An e-commerce warehouse center safety monitoring method based on deep learning

Abstract

The application discloses an e-commerce warehouse center safety monitoring method based on deep learning, and relates to the technical field of warehouse safety vision, comprising the following steps: carrying out space-time alignment fusion on structured scene data and laser point cloud data, obtaining a three-dimensional point cloud representation, carrying out point-level segmentation and high-risk area identification on the three-dimensional point cloud representation through a PointNet network, and obtaining a warehouse high-risk area; abstracting scene objects in the structured scene data into graph nodes, establishing an interaction relationship graph according to the three-dimensional space representation, analyzing the interaction relationship between the scene objects in combination with a space risk result, generating an interaction risk result and a three-dimensional position marker; generating a three-dimensional visual picture according to the interaction risk result and the three-dimensional position marker, labeling a dangerous area and a risk object, and forming a warehouse risk visual report. The application significantly improves the automation, refinement and visualization level of warehouse safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of warehouse safety vision technology, and in particular to a deep learning-based method for security monitoring of e-commerce warehouses. Background Technology

[0002] With the rapid development of e-commerce, warehousing centers play a crucial role in the circulation and distribution of goods. In recent years, with the maturity of deep learning technology, the application of convolutional neural networks in image recognition and object detection has been continuously expanding, significantly improving the level of automated monitoring based on video data. Simultaneously, with the development of LiDAR and 3D point cloud technology, warehouse safety monitoring is gradually transitioning from two-dimensional image analysis to three-dimensional spatial modeling. 3D point clouds can provide detailed information on the spatial structure of a scene and the geometric shape of objects, offering new technical support for spatial modeling and risk identification in complex warehouse environments.

[0003] Existing warehouse safety monitoring technologies based on single video data still have certain limitations in handling multi-source information fusion and spatial risk identification. Current methods often remain at the level of simple time alignment and information stitching during multi-source data fusion, failing to effectively achieve deep coupling between the semantic and geometric layers, making it difficult to accurately identify potentially high-risk areas in dynamic warehouse scenarios. Warehouse safety risks not only manifest as the dangerous state of a single object but may also involve interactive risks between multiple objects, such as the tilting of shelves and personnel approaching, or the relative movement of handling vehicles and obstacles. However, existing technologies have limited capabilities in analyzing such interactive risks. This, to some extent, restricts the interpretability and visualization level of risk monitoring results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based security monitoring method for e-commerce warehouse centers to address the problems of difficulty in deep coupling between the semantic and geometric layers of multi-source data and insufficient identification of interactive risks in warehouse scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based security monitoring method for e-commerce warehouse centers, which includes collecting high-definition video data and laser point cloud data of the warehouse, and processing them synchronously through timestamps to form multi-source monitoring data;

[0008] Convolutional neural networks are used to perform semantic segmentation and spatial information extraction on multi-source monitoring data to obtain structured scene data.

[0009] Structured scene data and laser point cloud data are spatiotemporally aligned and fused to obtain a 3D point cloud representation. PointNet network is then used to perform point-level segmentation and high-risk area identification on the 3D point cloud representation to obtain high-risk areas in the warehouse.

[0010] By combining high-risk storage areas with spatial location mapping tables, spatial risk results are generated.

[0011] The scene objects in the structured scene data are abstracted into graph nodes, and an interaction relationship graph is established based on the three-dimensional spatial representation. The interaction relationship between scene objects is analyzed in combination with the spatial risk results, and interaction risk results and three-dimensional position markers are generated.

[0012] Based on the interactive risk results and 3D location markers, a 3D visualization is generated, and hazardous areas and risk objects are marked to form a warehouse risk visualization report.

[0013] As a preferred embodiment of the deep learning-based e-commerce warehouse security monitoring method of the present invention, the step of forming multi-source monitoring data through timestamp synchronization processing is as follows:

[0014] Align the timestamps of warehouse high-definition video data and laser point cloud data to generate time-aligned warehouse data;

[0015] Perform video frame format unification and point cloud coordinate transformation on time-aligned warehouse data to generate formatted multi-source warehouse data;

[0016] Formatted multi-source warehouse data is combined according to time series to form multi-source monitoring data.

[0017] As a preferred embodiment of the deep learning-based e-commerce warehouse security monitoring method of the present invention, the specific steps for acquiring structured scene data are as follows:

[0018] The warehouse high-definition video data and laser point cloud data in the multi-source monitoring data are spatially mapped and fused according to time, and divided into voxel grids to generate voxel multi-source data tensors.

[0019] Voxelized multi-source data tensors are input into a convolutional neural network for semantic segmentation, and voxel-level semantic classification results are output.

[0020] The voxel-level semantic classification results are subjected to 3D coordinate back mapping and scene object extraction. The scene objects are organized according to category and spatial hierarchy to generate structured scene data.

[0021] As a preferred embodiment of the deep learning-based e-commerce warehouse security monitoring method of the present invention, the specific steps for obtaining the 3D point cloud representation are as follows:

[0022] The 3D position of each scene object in the structured scene data is compared with the laser point cloud data according to the timestamp, and the 3D position of the scene object is initially spatially registered to generate an aligned dataset.

[0023] Each record in the aligned dataset is split into a scene object and a corresponding point cloud subset, and the semantic labels of the scene object and the local geometric features of the corresponding point cloud subset are extracted respectively.

[0024] Semantic labels and local geometric features are used as two branches and input into a multi-layer interactive attention network for bidirectional information interaction and fusion to generate a fused feature set.

[0025] Non-uniform density sampling and local surface reconstruction are performed on the fused feature set, and all sampled and reconstructed point cloud subsets are integrated to generate a 3D point cloud representation.

[0026] As a preferred embodiment of the deep learning-based e-commerce warehouse security monitoring method of the present invention, the specific steps for identifying high-risk areas in the warehouse are as follows:

[0027] The 3D point cloud representation is input into the PointNet network, and the spatial geometric features and local neighborhood features of each point cloud are extracted to obtain a set of point-level features.

[0028] The point-level feature set is input into the classification layer for point-by-point discrimination, and the set of scene regions is generated by aggregation through point labels;

[0029] By classifying the scene regions and combining their spatial location, volume, and interaction with neighboring objects, risk characteristic analysis is performed to identify a preliminary set of high-risk region candidates with potential hazardous attributes.

[0030] The initial high-risk area candidate set is screened and confirmed through spatial consistency verification, threshold screening, and historical high-risk pattern matching, and the high-risk warehouse areas are output.

[0031] As a preferred embodiment of the deep learning-based security monitoring method for e-commerce warehouse centers described in this invention, the specific steps for combining high-risk warehouse areas with a spatial location mapping table to generate spatial risk results are as follows.

[0032] The center coordinates and boundary range of the high-risk storage area are extracted from the 3D point cloud representation to generate the corresponding spatial location identifier; the spatial location identifier is matched with the spatial location mapping table to obtain the mapping position relationship of the high-risk storage area.

[0033] Based on the mapping positional relationship, the spatial proximity relationship and relative position between each high-risk storage area and surrounding scene objects are analyzed to obtain the spatial interaction relationship of the high-risk storage area;

[0034] By combining spatial interaction relationships, the distribution characteristics of high-risk areas in warehousing, and the interaction risks of surrounding scene objects, spatial risk results for warehousing scenarios are generated.

[0035] As a preferred embodiment of the deep learning-based security monitoring method for e-commerce warehousing centers described in this invention, the spatial location mapping table is obtained by spatially dividing the three-dimensional point cloud representation in the warehousing scene, aligning the three-dimensional coordinates of each scene object with the actual layout information of the warehousing area, and recording the area identifier, location range, and object distance relationship in tabular form.

[0036] As a preferred embodiment of the deep learning-based security monitoring method for e-commerce warehouse centers described in this invention, the specific steps of abstracting scene objects in structured scene data into graph nodes and establishing an interaction relationship graph based on the three-dimensional spatial representation are as follows:

[0037] Extract the unique identifier, category attribute, and three-dimensional spatial coordinate information of each scene object from the structured scene data;

[0038] A graph node set is constructed by using unique identifiers as node indices, category attributes as node features, and 3D spatial coordinate information as node position vectors.

[0039] The spatial proximity relationship between scene objects is calculated based on the node position vector, and the candidate set of interaction edges is determined by combining the category attribute;

[0040] The interaction edges in the candidate set are combined with the graph node set to generate an interaction relationship graph.

[0041] As a preferred embodiment of the deep learning-based security monitoring method for e-commerce warehouse centers described in this invention, the specific steps for generating interactive risk results and three-dimensional location markers are as follows:

[0042] The spatial risk results are mapped to the nodes of the interaction graph to obtain the risk weight corresponding to each scene object;

[0043] Based on risk weights and the interaction edge features between graph nodes, the risk transmission relationship between scene objects is evaluated, and the interaction risk results are generated.

[0044] Based on the interaction risk results, mark the location of the corresponding scene object in three-dimensional space and output the three-dimensional location mark.

[0045] As a preferred embodiment of the deep learning-based e-commerce warehouse security monitoring method of the present invention, the specific steps for generating a warehouse risk visualization report are as follows:

[0046] By fusing the interactive risk results with the 3D location markers, the 3D spatial layout of the warehouse scene is reconstructed, and a 3D visualization is output.

[0047] The system highlights dangerous areas in the 3D visualization and marks risky objects to make them more prominent, then outputs the marked 3D visualization.

[0048] The system performs risk weight accumulation and spatial interaction analysis on dangerous areas and risky objects in the 3D visualization. When the risk level reaches the alarm condition, it triggers an early warning and generates a warehouse risk visualization report.

[0049] The beneficial effects of this invention are as follows: by synchronizing and multi-layering high-definition video and laser point cloud data in the warehouse, a high-precision three-dimensional point cloud representation is generated, which enables accurate identification of high-risk areas in the warehouse; at the same time, by abstracting structured scene objects into graph nodes and constructing an interaction relationship graph, the potential risks between multiple objects are quantified and a three-dimensional visualization is generated, which enables intuitive labeling and automatic early warning of dangerous areas and risky objects, thereby significantly improving the automation, refinement and visualization level of warehouse safety monitoring. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a deep learning-based security monitoring method for e-commerce warehouses.

[0052] Figure 2 A flowchart for the formation of multi-source monitoring data.

[0053] Figure 3 A flowchart for identifying high-risk areas in a warehouse.

[0054] Figure 4 A flowchart for generating visualization and early warning results of warehouse risks. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figures 1-4 This is one embodiment of the present invention, which provides a deep learning-based security monitoring method for e-commerce warehouse centers, comprising the following steps:

[0059] S1. Collect high-definition video data and laser point cloud data from the warehouse, and process them synchronously through timestamps to form multi-source monitoring data.

[0060] It should be noted that the high-definition video data of the warehouse includes image frame sequences and timestamp information of scene objects such as warehouse shelves, pallets, aisles, personnel and handling equipment.

[0061] Laser point cloud data includes the three-dimensional spatial coordinates, reflection intensity, and timestamp information of objects such as shelves, stacks, equipment, and personnel in the warehouse environment.

[0062] S1.1 Align the timestamps of the warehouse high-definition video data and laser point cloud data to generate time-aligned warehouse data.

[0063] Furthermore, the timestamps of the warehouse high-definition video data and laser point cloud data are unified and standardized, anomalies are removed, and frame-by-frame correspondence is achieved through nearest neighbor matching or time window aggregation. At the same time, motion compensation and extrinsic parameter transformation are combined to align the point cloud to the time point of the video frame, generating time-aligned warehouse data containing a unified timestamp, image frame and corresponding point cloud.

[0064] It should be noted that motion compensation refers to the process in which spatial offset still exists after time synchronization due to the relative motion of cameras or sensors during multi-source data acquisition. By estimating motion trajectory and velocity parameters, the position of image frames or point cloud data is dynamically corrected to ensure the spatial consistency of multi-source data at the same point in time.

[0065] External parameter transformation refers to the difference in spatial coordinate systems between different sensors. By obtaining the rotation matrix and translation vector of the camera and lidar through sensor calibration experiments, the coordinates of point cloud data are uniformly transformed into the video image coordinate system, thereby achieving cross-modal spatial alignment.

[0066] S1.2 Perform video frame format unification and point cloud coordinate transformation on time-aligned warehouse data to generate formatted multi-source warehouse data.

[0067] Furthermore, the encoding method, resolution, and frame rate of the high-definition video data in the time-aligned warehouse data are unified to ensure that the image frame sequence has a consistent timing and format; coordinate system transformation and normalization are performed on the laser point cloud data to unify the laser point cloud data into a unified spatial coordinate system; after the above processing, formatted multi-source warehouse data is obtained.

[0068] S1.3 Combine the formatted multi-source warehouse data according to the time series to form multi-source monitoring data.

[0069] Furthermore, the high-definition video data and laser point cloud data of the warehouse in the formatted multi-source warehouse data are arranged sequentially according to a unified timestamp to ensure that the information from different sources corresponds frame by frame on the timeline; the records in the time series are combined, and the image frames corresponding to the same time point are bound to the laser point cloud data to form a complete record; all time series records are collected in sequence to form multi-source monitoring data.

[0070] S2. Use convolutional neural networks to perform semantic segmentation and spatial information extraction on multi-source monitoring data to obtain structured scene data.

[0071] S2.1. Perform time-corresponding spatial mapping and fusion of warehouse high-definition video data and laser point cloud data from multi-source monitoring data, and divide them into voxelized meshes to generate voxelized multi-source data tensors.

[0072] Furthermore, based on timestamps, nearest neighbor matching or fixed-time-window aggregation methods are used to establish frame-to-point correspondences between warehouse high-definition video data frames and laser point cloud data records. When frame skipping or frame loss occurs, motion-compensated inter-frame displacement estimation is used for interpolation or frame supplementation to ensure temporal consistency. Camera intrinsic and extrinsic parameters are extracted, and extrinsic parameter transformation is performed to transform the laser point cloud data from the laser coordinate system to the camera coordinate system. The camera projection matrix is ​​used to project the three-dimensional coordinates onto the pixel plane of the warehouse high-definition video data to establish a pixel-point cloud correspondence. For each laser point cloud data point projected to a pixel, three-dimensional geometric features and laser reflection intensity are collected, and color information is read from the corresponding pixel. When multiple When multiple laser points are projected onto the same pixel, a strategy of merging pixel association information is adopted, either based on the nearest point in depth or by weighted averaging of reflection intensity. A regular voxelized mesh is divided in the 3D workspace according to a preset voxel size. Laser point cloud data points falling within the same voxel range and their corresponding pixel color information are aggregated into voxel features. Aggregation items include point count, 3D position mean, reflection intensity mean, and pixel color mean. All voxel features are then concatenated by tensor channels to form a voxelized multi-source data tensor. Examples of tensor channels include occupancy probability, point count, 3D position mean, reflection intensity mean, and pixel color mean. Voxel coordinates and timestamps are kept in a one-to-one correspondence to ensure spatiotemporal consistency.

[0073] It should be noted that the voxel size is usually set according to the three-dimensional spatial scale and monitoring accuracy requirements of the warehousing scenario. For example, the height of a single layer of storage rack or the width of the aisle can be used as a reference value. By balancing resolution and computing efficiency, an appropriate cube side length is selected so that the voxel division can cover the entire warehousing area.

[0074] S2.2 Input the voxelized multi-source data tensor into the convolutional neural network for semantic segmentation and output the voxel-level semantic classification result.

[0075] Furthermore, the voxelized multi-source data tensors are normalized and their dimensions adjusted so that each voxel feature can serve as the input channel for convolutional operations. Local spatial features are extracted through 3D convolutional layers in the convolutional neural network, redundant information is compressed and stable features (such as spatial structure continuity and voxel distribution patterns) are preserved through pooling layers, and high-level semantic features are gradually extracted through multi-layer convolution and nonlinear activation functions. Upsampling layers or deconvolutional layers are used to restore voxel-level spatial resolution, and the extracted high-level semantic features are mapped back to the original voxel coordinates. In the classification layer, the high-level semantic features of each voxel are discriminated voxel-by-voxel, and a probability distribution corresponding to the category set is output. Semantic categories are assigned to each voxel based on the label with the highest probability, forming voxel-level semantic classification results, including categories such as warehouse aisle areas, shelf areas, equipment areas, personnel areas, and potential obstacle areas.

[0076] It should be noted that training a convolutional neural network involves forming training samples from voxelized multi-source data tensors and corresponding voxel-level semantic labels, inputting them into the convolutional neural network for forward propagation to obtain prediction results, calculating the cross-entropy loss between the prediction results and the true labels, updating the network parameters through backpropagation, and iterating multiple times until the convolutional neural network converges and has a stable voxel-level semantic classification capability.

[0077] S2.3 Perform three-dimensional coordinate back mapping and scene object extraction on the voxel-level semantic classification results, organize each scene object according to category and spatial hierarchy, and generate structured scene data.

[0078] Furthermore, the voxel-level semantic classification results are mapped to the three-dimensional spatial coordinates of each voxel in the voxelized multi-source data tensor, achieving a one-to-one correspondence between voxel semantic categories and three-dimensional positions. In the three-dimensional coordinate space, voxels with the same semantic category and spatial adjacency are aggregated into continuous regions, and different object regions are separated by connectivity analysis or clustering methods. Spatial features such as boundary range, center coordinates, and spatial scale are extracted from each aggregated region, and the corresponding scene objects are identified by combining semantic category annotations. All scene objects are classified into categories such as shelves, pallets, personnel, and handling equipment according to semantic categories, and a hierarchical organizational structure is established based on the three-dimensional spatial positional relationships to generate complete structured scene data.

[0079] S3. The structured scene data and laser point cloud data are spatiotemporally aligned and fused to obtain a 3D point cloud representation. The PointNet network is used to perform point-level segmentation and high-risk area identification on the 3D point cloud representation to obtain high-risk areas in the warehouse.

[0080] It should be noted that existing methods typically involve simply overlaying structured scene data with laser point cloud data or aligning them through rigid body registration to obtain a 3D point cloud representation. Then, traditional 3D point cloud segmentation algorithms or single-branch deep learning networks are used to classify each point and identify potential high-risk areas. However, these methods are insufficient in the fusion of semantic information and local geometric features, making it difficult to fully utilize the semantic and spatial relationships of scene objects for accurate identification of high-risk areas.

[0081] Our invention generates an aligned dataset by precisely aligning structured scene data and laser point cloud data according to timestamps. Each scene object is then mapped to a corresponding subset of the point cloud, and semantic labels are extracted from the scene objects, while local geometric features are extracted from the corresponding point cloud subsets. The semantic labels and local geometric features are then used as dual-branch inputs to a multi-layer interactive attention network for bidirectional information exchange and fusion, generating a fused feature set. This fused feature set is then represented as a 3D point cloud through non-uniform density sampling and local surface reconstruction. Finally, the feature set is input into a PointNet point cloud neural network to extract point-level features and perform point-by-point discrimination and risk feature analysis, ultimately obtaining accurate high-risk areas in warehousing.

[0082] S3.1. Compare the 3D position of each scene object in the structured scene data with the laser point cloud data according to the timestamp, and perform preliminary spatial registration on the 3D position of the scene objects to generate an aligned dataset.

[0083] Furthermore, the timestamps carried by each scene object in the structured scene data are matched one-to-one with the timestamps of the laser point cloud data to ensure that the 3D position of the scene object is consistent with the laser point cloud data at the same time point. Based on the time correspondence, the 3D position of the scene object is mapped to the spatial coordinate system of the laser point cloud data, and the spatial reference is unified through extrinsic parameter information. The spatial proximity constraint is used to perform preliminary spatial registration between the 3D position of the scene object and the laser point cloud data, so that the scene object and the point cloud maintain a reasonable correspondence in 3D space, and the aligned dataset is output.

[0084] It should be noted that extrinsic parameters refer to the external parameters between the camera or sensor, including rotation matrices and translation vectors.

[0085] Spatial proximity constraints refer to constraints based on the degree of proximity of spatial locations. For example, the Euclidean distance or neighborhood relationship between the 3D position of a scene object and laser point cloud data points can be used to determine the matching relationship.

[0086] S3.2. Split each record in the aligned dataset into a scene object and a corresponding point cloud subset, and extract the semantic labels of the scene object and the local geometric features of the corresponding point cloud subset respectively.

[0087] Furthermore, each record in the aligned dataset is decomposed according to the scene object, and each scene object is extracted independently from the record. At the same time, a set of points corresponding to the 3D position space of the scene object is selected from the laser point cloud data to form a corresponding point cloud subset, and a one-to-one correspondence is established between the scene object and the corresponding point cloud subset. The semantic category labels (e.g., warehouse aisle area, shelf area, equipment area, personnel area, and potential obstacle area) contained in the voxel-level semantic classification results are directly read from the scene object. A set of 3D points belonging to the scene object is selected from the corresponding point cloud subset. Geometric shape information is obtained by statistically analyzing the spatial distribution of 3D points. The density distribution of the point set reflects the local point density. The neighborhood point set is determined by the K-nearest neighbor search method based on Euclidean distance. The covariance matrix of the neighborhood points is calculated by the principal component analysis method and the eigenvalues ​​are extracted to reflect the local curvature changes. The local geometric features are obtained by combining the normal vector estimation based on the least squares method.

[0088] S3.3. Semantic labels and local geometric features are used as two branches and input into a multi-layer interactive attention network for bidirectional information interaction and fusion to generate a fused feature set.

[0089] Furthermore, the semantic labels of each scene object and the local geometric features of the corresponding laser point cloud subset are respectively used as dual-branch inputs to the multi-layer interactive attention network. Through the multi-layer attention mechanism, feature query, key-value matching, and weighted information transmission are performed between the two branches to achieve bidirectional interactive fusion of semantic information and geometric features. In each attention layer, the feature vector of the semantic label branch is updated to be affected by local geometric features, and the feature vector of the local geometric feature branch is updated to be affected by semantic labels. Global and local spatial interaction features are extracted layer by layer, and a fused feature set is output.

[0090] It should be noted that training the multi-layer interactive attention network involves taking the semantic labels of each scene object and the local geometric features of the corresponding laser point cloud subset as input, using the supervision labels generated by voxel-level semantic annotation and point cloud local geometric feature extraction, and inputting them into the multi-layer interactive attention network for forward propagation to obtain the predicted fusion features. The difference between the predicted fusion features and the supervision labels is measured by a defined loss function, and the parameters of the multi-layer interactive attention network are updated using a backpropagation algorithm. After multiple iterations, the multi-layer interactive attention network converges and can stably generate the fusion feature set.

[0091] S3.4 Perform non-uniform density sampling and local surface reconstruction on the fused feature set, and integrate all sampled and reconstructed point cloud subsets to generate a three-dimensional point cloud representation.

[0092] Furthermore, for the laser point cloud subset corresponding to each scene object in the fusion feature set, non-uniform density sampling points are selected based on the differences in point distribution density to preserve details in high-density areas and reduce redundant points in low-density areas. Surface reconstruction is performed in the local neighborhood of each sampling point to generate local surface fragments that reflect the local spatial shape features of the scene object. All local surface fragments and sampling point sets are integrated and aligned according to their three-dimensional spatial positions to generate a continuous three-dimensional point cloud representation.

[0093] S3.5 Input the 3D point cloud representation into the PointNet network, extract the spatial geometric features and local neighborhood features of each point cloud, and obtain a set of point-level features.

[0094] Furthermore, the 3D point set of each scene object in the 3D point cloud representation is input into the PointNet network. The spatial geometric features of the position coordinates of each 3D point are encoded through a point-by-point feature extraction layer. At the same time, the spatial coordinates of each point in the local neighborhood are input into a multilayer perceptron. The feature representation of the point is calculated using layer-by-layer nonlinear mapping. Then, weight sharing is used to ensure that the feature extraction method of different points is consistent. Finally, the features of the neighborhood points are aggregated (such as max pooling) to extract the local spatial distribution pattern and form local neighborhood features. In each layer of the PointNet network, local and global spatial structure information is captured through max pooling or global feature aggregation. Finally, the point-level feature set of each 3D point is output.

[0095] It should be noted that the PointNet network is trained by using supervised labels generated from voxelized multi-source data tensors and fused feature sets. Point-level features are extracted through forward propagation and the error between the predicted features and the supervised labels is calculated. The PointNet network parameters are updated using the backpropagation algorithm. This process is repeated multiple times until the PointNet network converges, ensuring that the spatial geometric features and local neighborhood features of each 3D point can be stably extracted.

[0096] S3.6 Input the point-level feature set into the classification layer for point-by-point discrimination, and generate a scene region set by aggregating the point labels.

[0097] Furthermore, the point-level feature set of each 3D point is input into the classification layer of the PointNet network, and a class label is assigned to each 3D point using a point-by-point classification method. 3D points with the same class are aggregated according to spatial proximity and class consistency to form continuous spatial regions. During the aggregation process, the geometric structure and semantic information of each scene object are preserved, and a set of scene regions is output.

[0098] It should be noted that spatial proximity relationships are established by calculating the Euclidean distance of each 3D point in the 3D coordinate system or by using the k-Nearest Neighbors (k-NN) method to determine the set of neighboring points of each point within a local range.

[0099] S3.7. The scene area set is labeled according to category and combined with spatial location, volume and interaction relationship of neighboring objects to conduct risk feature analysis, and identify a preliminary high-risk area candidate set with potential dangerous attributes.

[0100] Furthermore, the three-dimensional spatial location and volume range of each scene area are associated with category labels to characterize the spatial scale features of the scene area. By comparing the three-dimensional distance, overlapping volume, and relative positional relationship between each scene area and surrounding scene areas, and considering mutual occlusion or proximity, the spatial proximity and potential interaction between areas are evaluated to form interaction features. For example, there may be stacking collision risks between adjacent shelves, occlusion and interference risks between mobile pallets and fixed shelves, and the risk of passage obstruction due to the overlap of stacking areas and aisles. Based on the combination of category labels, spatial location, volume range, and interaction features of neighboring objects, risk features with potential dangerous attributes are extracted, and a preliminary high-risk area candidate set is selected based on the risk features.

[0101] S3.8. The preliminary high-risk area candidate set is screened and confirmed through spatial consistency verification, threshold screening and historical high-risk pattern matching, and the high-risk warehouse areas are output.

[0102] Furthermore, for each scene region in the initial high-risk area candidate set, the point cloud coverage and spatial distribution consistency at the same 3D location are compared in a continuous time series. If the positional offset and spatial morphological changes are within an acceptable range over multiple frames, the scene region is retained; otherwise, it is excluded. Subsequently, volume, distance to neighboring objects, and interaction features are extracted for each scene region. The risk feature values ​​are then compared with the high-risk determination threshold, and only scene regions with risk feature values ​​higher than the high-risk determination threshold are retained. The scene regions in the 3D point cloud representation are spatially overlapped and morphologically similar to the high-risk regions marked in the historical high-risk patterns. Scene regions with high matching degree and spatial consistency are confirmed and output as high-risk warehouse areas.

[0103] It should be noted that the high-risk determination threshold is determined by analyzing the distribution of volume, distance to nearby objects, and interaction characteristics in past high-risk events. A reference value that can effectively distinguish between high-risk and low-risk areas is selected as the preset high-risk determination threshold (the value range is usually between 0 and 1). For example, sample values ​​that are slightly higher or lower than the characteristic range of historical high-risk areas can be selected as risk screening criteria.

[0104] S4. Combine high-risk storage areas with spatial location mapping tables to generate spatial risk results.

[0105] S4.1 Extract the center coordinates and boundary range of the high-risk storage area in the 3D point cloud representation and generate the corresponding spatial location identifier; match the spatial location identifier with the spatial location mapping table to obtain the mapping position relationship of the high-risk storage area.

[0106] Furthermore, the system identifies the 3D point cloud boundary points of each high-risk area in the warehouse and determines the farthest and nearest spatial locations within the boundary range. Then, it obtains the average position of the 3D point cloud within the boundary range as the center coordinates, and combines the center coordinates and the boundary range to form the corresponding spatial location identifier. The spatial location identifier is matched with the area identifier, location range, and object distance relationship recorded in the spatial location mapping table. By comparing the center coordinates and boundary range with the spatial information of each area in the spatial location mapping table, the system confirms the corresponding position of each high-risk warehouse area in the spatial location mapping table and outputs the mapping position relationship set for each high-risk warehouse area.

[0107] It should be noted that the spatial location mapping table is obtained by dividing the space of the 3D point cloud representation in the warehousing scene, aligning the 3D coordinates of each scene object with the actual layout information of the warehousing area, and recording the area identification, location range and object distance relationship in tabular form.

[0108] S4.2. Based on the mapping positional relationship, analyze the spatial proximity and relative position between each high-risk storage area and the surrounding scene objects to obtain the spatial interaction relationship of the high-risk storage area.

[0109] Furthermore, based on the mapping positional relationship of high-risk storage areas, the scene objects around each high-risk area are identified, and the three-dimensional spatial coordinates of the scene objects are extracted. By comparing the center coordinate distance, boundary range, and spatial direction of the high-risk area and the surrounding scene objects, the proximity relationship and relative positional relationship between the high-risk area and each scene object are determined. The spatial proximity relationship and relative positional relationship between all high-risk areas and the surrounding scene objects are integrated to form the spatial interaction relationship of high-risk storage areas.

[0110] S4.3 Combine spatial interaction relationships, the distribution characteristics of high-risk areas in the warehouse, and the interaction risks of surrounding scene objects to generate spatial risk results for the warehouse scene.

[0111] Furthermore, the spatial interaction relationships of high-risk storage areas are integrated with the distribution characteristics of each high-risk area, including location density, volume distribution, and the number of neighboring objects. At the same time, the interaction risk characteristics of surrounding scene objects are extracted, such as their proximity to high-risk areas and potential interference relationships. The spatial interaction relationships, the distribution characteristics of high-risk storage areas, and the interaction risk information of surrounding scene objects (including proximity distance, spatial coverage overlap, relative position changes, and interaction impact weights) are correlated and fused to form the spatial risk results of the storage scene.

[0112] S5. Abstract scene objects in structured scene data into graph nodes, and establish an interaction relationship graph based on the three-dimensional spatial representation. Combine the spatial risk results to analyze the interaction relationships between scene objects and generate interaction risk results and three-dimensional position markers.

[0113] S5.1 Extract the unique identifier, category attribute, and three-dimensional spatial coordinate information of each scene object from the structured scene data.

[0114] Furthermore, the unique identifier of the scene object in the structured scene data is read to distinguish different scene object instances, and the category attributes of the scene object (such as shelves, pallets, personnel or handling equipment) are obtained to label the scene object type; at the same time, the three-dimensional coordinate information of the scene object in three-dimensional space is extracted, including the center position and boundary range, to represent the specific spatial position of the scene object in the warehousing scene.

[0115] S5.2 Construct a graph node set by using the unique identifier as the node index, the category attribute as the node feature, and the three-dimensional spatial coordinate information as the node position vector.

[0116] Furthermore, for each scene object extracted from the structured scene data, the unique identifier of the scene object is used as the node index of the corresponding graph node in the interaction relationship graph to ensure that the graph node can uniquely correspond to each scene object; the category attribute of the scene object is used as the node feature of the graph node to represent the semantic type information of the scene object; the three-dimensional coordinate information of the scene object is used as the position vector of the graph node to reflect the specific position of the scene object in three-dimensional space; all scene objects in the structured scene data are processed in sequence, and all graph nodes and their node indexes, node features and position vectors are integrated to generate a complete set of graph nodes.

[0117] S5.3 Calculate the spatial proximity relationship between scene objects based on the node position vector, and determine the candidate set of interaction edges by combining the category attributes.

[0118] Furthermore, the node position vectors of the graph nodes are read sequentially from the graph node set, let the first node be... The node position vectors corresponding to each scene object are: , No. The node position vectors corresponding to each scene object are: The Euclidean distance between the two is calculated as a spatial proximity index, and the expression is:

[0119] ;

[0120] In the formula, It is a scene object With scene objects The Euclidean distance; It is a node position vector and The Euclidean norm between them, i.e., the linear distance in space; It is the first The X, Y, and Z coordinate components of a scene object in three-dimensional space; It is the first The X, Y, and Z coordinate components of a scene object in three-dimensional space;

[0121] If Euclidean distance Less than the preset neighbor threshold Then determine the scene object With scene objects Spatial proximity exists The expression is:

[0122] ;

[0123] Repeat the above calculation for all scene object pairs in the graph node set to generate a complete spatial proximity matrix. Matrix elements Represents scene object With scene objects Elements have spatial proximity relationships; an element value of 0 indicates that they are not adjacent.

[0124] The scene objects corresponding to the graph nodes are compared by 3D distance, overlapping volume, and relative position, and spatial proximity is determined by combining mutual occlusion or close proximity conditions to obtain graph node pairs that satisfy spatial proximity. The graph node pairs that satisfy spatial proximity are matched with category interaction rules, such as shelf-pallet, personnel-handling equipment, etc. Only graph node pairs that satisfy both spatial proximity and category interaction rules are retained as the candidate set of interaction edges. All graph node pairs in the graph node set are processed repeatedly to finally form a complete candidate set of interaction edges.

[0125] It should be noted that the proximity threshold is determined based on the spatial scale and interaction requirements of different types of scene objects in the warehousing scenario. Typically, it refers to the spatial characteristics of scene objects such as size, aisle width, or shelf height. By analyzing historical scene layouts and interaction events, a distance value is selected that can both cover potential interaction objects and avoid too many redundant edges. For example, the value is between 0.5 meters and 2 meters as the proximity threshold.

[0126] Category interaction rules are obtained by referencing known interaction patterns of different categories of scene objects in the warehousing scenario.

[0127] S5.4 Combine the interaction edges in the candidate set with the graph node set to generate an interaction relationship graph.

[0128] Furthermore, each interaction edge in the candidate set is examined to confirm that both ends of the interaction edge exist in the graph node set and that the graph node pair conditions of spatial proximity and category interaction rules are met simultaneously, thus determining it as a valid interaction edge. Feature information is added to each valid interaction edge, including spatial distance between nodes, category matching, and historical interaction frequency, and weights or importance indicators are assigned to the interaction edges based on the added features. After completing the feature construction of all interaction edges, the interaction edge set is filtered to remove interaction edges with low confidence or that do not meet the consistency requirements. Finally, the graph node set is combined with the filtered interaction edge set to output an interaction relationship graph containing complete node and interaction edge features.

[0129] S5.5 Map the spatial risk results to each graph node in the interaction relationship graph to obtain the risk weight corresponding to each scene object.

[0130] Furthermore, the system reads the set of graph nodes in the spatial risk results and interaction relationship graph, establishing a correspondence between the spatial risk results and the set of graph nodes; it traverses the set of graph nodes, obtains the node position vector of each graph node, and matches the node position vector with the spatial position in the spatial risk results; when a match is successful, the risk value in the spatial risk results is assigned to the corresponding graph node, forming the risk weight of the graph node; after completing the risk value assignment of all graph nodes, the risk weights can be normalized or weighted, and the complete result containing the risk weight corresponding to each scene object can be output.

[0131] S5.6. Based on risk weights and the interaction edge features between graph nodes, assess the risk transmission relationship between scene objects and generate interaction risk results.

[0132] Furthermore, the risk weights of graph nodes are normalized to obtain comparable normalized risk weight vectors, eliminating differences caused by different dimensions and numerical ranges. Elements such as spatial distance, category matching, historical interaction frequency, and edge feature importance are extracted from the interaction edge features. These elements are then synthesized into edge weights for each interaction edge according to a pre-defined fusion rule. The fusion rule can use the product of exponential decay of spatial distance and category matching priority, combined with a weighted adjustment based on historical interaction frequency, to reflect the true interaction impact. A weighted adjacency matrix is ​​constructed using the edge weights, and the matrix is ​​normalized row-wise or column-wise to ensure the numerical stability of the risk transmission process. Iterative propagation is implemented based on the normalized risk weight vector and the normalized weighted adjacency matrix: in each iteration, the current risk value of each graph node is allocated to adjacent graph nodes according to the corresponding edge weight in the normalized weighted adjacency matrix. The allocation result is then merged with the original normalized risk weights of the adjacent graph nodes through weighted summation or weighted average. This iteration is repeated until the risk values ​​converge. The risk value of each graph node after the iteration is taken as the interaction risk result.

[0133] It should be noted that the fusion rules were determined by statistically analyzing historical warehousing scenario risk event records and interaction behavior samples, combined with spatial distance attenuation patterns, risk transmission relationships between scenario object categories, and the impact of interaction frequency on risk accumulation.

[0134] S5.7. Based on the interaction risk results, mark the positions of the corresponding scene objects in three-dimensional space and output the three-dimensional position marks.

[0135] Furthermore, the three-dimensional coordinates of each scene object are extracted from the structured scene data, including the center position and boundary range. Based on the three-dimensional coordinates of the scene object, a position mark is generated for each scene object in three-dimensional space. The risk weight is associated with the position mark so that the position mark can reflect the risk level of the corresponding scene object. The three-dimensional position marks of all scene objects are summarized to form a complete three-dimensional position mark.

[0136] S6. Generate a 3D visualization based on the interactive risk results and 3D location markers, and mark dangerous areas and risk objects to form a warehouse risk visualization report.

[0137] S6.1. The interactive risk results are fused with the three-dimensional location markers to reconstruct the three-dimensional spatial layout of the warehousing scene and output a three-dimensional visualization.

[0138] Furthermore, the scene objects in the 3D location markers are mapped with visual attributes according to risk weights, such as reflecting the risk level through color, brightness, or size. All scene objects are arranged in spatial position in the 3D coordinate space, maintaining the original 3D layout and hierarchical relationship. The set of scene objects mapped with risk information is integrated into a complete 3D spatial model, forming a 3D visualization screen that can be directly presented.

[0139] S6.2 Highlight the dangerous areas in the 3D visualization and make the risky objects prominent, then output the annotated 3D visualization.

[0140] Furthermore, the 3D coordinate range and spatial location marker information of high-risk storage areas are extracted from the 3D visualization. Based on the risk weights in the interactive risk results, the high-risk storage areas are highlighted in the 3D visualization, for example, by using color deepening, transparency adjustment, or texture changes to present the risk level. High-risk scene objects are marked with significant marks, including adding borders, symbols, or labels to the 3D visualization to identify object categories and risk characteristics. All highlighted high-risk storage areas and significantly marked risk objects are integrated according to the original 3D spatial layout to form a marked 3D visualization.

[0141] S6.3. Accumulate risk weights and perform spatial interaction analysis on dangerous areas and risk objects in the 3D visualization. When the risk level reaches the alarm condition, trigger an early warning and generate a warehouse risk visualization report.

[0142] Furthermore, scene objects that spatially overlap with or are spatially adjacent to hazardous areas and risky objects are weighted and summed according to the edge weights in the spatial interaction relationship. The edge weights are calculated by multiplying the exponential decay of spatial distance and the category matching priority, combined with the weighted adjustment of historical interaction frequency. The cumulative risk weights of hazardous areas and risky objects are normalized to obtain comparable risk level indicators. When the risk level indicators of hazardous areas and risky objects reach the alarm conditions or the similarity with historical high-risk patterns exceeds the risk threshold (the risk threshold usually ranges from 0 to 1; for example, when the calculated result is greater than 0.7, it is judged as high risk, and when it is less than 0.3, it is judged as low risk), an early warning is triggered, and the three-dimensional location markers, cumulative risk weights, and salience visualization attributes of hazardous areas and risky objects are written together to generate a warehouse risk visualization report.

[0143] It should be noted that the risk threshold is obtained by statistically modeling accident samples and normal samples from historical monitoring data, and by combining the distribution patterns and classification boundaries of different risk characteristics to extract the critical value that can effectively distinguish between high-risk and safe states.

[0144] Alarm conditions refer to the trigger limits reached by risk indicators of dangerous areas or risky objects; alarm conditions are determined by statistical analysis and modeling of the differences between historical accident data and normal operation data.

[0145] This embodiment also provides a computer device applicable to the deep learning-based e-commerce warehouse security monitoring method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based e-commerce warehouse security monitoring method proposed in the above embodiment.

[0146] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0147] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the deep learning-based security monitoring method for e-commerce warehouse centers as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0148] In summary, this invention achieves accurate identification of high-risk areas in warehouses by: time-synchronizing and multi-layer interactively fusing high-definition video and laser point cloud data in the warehouse to generate a high-precision 3D point cloud representation; simultaneously, it abstracts structured scene objects into graph nodes and constructs an interaction relationship graph, quantifies the potential risks between multiple objects, and generates a 3D visualization, enabling intuitive labeling and automatic early warning of dangerous areas and risky objects, thereby significantly improving the automation, precision, and visualization level of warehouse safety monitoring.

[0149] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based security monitoring method for e-commerce warehouse centers, characterized in that: include, Collect high-definition video data and laser point cloud data from the warehouse, and process them synchronously through timestamps to form multi-source monitoring data; The convolutional neural network is used to perform semantic segmentation and spatial information extraction on multi-source monitoring data to obtain structured scene data. The specific steps are as follows. The warehouse high-definition video data and laser point cloud data in the multi-source monitoring data are spatially mapped and fused according to time, and divided into voxel grids to generate voxel multi-source data tensors. Voxelized multi-source data tensors are input into a convolutional neural network for semantic segmentation, and voxel-level semantic classification results are output. The voxel-level semantic classification results are subjected to 3D coordinate back mapping and scene object extraction. The scene objects are organized according to category and spatial hierarchy to generate structured scene data. Structured scene data and laser point cloud data are spatiotemporally aligned and fused to obtain a 3D point cloud representation. PointNet network is then used to perform point-level segmentation and high-risk area identification on the 3D point cloud representation to obtain high-risk areas in the warehouse. By combining high-risk storage areas with spatial location mapping tables, spatial risk results are generated. The scene objects in the structured scene data are abstracted into graph nodes, and an interaction relationship graph is established based on the three-dimensional spatial representation. The interaction relationship between scene objects is analyzed in combination with the spatial risk results, and interaction risk results and three-dimensional position markers are generated. Based on the interactive risk results and 3D location markers, a 3D visualization is generated, and hazardous areas and risk objects are marked to form a warehouse risk visualization report.

2. The deep learning-based security monitoring method for e-commerce warehouse centers as described in claim 1, characterized in that: The process of synchronizing data using timestamps to generate multi-source monitoring data involves the following steps: Align the timestamps of warehouse high-definition video data and laser point cloud data to generate time-aligned warehouse data; Perform video frame format unification and point cloud coordinate transformation on time-aligned warehouse data to generate formatted multi-source warehouse data; Formatted multi-source warehouse data is combined according to time series to form multi-source monitoring data.

3. The deep learning-based security monitoring method for e-commerce warehousing centers as described in claim 1, characterized in that: The specific steps for obtaining the 3D point cloud representation are as follows: The 3D position of each scene object in the structured scene data is compared with the laser point cloud data according to the timestamp, and the 3D position of the scene object is initially spatially registered to generate an aligned dataset. Each record in the aligned dataset is split into a scene object and a corresponding point cloud subset, and the semantic labels of the scene object and the local geometric features of the corresponding point cloud subset are extracted respectively. Semantic labels and local geometric features are used as two branches and input into a multi-layer interactive attention network for bidirectional information interaction and fusion to generate a fused feature set. Non-uniform density sampling and local surface reconstruction are performed on the fused feature set, and all sampled and reconstructed point cloud subsets are integrated to generate a 3D point cloud representation.

4. The deep learning-based security monitoring method for e-commerce warehouse centers as described in claim 1, characterized in that: The specific steps for identifying high-risk storage areas are as follows. The 3D point cloud representation is input into the PointNet network, and the spatial geometric features and local neighborhood features of each point cloud are extracted to obtain a set of point-level features. The point-level feature set is input into the classification layer for point-by-point discrimination, and the set of scene regions is generated by aggregation through point labels; By classifying the scene regions and combining their spatial location, volume, and interaction with neighboring objects, risk characteristic analysis is performed to identify a preliminary set of high-risk region candidates with potential hazardous attributes. The initial high-risk area candidate set is screened and confirmed through spatial consistency verification, threshold screening, and historical high-risk pattern matching, and the high-risk warehouse areas are output.

5. The deep learning-based security monitoring method for e-commerce warehousing centers as described in claim 4, characterized in that: The specific steps for combining high-risk storage areas with a spatial location mapping table to generate spatial risk results are as follows. The center coordinates and boundary range of the high-risk storage area are extracted from the 3D point cloud representation to generate the corresponding spatial location identifier; the spatial location identifier is matched with the spatial location mapping table to obtain the mapping position relationship of the high-risk storage area. Based on the mapping positional relationship, the spatial proximity relationship and relative position between each high-risk storage area and surrounding scene objects are analyzed to obtain the spatial interaction relationship of the high-risk storage area; By combining spatial interaction relationships, the distribution characteristics of high-risk areas in warehousing, and the interaction risks of surrounding scene objects, spatial risk results for warehousing scenarios are generated.

6. The deep learning-based security monitoring method for e-commerce warehousing centers as described in claim 5, characterized in that: The spatial location mapping table is obtained by dividing the space of the three-dimensional point cloud representation in the warehousing scene, aligning the three-dimensional coordinates of each scene object with the actual layout information of the warehousing area, and recording the area identifier, location range and object distance relationship in tabular form.

7. The deep learning-based security monitoring method for e-commerce warehousing centers as described in claim 1, characterized in that: The specific steps for abstracting scene objects from structured scene data into graph nodes and establishing an interaction relationship graph based on a three-dimensional spatial representation are as follows. Extract the unique identifier, category attribute, and three-dimensional spatial coordinate information of each scene object from the structured scene data; A graph node set is constructed by using unique identifiers as node indices, category attributes as node features, and 3D spatial coordinate information as node position vectors. The spatial proximity relationship between scene objects is calculated based on the node position vector, and the candidate set of interaction edges is determined by combining the category attribute; The interaction edges in the candidate set are combined with the graph node set to generate an interaction relationship graph.

8. The deep learning-based security monitoring method for e-commerce warehousing centers as described in claim 1, characterized in that: The specific steps for generating interactive risk results and 3D location markers are as follows. The spatial risk results are mapped to the nodes of the interaction graph to obtain the risk weight corresponding to each scene object; Based on risk weights and the interaction edge features between graph nodes, the risk transmission relationship between scene objects is evaluated, and the interaction risk results are generated. Based on the interaction risk results, mark the location of the corresponding scene object in three-dimensional space and output the three-dimensional location mark.

9. The deep learning-based security monitoring method for e-commerce warehouse centers as described in claim 1, characterized in that: The specific steps for generating a warehouse risk visualization report are as follows: By fusing the interactive risk results with the 3D location markers, the 3D spatial layout of the warehouse scene is reconstructed, and a 3D visualization is output. The system highlights dangerous areas in the 3D visualization and marks risky objects to make them more prominent, then outputs the marked 3D visualization. The system performs risk weight accumulation and spatial interaction analysis on dangerous areas and risky objects in the 3D visualization. When the risk level reaches the alarm condition, it triggers an early warning and generates a warehouse risk visualization report.

Images