor of the arc surface loss function is 0.9, the weight factor of the triple loss function is 0.3, and the weight factor of the center loss function is 0.01.

[0009] In one possible implementation of the first aspect, the step of matching the multidimensional feature vector with employee feature vectors in the employee feature database, and determining the personnel type corresponding to each human point cloud based on the matching result, specifically includes: In three consecutive frames, if the cosine similarity between the multidimensional feature vector and an employee feature vector in the employee feature library is greater than or equal to a preset threshold, the corresponding personnel type of the human point cloud associated with the multidimensional feature vector is an enterprise employee; if the cosine similarity between the multidimensional feature vector and an employee feature vector in the employee feature library is less than the preset threshold, the corresponding personnel type of the human point cloud associated with the multidimensional feature vector is a visitor; enterprise employees include several levels, and different levels of enterprise employees can access different enterprise databases.

[0010] In one possible implementation of the first aspect, determining the data generation library based on the personnel type corresponding to each of the human point clouds, and determining the chart scaling ratio based on the point cloud distance between each of the human point clouds and the display screen, specifically includes: If the person type corresponding to the human point cloud is a visitor, the lowest-level enterprise data sub-database is used as the data generation database; if the person type corresponding to all the human point clouds is an enterprise employee, the lowest-level enterprise employee is determined among all the person types corresponding to the human point clouds, and the enterprise data sub-database corresponding to that level of enterprise employee is selected as the data generation database; the enterprise data sub-database corresponding to the lowest-level enterprise employee is one level higher than the lowest-level enterprise data sub-database; the lower-level enterprise data sub-database is obtained by distilling the higher-level enterprise data sub-database. Calculate the average point cloud distance between each of the human body point clouds and the display screen; Select the chart scaling ratio based on the average point cloud distance.

[0011] In one possible implementation of the first aspect, determining several metadata items in the data generation library based on the question-and-answer prompts input to the display screen; and converting the several metadata items into a visual chart displayed on the display screen according to the chart scaling ratio, specifically includes: The question-and-answer prompts input to the display screen are semantically decomposed to obtain semantic vectors; In the data generation library, for each metadata, if the cosine similarity between the word sense vector and the data vector of the metadata in the data generation library is greater than a preset threshold, the metadata is used as a visualization target. For each visualization target, determine the vector subspace to which the metadata belongs; the vector space includes a comparison vector subspace, a trend vector subspace, a proportion vector subspace, a distribution vector subspace, and a relationship vector subspace; based on the data labels corresponding to all the data vectors in the vector space, perform K-means clustering on all the data vectors to obtain several cluster centers, and one cluster center corresponds to a center vector of a vector space; Metadata belonging to the comparison vector subspace is converted into a bar chart displayed on the screen according to the chart scaling ratio; metadata belonging to the trend vector subspace is converted into a line chart displayed on the screen according to the chart scaling ratio; metadata belonging to the proportion vector subspace is converted into a pie chart displayed on the screen according to the chart scaling ratio; metadata belonging to the distribution vector subspace is converted into a histogram displayed on the screen according to the chart scaling ratio; and metadata belonging to the relationship vector subspace is converted into a scatter plot displayed on the screen according to the chart scaling ratio.

[0012] A second aspect of this application provides a data visualization device for enterprise scenarios, including: The point cloud module is used to acquire images captured by several cameras, and to perform point cloud modeling on each human body based on the captured images to obtain several human body point clouds. When performing point cloud modeling, the origin of the point cloud coordinate system is set at the geometric center of the display screen, and the X-axis direction of the point cloud coordinate system is consistent with the orientation of the display screen. The overlapping range of adjacent cameras is within a preset range, and the camera whose lens orientation is consistent with the orientation of the display screen is the main camera. The detection module is used to perform face detection and alignment on the images captured by the main camera using the MTCNN model to obtain several face images; each face image corresponds to a human point cloud. The extraction module is used to extract features from each face image using a pre-trained ArcFace model, resulting in several multi-dimensional feature vectors. The matching module is used to match the multidimensional feature vector with the employee feature vector in the employee feature library, and determine the personnel type corresponding to each human point cloud based on the matching result; The determination module is used to determine the data generation library based on the personnel type corresponding to each of the human point clouds; and to determine the chart scaling ratio based on the point cloud distance between each of the human point clouds and the display screen. The display module is used to determine several metadata items in the data generation library based on the question-and-answer prompts input to the display screen; and to convert the several metadata items into a visual chart displayed on the display screen according to the chart scaling ratio.

[0013] A third aspect of this application provides a terminal device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a data visualization method for enterprise scenarios as described above.

[0014] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a data visualization method for enterprise scenarios as described above.

[0015] Compared to existing technologies, this invention achieves intelligent data visualization within the constraints of a single display screen through a three-tiered linkage of spatial perception, identity recognition, and adaptive rendering. Its core value lies in a unified content strategy based on minimum access permissions and adaptive rendering based on average distance: when executives and clients are in the same room, the system automatically displays data content that meets the minimum access requirements, while optimizing the display effect according to the viewing distance. This ensures enterprise data security and compliance, improves long-distance readability, enhances meeting decision-making efficiency, and truly realizes "one screen, unified security, and adaptive experience" for intelligent enterprise data Q&A interaction. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a data visualization method for enterprise scenarios according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a data visualization device for enterprise scenarios provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To resolve the above issues, please refer to [link / reference]. Figure 1 An embodiment of the present invention provides a data visualization method for enterprise scenarios, comprising: S10. Acquire images captured by several cameras, and perform point cloud modeling on each human body based on the captured images to obtain several human body point clouds; when performing point cloud modeling, the origin of the point cloud coordinate system is set at the geometric center of the display screen, and the X-axis direction of the point cloud coordinate system is consistent with the orientation of the display screen; the overlapping range of adjacent cameras is within a preset range, and the camera whose lens orientation is consistent with the orientation of the display screen is the main camera.

[0019] S11. Use the MTCNN model to perform face detection and alignment on the images captured by the main camera to obtain several face images; each face image corresponds to a human point cloud.

[0020] S12. Use the pre-trained ArcFace model to extract features from each face image to obtain several multi-dimensional feature vectors.

[0021] S13. Match the multidimensional feature vector with the employee feature vector in the employee feature library, and determine the personnel type corresponding to each human point cloud based on the matching result.

[0022] S14. Determine the data generation library based on the personnel type corresponding to each of the human point clouds; determine the chart scaling ratio based on the point cloud distance between each of the human point clouds and the display screen.

[0023] S15. Based on the question-and-answer prompts input to the display screen, determine several metadata in the data generation library; convert the several metadata into a visual chart displayed on the display screen according to the chart scaling ratio.

[0024] S10-S11 describes the spatial perception and face processing process: Several cameras are deployed around the display screen. Generally, the overlap of the shooting areas of adjacent cameras is strictly controlled between 30% and 60% to ensure no blind spots in spatial coverage. Then, a unified point cloud coordinate system is established: the origin O is precisely set at the geometric center of the display screen, and the positive X-axis direction is strictly consistent with the orientation of the display screen. The main camera is selected as the device with the smallest angle between the lens optical axis and the X-axis (usually <15°). A high-precision human point cloud is generated by fusing the depth images of each camera through a multi-view stereo matching algorithm. The MTCNN model performs face detection on the RGB image of the main camera, outputting the face bounding box and the coordinates of 5 key points (center of the eyes, tip of the nose, and corners of the mouth). After affine transformation, a 112×112 standardized aligned image is generated, and the 2D face position is projected into 3D space through the camera intrinsic parameter matrix K and extrinsic parameter matrix [R|t], establishing a correspondence with the center of the human point cloud at a distance of <1.5 meters.

[0025] This improves the accuracy of face detection in complex office environments. The collaborative work of multiple cameras eliminates blind spots and provides a precise spatial basis for subsequent identity recognition and adaptive visualization. When a user approaches from a distance, the system can track and update their spatial location information in real time.

[0026] S12-S13 is the identity recognition process, which uses an ArcFace deep neural network pre-trained on public datasets such as MS1M to encode the aligned face image into a 512-dimensional feature vector f∈ℝ. 5¹². Perform L2 normalization on the feature vectors: f' = f / ||f||², ensuring the feature vectors lie on the unit hypersphere. Calculate the cosine similarity between f' and each vector gᵢ in the employee feature database: sim(f',gᵢ) = f'·gᵢ. Set a threshold τ1 = 0.65; only when the similarity of three consecutive frames is ≥ τ1 is the corresponding human point cloud identified as an employee; otherwise, it is identified as a visitor.

[0027] This identity recognition mechanism boasts extremely high accuracy and robustness, maintaining stable and reliable recognition performance under various office scenarios. The continuous multi-frame verification strategy effectively reduces the risk of misjudgment, and the feature matching speed of the large-scale employee database meets the needs of real-time interaction. It also exhibits excellent cross-time recognition stability, providing accurate identity criteria for data access control. Even under conditions of drastic changes in lighting or when users wear common accessories, it can accurately distinguish the identity characteristics of users with different access levels.

[0028] S14-S15 describes the adaptive visualization generation process for enterprise data: Assume there are n observers in the scene, with permission levels L1, L2, ..., L... n (Employees are set to different levels, and visitors are defined as the lowest level.) The system selects the lowest permission level, L. min =min(L1,...,L n The corresponding sub-database serves as the data generation source, ensuring data security. The average distance d̄ from the center of all human point clouds to the origin O of the display screen is calculated as: d̄ = (d1 + d2 + ... + d...). n ) / n, where dᵢ represents the Euclidean distance from the center of the i-th human point cloud to the origin O. The chart scaling ratio s is dynamically determined based on d̄: a larger scaling ratio is used when d̄ is small, and a smaller scaling ratio is used when d̄ is large. Semantic parsing is performed on the input question-and-answer prompts to extract the query intent, relevant metadata is retrieved from the selected data generation library, the optimal chart type is automatically selected based on data characteristics, and all visualization elements are adjusted according to the ratio s.

[0029] This visualization mechanism balances data security and user experience on a single screen. When people with different access levels are in the same space, it automatically switches to the data view that meets the minimum access requirements, effectively preventing the leakage of sensitive information. Adaptive rendering based on viewing distance significantly improves the readability of content at different distances. When viewed from a distance, it optimizes fonts and layout to ensure clear information delivery, while providing rich details when viewed up close. The semantic parsing engine accurately understands the user's query intent and quickly converts natural language into professional charts, greatly reducing the barrier to entry and significantly improving the efficiency of meeting decisions. This allows users from different backgrounds to obtain the best information acquisition experience in the same display environment.

[0030] For example, the step of performing point cloud modeling on each human body based on the captured images to obtain several human body point clouds specifically includes: The ICP algorithm is used to align images captured by different cameras to obtain aligned point clouds.

[0031] PointNet was used to separate individual human point clouds from the aligned point cloud.

[0032] The ICP (Iterative Closest Point) algorithm achieves accurate registration of point clouds from multiple perspectives through iterative optimization. Let the system contain k cameras, and let the original point clouds captured by each camera be denoted as P1, P2, ..., P... k Each point cloud consists of N 3D points: Pᵢ = {pᵢ1, pᵢ2, ..., pᵢ} n}, where pᵢⱼ = (x, y, z) represents the three-dimensional coordinates of a point. The algorithm first establishes a global coordinate system with the geometric center of the display screen as the origin O and the X-axis aligned with the screen orientation. For each pair of point clouds from adjacent cameras (such as Pᵢ and Pⱼ), ICP calculates the optimal rigid transformation matrix Tᵢⱼ = [Rᵢⱼ|tᵢⱼ], where Rᵢⱼ is a 3×3 rotation matrix and tᵢⱼ is a 3×1 translation vector, minimizing the mean square error between the transformed point cloud and the target point cloud: Tᵢⱼ* = argminΣ||Rᵢⱼ·pᵢ + tᵢⱼ - pⱼ'||², where pⱼ' is the nearest neighbor of pᵢ in Pⱼ. To enhance robustness, the algorithm introduces the RANSAC strategy to remove outlier correspondences and ensures seamless fusion of all point clouds in a unified coordinate system through global optimization. This alignment process achieves precise spatial unification of multi-view point cloud data, effectively eliminating viewpoint differences and measurement errors between different cameras, ensuring the consistency and continuity of the spatial coordinate system, and providing a geometrically accurate 3D scene foundation for human detection. Even under conditions of changing lighting or partial occlusion, it maintains the stability of spatial mapping, enabling the system to accurately understand the user's true position and posture in physical space.

[0033] The PointNet network architecture directly processes aligned point cloud data without voxelization or projection. The input is the fused global point cloud P = {p1, p2, ..., p...} mEach point pᵢ contains coordinates (x, y, z) and features such as a normal vector. The network first learns the optimal spatial alignment transformation through an input transformation network T-Net to improve robustness to rigid transformations. Then, a shared multilayer perceptron (MLP) independently processes each point to extract local features. Global shape features are aggregated through max pooling. Finally, combining local and global features, a fully connected layer predicts an instance label for each point. For human segmentation, the network outputs the class probability distribution for each point and uses a clustering algorithm (such as DBSCAN) to aggregate high-probability human points into independent instances. To handle occlusion scenarios, the network introduces an attention mechanism to enhance feature extraction capabilities for key human body parts (such as the head and torso), ensuring accurate segmentation boundaries. This segmentation process achieves high-precision separation of human instances in complex scenes, effectively handling mutual occlusion, pose changes, and background interference between people, generating a complete and continuous human point cloud model, providing a reliable geometric basis for subsequent face-body association, enabling the system to accurately identify the independent spatial position and contour of each user, and maintaining the continuity and accuracy of individual recognition even in densely populated or complex situations, providing precise spatial perception capabilities for human-computer interaction.

[0034] For example, before extracting features from each face image using a pre-trained ArcFace model to obtain several multi-dimensional feature vectors, the process specifically includes: Collect and clean a set of facial images of the company's employees, where each employee has more than three facial images, and more than one of the facial images was taken in front of the display screen; The base ArcFace model was trained using ArcFace loss on a publicly available face dataset; By combining the set of facial images of the company's employees, the basic ArcFace model is adjusted using multi-task training.

[0035] During the enterprise's facial image collection phase, each employee is required to provide at least three photos under different conditions, one of which must be taken in a real-world usage environment (in front of a display screen) to capture lighting and pose features in a real interactive scenario. Image cleaning employs a multi-level filtering mechanism: first, MTCNN is used for face detection and quality scoring to remove samples with excessively small faces, excessive blur, or missing key points; second, an illumination normalization algorithm is applied to eliminate the effects of extreme brightness; finally, manual review ensures the accuracy of identity labeling. The base model is trained on a publicly available large-scale face dataset (such as MS1M) using the ArcFace loss function: for a normalized feature vector x and classification weights W, with the true class y, the loss is defined as L = -log(exp(s·cos(θ_y + m))) / Σ_j exp(s·cos(θ_j + δ_{j,y}·m)), where θ_y = arccos(W_y^T·x) is the angle between the feature and the weight, m ​​is the angle margin (usually 0.5 radians), s is the feature scaling factor (usually 64), and δ is the indicator function. This loss imposes a margin in the angular space, making intra-class features compact and inter-class features separate.

[0036] The data preparation and basic training process establishes a foundation for face recognition with broad generalization capabilities. Real-world data collection ensures the model adapts to specific enterprise environments, while a rigorous data cleaning mechanism guarantees the high quality and consistency of training data. Training on large-scale public datasets endows the model with robust understanding of various facial variations. The introduction of angular margins significantly enhances the discriminative power of the feature space, laying a solid foundation for subsequent customized fine-tuning by enterprises. This enables the model to effectively cope with practical challenges in the office environment, such as changes in lighting, diverse poses, and partial occlusion.

[0037] For example, the step of combining the set of facial images of the company's employees and adjusting the base ArcFace model using multi-task training specifically includes: Load the set of facial images of the company's employees, and occlude and rotate the faces in the images within the set by a preset ratio to obtain a fine-tuned training set; The weighted sum of the arc surface loss function, the triple loss function, and the center loss function is used as the optimization objective. The base ArcFace model is trained using the fine-tuned training set to obtain a pre-trained ArcFace model. The weight factor of the arc surface loss function is 0.9, the weight factor of the triple loss function is 0.3, and the weight factor of the center loss function is 0.01.

[0038] Based on the basic ArcFace model, targeted fine-tuning was performed using a dataset of enterprise employees. First, scene-based enhancement was implemented on the images of enterprise employees: a virtual mask was added with a 30% probability to cover the lower half of the face, and random rotations of ±20° were performed with a 25% probability to simulate common changes in real office scenarios. Fine-tuning employs a combination of multi-task loss functions: L_total = λ1·L_arcface + λ2·L_triplet + λ3·L_center, where L_arcface is the standard ArcFace loss (λ1=0.9), applying margins in the angle space; L_triplet is the triplet loss (λ2=0.3), defined as max(d(x_a,x_p)- d(x_a,x_n) + α, 0), where d is the cosine distance, α is the distance margin (usually 0.3), and x_a, x_p, and x_n represent anchor points, positive samples, and negative samples, respectively; L_center is the class center loss (λ3=0.01), constraining similar features to cluster inwards. Training employs a two-stage strategy: initially using a high learning rate to quickly adapt to the enterprise data distribution, and later reducing the learning rate to fine-tune the decision boundary, while dynamically increasing the margin parameter m for easily confused employee pairs (such as similar appearances) to strengthen the ability to distinguish identities.

[0039] For example, the step of matching the multidimensional feature vector with the employee feature vector in the employee feature database, and determining the personnel type corresponding to each human point cloud based on the matching result, specifically includes: In three consecutive frames, if the cosine similarity between the multidimensional feature vector and an employee feature vector in the employee feature library is greater than or equal to a preset threshold, the corresponding personnel type of the human point cloud associated with the multidimensional feature vector is an enterprise employee; if the cosine similarity between the multidimensional feature vector and an employee feature vector in the employee feature library is less than the preset threshold, the corresponding personnel type of the human point cloud associated with the multidimensional feature vector is a visitor; enterprise employees include several levels, and different levels of enterprise employees can access different enterprise databases.

[0040] The system dynamically matches the real-time extracted facial feature vectors with a pre-registered employee feature database. Let the normalized feature vector extracted in the current frame be f∈ℝ. 5 ¹², The employee feature library contains n normalized feature vectors {g1, g2, ..., g n}. Calculate cosine similarity: sim(f, gᵢ) = f·gᵢ (since the vectors have been normalized, the dot product is the cosine value). Set a preset threshold τ2 (usually 0.6-0.7) as the decision boundary for identity matching. The system adopts a continuous frame verification mechanism: only when the similarity between f and the same employee feature gᵢ is ≥τ2 in three consecutive frames (k=3) is the human point cloud determined to correspond to an employee of the enterprise; otherwise, it is determined to be a visitor. Employee level classification is based on similarity confidence: for confirmed enterprise employees, they are classified into different permission levels according to the average similarity of three consecutive frames. The higher the similarity, the higher the matching confidence, and the higher the corresponding permission level.

[0041] This determination mechanism significantly improves the reliability and stability of identity recognition. Continuous frame verification effectively filters out instantaneous interference and accidental mismatches, greatly reducing the risk of incorrect identity determination. Cosine similarity measurement provides a scale-invariant identity measurement standard in the normalized feature space, adapting to different lighting and imaging conditions. The dynamic threshold mechanism balances recognition sensitivity and security. The multi-level permission design realizes fine-grained data access control, ensuring that sensitive information is only open to authorized personnel, while maintaining the system's response speed and user experience smoothness, laying an accurate identity foundation for subsequent differentiated data display.

[0042] In this embodiment, the personnel type determination result is directly mapped to the data access permission system. The system predefines a multi-level enterprise database, with levels L (L=1,2,...,m) corresponding to different data granularities and sensitivities; the larger the L value, the higher the permission. When determined to be an enterprise employee, the corresponding data subdatabase with the appropriate permission is dynamically selected based on their level L; when determined to be a visitor, only access to the anonymized dataset with the lowest permission (L=0) is allowed. The permission decision follows the "least common permission" principle: in multi-person scenarios, the system automatically selects the lowest permission level among all personnel present as the permission benchmark for the current data display, ensuring that no user can access data content beyond their permissions.

[0043] This permission mechanism achieves strict data security boundaries on a single display screen. Even if high-privilege and low-privilege users view the data simultaneously, the system can automatically ensure compliant data display and completely eliminate the risk of unauthorized data leakage. Dynamic permission mapping enables the system to adjust data sensitivity in real time based on the composition of the users present, without manual intervention. The multi-level data sub-database design balances information value and security requirements, allowing users with different roles to obtain information content matching their permissions in the same environment. This not only ensures the security of enterprise data assets but also maintains collaborative efficiency, truly achieving an organic unity of security and user experience.

[0044] For example, the step of determining a data generation library based on the personnel type corresponding to each of the human point clouds, and determining the chart scaling ratio based on the point cloud distance between each of the human point clouds and the display screen, specifically includes: If the person type corresponding to the human point cloud is a visitor, the lowest-level enterprise data sub-database is used as the data generation database; if the person type corresponding to all the human point clouds is an enterprise employee, the lowest-level enterprise employee is determined among all the person types corresponding to the human point clouds, and the enterprise data sub-database corresponding to that level of enterprise employee is selected as the data generation database; the enterprise data sub-database corresponding to the lowest-level enterprise employee is one level higher than the lowest-level enterprise data sub-database; the lower-level enterprise data sub-database is obtained by distilling the higher-level enterprise data sub-database. Calculate the average point cloud distance between each of the human body point clouds and the display screen; Select the chart scaling ratio based on the average point cloud distance.

[0045] The system determines data access permissions in real time based on the identities of the people present. Assume there are n human point clouds in the scene, with the corresponding set of person types as T = {t1, t2, ..., t...}. n}, where tᵢ represents the type of the i-th person (visitor or employee of different levels). Data access control decisions follow the "lowest common access" principle: when there are visitor types in set T, the system selects the data sub-database D0 corresponding to the lowest access level L0 as the data generation source; when all members of T are employees, the system selects the lowest employee level L m ᵢ n = min({lᵢ | tᵢ represents company employees}) corresponds to the subdatabase D min The data sub-database adopts a hierarchical distillation architecture: the higher-level sub-database D... h Includes complete raw data, low-level sub-database D l Knowledge is generated from higher-level sub-repositories through knowledge distillation. The distillation process includes data anonymization (such as numerical range division and detailed aggregation), sensitive field filtering, and retention of statistical characteristics. This ensures that lower-level repositories maintain data distribution consistency while reducing information granularity. For example, D1 (the lowest-level employee repository) contains more granular business metrics than D0 (the visitor repository).

[0046] This permission decision-making mechanism achieves seamless data security boundaries in multi-person collaboration scenarios. It automatically and dynamically adjusts data sensitivity based on the composition of the participants, ensuring that no user can access content beyond the minimum permissions. This completely eliminates the risk of data leakage caused by permission switching delays in traditional systems. The hierarchical distillation architecture maintains data value while ensuring security, allowing users with different permissions to access information density matching their roles on the same screen. This protects the company's core data assets and avoids the interactive interruptions caused by frequent manual view switching, significantly improving team collaboration efficiency and decision-making quality.

[0047] The system optimizes the visualization effect through spatial calculation. Let the center coordinates of n human point clouds in the scene be Pᵢ = (xᵢ, yᵢ, zᵢ), and the origin of the display screen coordinate system be O = (0, 0, 0). Then the distance between each human point cloud and the display screen is dᵢ = √(xᵢ² + yᵢ² + zᵢ²). The system calculates the average viewing distance d̄ = (Σdᵢ) / n and determines the chart scaling ratio s accordingly. The scaling ratio mapping function is defined as a piecewise function: when d̄ is in the near-distance interval [d min When d̄ is in the mid-range interval (d1, d2), s takes a larger value s1 (e.g., 1.5) to display more details; when d̄ is in the mid-range interval (d1, d2), s takes a standard value s2 (e.g., 1.0); when d̄ is in the long-range interval (d2, d3), s takes a larger value s1 (e.g., 1.5) to display more details. max When [the value is] , s takes a smaller value s3 (e.g., 0.6) to enhance readability. This mapping function takes into account the characteristics of human vision and optimizes the balance between information density and recognizability at different distances.

[0048] This spatial adaptation mechanism enables visualized content to intelligently respond to changes in the viewing environment, automatically adjusting the way information is presented based on the distribution of people. When viewed from a distance, it improves readability by increasing font size, simplifying labels, and aggregating data, while providing rich details and refined interactions when viewed up close, ensuring that users in different positions have the best information acquisition experience. Dynamic scaling avoids the readability problems caused by fixed layouts in traditional systems, significantly reducing the cognitive load on users and improving data understanding efficiency and decision-making quality.

[0049] For example, the step of determining several metadata items in the data generation library based on the question-and-answer prompts input to the display screen, and converting the several metadata items into a visual chart displayed on the display screen according to the chart scaling ratio, specifically includes: The question-and-answer prompts input to the display screen are semantically decomposed to obtain semantic vectors; In the data generation library, for each metadata, if the cosine similarity between the word sense vector and the data vector of the metadata in the data generation library is greater than a preset threshold, the metadata is used as a visualization target. For each visualization target, determine the vector subspace to which the metadata belongs; the vector space includes a comparison vector subspace, a trend vector subspace, a proportion vector subspace, a distribution vector subspace, and a relationship vector subspace; based on the data labels corresponding to all the data vectors in the vector space, perform K-means clustering on all the data vectors to obtain several cluster centers, and one cluster center corresponds to a center vector of a vector space; Metadata belonging to the comparison vector subspace is converted into a bar chart displayed on the screen according to the chart scaling ratio; metadata belonging to the trend vector subspace is converted into a line chart displayed on the screen according to the chart scaling ratio; metadata belonging to the proportion vector subspace is converted into a pie chart displayed on the screen according to the chart scaling ratio; metadata belonging to the distribution vector subspace is converted into a histogram displayed on the screen according to the chart scaling ratio; and metadata belonging to the relationship vector subspace is converted into a scatter plot displayed on the screen according to the chart scaling ratio.

[0050] The first step is semantic parsing and metadata matching: performing deep semantic understanding on the natural language query input by the user. Let the question-answering prompt be Q, which is converted into a high-dimensional word sense vector v_q ∈ ℝᵈ using a pre-trained language model (such as BERT). Each metadata element mⱼ in the data generation library has a corresponding data vector v_dⱼ ∈ ℝᵈ, which encodes the semantic features, statistical attributes, and business meaning of the metadata through offline training. The system calculates the cosine similarity between the word sense vector and each metadata vector: sim(v_q, v_dⱼ) = v_q·v_dⱼ / (||v_q||·||v_dⱼ||). When the similarity exceeds a preset threshold τ3, the metadata is marked as a visual target. The threshold τ3 controls the matching sensitivity; a high τ3 ensures accurate matching, while a low τ3 enhances semantic generalization ability. This mechanism enables intelligent conversion of natural language into data requirements, accurately understanding the true intent of user queries and finding relevant data sources even when queries are imprecise or contain business terminology. The construction of the semantic vector space gives the system context awareness, allowing it to distinguish the different business meanings of similar words (such as the difference between "growth" in financial and user scenarios). The threshold control mechanism balances matching accuracy and recall, ensuring that relevant data is not missed while avoiding interference from irrelevant data, significantly lowering the threshold for users to retrieve data and enabling non-technical personnel to efficiently obtain the information they need.

[0051] The second step is semantic space classification and chart mapping: mapping the metadata of the visualization target to a predefined semantic vector subspace. The entire vector space is divided into five subspaces: comparison space S_c (including comparison and ranking semantics), trend space S_t (including time change and increase / decrease semantics), proportion space S_p (including proportion and composition semantics), distribution space S_d (including distribution pattern and dispersion semantics), and relation space S_r (including correlation and dependency semantics). For each subspace S, all historical data vectors and their corresponding labels are collected, and the K-means clustering algorithm is applied to obtain k cluster centers c1, c2, ..., c_k, each center representing a typical semantic pattern of the space. For the visualization target vector v_d, its distance to the center vectors of each space is calculated, and it is assigned to the nearest subspace. The subspace type is then mapped to the corresponding chart: S_c → bar chart, S_t → line chart, S_p → donut chart, S_d → histogram, S_r → scatter plot. Finally, the visual parameters of the chart (font size, line width, label density, etc.) are adjusted according to the previously determined scaling ratio s. This classification mapping mechanism realizes the automatic conversion of data features to the best visualization form, ensuring that each data pattern is presented in a way that best conforms to human cognitive habits; semantic space clustering enables the system to understand the deep intent of complex queries (such as "market share change" which contains both percentage and trend, and the system selects the appropriate chart based on the dominant semantics); adaptive scaling ensures optimal readability at different viewing distances, optimizing the visual experience while maintaining information integrity; this intelligent mapping greatly reduces the cognitive burden on users in chart selection, making the data insight process more intuitive and smooth, and truly realizing a natural interactive experience of "what you think is what you get".

[0052] For example, in a real-world scenario: In a corporate decision-making meeting room, the marketing director asks the smart display screen, "Please show the percentage of quarterly sales revenue for each product line in East China." At this time, there are also two external partners (visitors) in the room. The system has already determined to use a visitor-level anonymized data sub-database, with an average viewing distance of 2.1 meters and a zoom ratio set to 0.7.

[0053] The processing flow is as follows: Semantic parsing: The system decomposes the query into key semantic units "East China Region" (geographic dimension), "Product Line" (category dimension), "Previous Quarter" (time range), "Sales Revenue" (metric), and "Percentage" (analysis type), generating a semantic vector v_q; Metadata matching: Calculate the similarity between v_q and all metadata vectors in the de-identification library, and match metadata such as "region_de-identification", "product category_aggregation", "sales cycle_quarter", and "amount_range". Spatial classification: The dominant semantics of "proportion status" points to the proportion vector subspace S_p, whose center vector is obtained by clustering historical proportion class queries; Visualization generation: Select a pie chart as the basic chart type. Due to the scaling ratio s=0.7 (medium to long distance), the title font size is automatically increased by 40%, the number of labels is reduced, and the specific amount is replaced with the percentage range (such as "30%-40%)", highlighting the top 3 product categories.

[0054] Finally, the screen displays a concise and clear pie chart, clearly showing the sales distribution of the East China product line after anonymization. Large font titles and simplified label design ensure readability from a distance, and key data points are highlighted with color. When the partner leaves, the system seamlessly switches to a detailed data view, restoring specific amounts and complete product categories, while adjusting to a denser information layout due to the reduced average distance. This intelligent visualization generation mechanism allows users from different backgrounds to obtain the best data insight experience matching their permissions and location in the same environment, significantly improving decision-making efficiency and collaboration quality.

[0055] Compared to existing technologies, this invention achieves intelligent data visualization within the constraints of a single display screen through a three-tiered linkage of spatial perception, identity recognition, and adaptive rendering. Its core value lies in a unified content strategy based on minimum access permissions and adaptive rendering based on average distance: when executives and clients are in the same room, the system automatically displays data content that meets the minimum access requirements, while optimizing the display effect according to the viewing distance. This ensures enterprise data security and compliance, improves long-distance readability, enhances meeting decision-making efficiency, and truly realizes "one screen, unified security, and adaptive experience" for intelligent enterprise data Q&A interaction.

[0056] One embodiment of this application provides a data visualization device for enterprise scenarios, including: a point cloud module 20, a detection module 21, an extraction module 22, a matching module 23, a determination module 24, and a display module 25.

[0057] The point cloud module 20 is used to acquire images captured by several cameras, and to perform point cloud modeling on each human body based on the captured images to obtain several human body point clouds. When performing point cloud modeling, the origin of the point cloud coordinate system is set at the geometric center of the display screen, and the X-axis direction of the point cloud coordinate system is consistent with the orientation of the display screen. The overlapping range of adjacent cameras is within a preset range, and the camera whose lens orientation is consistent with the orientation of the display screen is the main camera. The detection module 21 is used to perform face detection and alignment on the images captured by the main camera using the MTCNN model to obtain a number of face images; each face image corresponds to a human point cloud. The extraction module 22 is used to extract features from each face image using a pre-trained ArcFace model to obtain several multi-dimensional feature vectors; Matching module 23 is used to match the multidimensional feature vector with the employee feature vector in the employee feature library, and determine the personnel type corresponding to each human point cloud based on the matching result; The determination module 24 is used to determine the data generation library based on the personnel type corresponding to each of the human point clouds; and to determine the chart scaling ratio based on the point cloud distance between each of the human point clouds and the display screen. Display module 25 is used to determine several metadata in the data generation library based on the question and answer prompts input to the display screen; and to convert the several metadata into a visual chart displayed on the display screen according to the chart scaling ratio.

[0058] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the drug storage management device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0059] Compared to existing technologies, this invention achieves intelligent data visualization within the constraints of a single display screen through a three-tiered linkage of spatial perception, identity recognition, and adaptive rendering. Its core value lies in a unified content strategy based on minimum access permissions and adaptive rendering based on average distance: when executives and clients are in the same room, the system automatically displays data content that meets the minimum access requirements, while optimizing the display effect according to the viewing distance. This ensures enterprise data security and compliance, improves long-distance readability, enhances meeting decision-making efficiency, and truly realizes "one screen, unified security, and adaptive experience" for intelligent enterprise data Q&A interaction.

[0060] One embodiment of this application provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a data visualization method for enterprise scenarios as described above.

[0061] One embodiment of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a data visualization method for enterprise scenarios as described above.

[0062] The computer device may be a smartphone, tablet, desktop computer, or cloud server, among other computing devices. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the figures are merely examples of computer devices and do not constitute a limitation on the computer device. It may include more or fewer components than illustrated, or a combination of certain components, or different components, such as input / output devices, network access devices, etc.

[0063] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0064] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0065] This application provides a computer program product that, when run on a computer device, enables the computer device to execute the steps described in the various method embodiments above.

[0066] In the several embodiments provided in this application, it will be understood that each block in the flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the figures. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved.

[0067] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0068] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A data visualization method for enterprise scenarios, characterized in that, include: The system acquires images captured by several cameras, and performs point cloud modeling on each human body based on the captured images to obtain several human body point clouds. When performing point cloud modeling, the origin of the point cloud coordinate system is set at the geometric center of the display screen, and the X-axis direction of the point cloud coordinate system is consistent with the orientation of the display screen. The overlapping range of adjacent cameras is within a preset range, and the camera whose lens orientation is the same as that of the display screen is the main camera; The MTCNN model is used to perform face detection and alignment on the images captured by the main camera to obtain several face images; each face image corresponds to a human point cloud. The pre-trained ArcFace model is used to extract features from each face image to obtain several multi-dimensional feature vectors; The multidimensional feature vector is matched with the employee feature vector in the employee feature library, and the personnel type corresponding to each human point cloud is determined according to the matching result. A data generation library is determined based on the personnel type corresponding to each human point cloud; the chart scaling ratio is determined based on the point cloud distance between each human point cloud and the display screen. Based on the question-and-answer prompts input to the display screen, several metadata items are determined in the data generation library; the metadata items are then converted into a visual chart displayed on the display screen according to the chart scaling ratio.

2. The data visualization method for enterprise scenarios as described in claim 1, characterized in that, The step of performing point cloud modeling on each human body based on the captured images to obtain several human body point clouds specifically includes: The ICP algorithm is used to align images captured by different cameras to obtain aligned point clouds; PointNet was used to separate individual human point clouds from the aligned point cloud.

3. The data visualization method for enterprise scenarios as described in claim 1, characterized in that, Before using the pre-trained ArcFace model to extract features from each face image and obtain several multi-dimensional feature vectors, the process specifically includes: Collect and clean a set of facial images of the company's employees, where each employee has more than three facial images, and more than one of the facial images was taken in front of the display screen; The base ArcFace model was trained using ArcFace loss on a publicly available face dataset; By combining the set of facial images of the company's employees, the basic ArcFace model is adjusted using multi-task training.

4. The data visualization method for enterprise scenarios as described in claim 3, characterized in that, The step of combining the set of facial images of the company's employees and adjusting the base ArcFace model using multi-task training specifically includes: Load the set of facial images of the company's employees, and occlude and rotate the faces in the images within the set by a preset ratio to obtain a fine-tuned training set; The weighted sum of the arc surface loss function, the triple loss function, and the center loss function is used as the optimization objective. The base ArcFace model is trained using the fine-tuned training set to obtain a pre-trained ArcFace model. The weight factor of the arc surface loss function is 0.9, the weight factor of the triple loss function is 0.3, and the weight factor of the center loss function is 0.

01.

5. The data visualization method for enterprise scenarios as described in claim 1, characterized in that, The step of matching the multidimensional feature vector with the employee feature vector in the employee feature database, and determining the personnel type corresponding to each human point cloud based on the matching result, specifically includes: In three consecutive frames, if the cosine similarity between the multidimensional feature vector and an employee feature vector in the employee feature library is greater than or equal to a preset threshold, the corresponding personnel type of the human point cloud associated with the multidimensional feature vector is an enterprise employee; if the cosine similarity between the multidimensional feature vector and an employee feature vector in the employee feature library is less than the preset threshold, the corresponding personnel type of the human point cloud associated with the multidimensional feature vector is a visitor; enterprise employees include several levels, and different levels of enterprise employees can access different enterprise databases.

6. The data visualization method for enterprise scenarios as described in claim 5, characterized in that, The data generation library is determined based on the personnel type corresponding to each of the human point clouds; The chart scaling ratio is determined based on the distance between each human point cloud and the point cloud of the display screen, specifically including: If the person type corresponding to the human point cloud is a visitor, the lowest-level enterprise data sub-database is used as the data generation database; if the person type corresponding to all the human point clouds is an enterprise employee, the lowest-level enterprise employee is determined among all the person types corresponding to the human point clouds, and the enterprise data sub-database corresponding to that level of enterprise employee is selected as the data generation database; the enterprise data sub-database corresponding to the lowest-level enterprise employee is one level higher than the lowest-level enterprise data sub-database; the lower-level enterprise data sub-database is obtained by distilling the higher-level enterprise data sub-database. Calculate the average point cloud distance between each of the human body point clouds and the display screen; Select the chart scaling ratio based on the average point cloud distance.

7. The data visualization method for enterprise scenarios as described in claim 1, characterized in that, The process of determining several metadata items in the data generation library based on the question-and-answer prompts input to the display screen, and converting the several metadata items into a visual chart displayed on the display screen according to the chart scaling ratio, specifically includes: The question-and-answer prompts input to the display screen are semantically decomposed to obtain semantic vectors; In the data generation library, for each metadata, if the cosine similarity between the word sense vector and the data vector of the metadata in the data generation library is greater than a preset threshold, the metadata is used as a visualization target. For each visualization target, determine the vector subspace to which the metadata belongs; the vector space includes a comparison vector subspace, a trend vector subspace, a proportion vector subspace, a distribution vector subspace, and a relationship vector subspace; based on the data labels corresponding to all the data vectors in the vector space, perform K-means clustering on all the data vectors to obtain several cluster centers, and one cluster center corresponds to a center vector of a vector space; Metadata belonging to the comparison vector subspace is converted into a bar chart displayed on the screen according to the chart scaling ratio; metadata belonging to the trend vector subspace is converted into a line chart displayed on the screen according to the chart scaling ratio; metadata belonging to the proportion vector subspace is converted into a pie chart displayed on the screen according to the chart scaling ratio; metadata belonging to the distribution vector subspace is converted into a histogram displayed on the screen according to the chart scaling ratio; and metadata belonging to the relationship vector subspace is converted into a scatter plot displayed on the screen according to the chart scaling ratio.

8. A data visualization device for enterprise scenarios, characterized in that, include: The point cloud module is used to acquire images captured by several cameras, and to perform point cloud modeling on each human body based on the captured images to obtain several human body point clouds. When performing point cloud modeling, the origin of the point cloud coordinate system is set at the geometric center of the display screen, and the X-axis direction of the point cloud coordinate system is consistent with the orientation of the display screen. The overlapping range of adjacent cameras is within a preset range, and the camera whose lens orientation is the same as that of the display screen is the main camera; The detection module is used to perform face detection and alignment on the images captured by the main camera using the MTCNN model to obtain several face images; each face image corresponds to a human point cloud. The extraction module is used to extract features from each face image using a pre-trained ArcFace model, resulting in several multi-dimensional feature vectors. The matching module is used to match the multidimensional feature vector with the employee feature vector in the employee feature library, and determine the personnel type corresponding to each human point cloud based on the matching result; The determination module is used to determine the data generation library based on the personnel type corresponding to each of the human point clouds; The chart scaling ratio is determined based on the distance between each human point cloud and the point cloud of the display screen; The display module is used to determine several metadata items in the data generation library based on the question-and-answer prompts input to the display screen; and to convert the several metadata items into a visual chart displayed on the display screen according to the chart scaling ratio.

9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a data visualization method for enterprise scenarios as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a data visualization method for enterprise scenarios as described in any one of claims 1 to 7.

Images