A logistics warehousing monitoring system based on artificial intelligence

Abstract

The present invention relates to the field of warehouse monitoring technology, specifically a logistics warehouse monitoring system based on artificial intelligence, including a dynamic update unit, a behavior detection unit, a regional evaluation unit, a regional early warning unit, an abnormal prediction unit and a monitoring interface unit, and also includes: a spatial modeling unit, the spatial modeling unit is used to divide the target warehouse into regions based on the spatial connectivity of the target warehouse, so as to divide the target warehouse into multiple sub-regions, construct corresponding multiple digital grids based on the multiple sub-regions, and construct spatial topological relationship primitives based on the multiple digital grids. Through video monitoring and dynamic target detection, the system of the present invention can monitor the dynamics of personnel and goods in the storage area in real time, improve the timeliness and accuracy of management, and can identify and classify the behavior of staff through the behavior detection unit, help identify potential abnormal behavior, and improve work efficiency and safety.

Description

Technical Field

[0001] The present invention relates to the technical field of warehousing monitoring, and in particular to a logistics warehousing monitoring system based on artificial intelligence. Background Art

[0002] Artificial intelligence is a branch of computer science that aims to create machines and systems that can simulate human intelligence. Logistics warehousing refers to the storage, management and distribution process of goods or goods in logistics management.

[0003] Traditional monitoring systems often rely on manual inspections or scheduled checks, resulting in delayed information updates and difficulty in detecting abnormal situations in a timely manner. Traditional systems are usually unable to perform behavior recognition and analysis and can only provide simple video surveillance, unable to deeply understand and analyze staff behavior. Traditional systems usually do not consider spatial connectivity and lack dynamic evaluation and optimization of storage areas, which may lead to insufficient space utilization. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a logistics warehousing monitoring system based on artificial intelligence.

[0005] The technical solution adopted to solve the above technical problems is: an artificial intelligence-based logistics warehousing monitoring system, including a dynamic update unit, a behavior detection unit, a regional evaluation unit, a regional early warning unit, an abnormality prediction unit and a monitoring interface unit, and also includes:

[0006] A spatial modeling unit, the spatial modeling unit is used to divide the target warehouse into multiple sub-areas based on the spatial connectivity of the target warehouse, and to construct a corresponding multiple digital grids based on the multiple sub-areas, and to construct a spatial topological relationship graph based on the multiple digital grids, wherein the spatial topological relationship graph includes multiple nodes and multiple connecting edges, the nodes represent the digital grids, and the connecting edges represent the spatial connectivity between the sub-areas, and the spatial topological relationship graph is transmitted to the dynamic update unit;

[0007] A video monitoring unit, wherein the video monitoring unit is used to set a video acquisition device in the plurality of sub-areas to acquire video of the sub-areas to obtain scene videos, and transmit the scene videos to the target detection unit;

[0008] A target detection unit, which receives the scene video transmitted by the video monitoring unit, and performs target detection on the video frame of the scene video based on the trained target detection model to obtain a target bounding box of the video frame, wherein the target includes a human body of a staff member and a cargo, and transmits the target bounding box of the video frame to a target tracking unit;

[0009] A target tracking unit receives the target bounding box of the video frame transmitted by the target detection unit, and tracks the target in the scene video based on the target bounding box of the video frame and a preset template image to obtain a dynamic target bounding box in the scene video, and transmits the dynamic target bounding box in the scene video to the behavior detection unit.

[0010] Preferably, the behavior detection unit receives the dynamic target bounding box in the scene video transmitted by the target tracking unit, and classifies the dynamic target bounding box based on a trained behavior detection model to obtain the behavior label corresponding to the dynamic target bounding box, that is, obtains the behavior label sequence corresponding to the staff, and transmits the behavior label sequence corresponding to the staff to the regional evaluation unit.

[0011] Preferably, the area evaluation unit receives the behavior label sequence corresponding to the staff transmitted by the behavior detection unit, and counts the behavior label sequences corresponding to all the staff in the sub-area to obtain a behavior label quantity sequence in the sub-area, and evaluates the sub-area based on the behavior label quantity sequence in the sub-area and the behavior label sequences corresponding to all the staff in the sub-area to obtain an abnormality coefficient of the sub-area, and transmits the abnormality coefficient of the sub-area to the dynamic update unit.

[0012] Preferably, the dynamic update unit receives the spatial topological relationship diagram transmitted by the spatial modeling unit and the anomaly coefficient of the sub-region transmitted by the regional evaluation unit, and assigns weights of multiple nodes in the spatial topological relationship diagram based on the anomaly coefficient of the sub-region to obtain a weighted spatial topological relationship diagram, and transmits the weighted spatial topological relationship diagram to the regional early warning unit, the anomaly prediction unit and the monitoring interface unit.

[0013] Preferably, the regional warning unit receives the weighted spatial topological relationship graph transmitted by the dynamic update unit, and compares the weighted spatial topological relationship graph with a preset standard weighted spatial topological relationship graph to obtain difference nodes, obtains corresponding sub-regions based on the difference nodes, and performs regional warning on the sub-regions corresponding to the difference nodes based on the sound and light alarm equipment.

[0014] Preferably, the anomaly prediction unit receives the weighted spatial topological relationship graph transmitted by the dynamic update unit, and performs anomaly prediction on the weighted spatial topological relationship graph based on a trained anomaly prediction model to obtain the weighted spatial topological relationship graph at the next preset moment, and transmits the weighted spatial topological relationship graph at the next preset moment to the monitoring interface unit.

[0015] Preferably, the monitoring interface unit receives the weighted spatial topological relationship diagram transmitted by the dynamic update unit and the weighted spatial topological relationship diagram at the next preset moment transmitted by the abnormality prediction unit, and dynamically displays the weighted spatial topological relationship diagram and the weighted spatial topological relationship diagram at the next preset moment in the intelligent monitoring device.

[0016] Preferably, tracking the target in the scene video based on the target bounding box of the video frame and a preset template image to obtain a dynamic target bounding box in the scene video includes:

[0017] Cropping an image block based on the center position of the target bounding box of the aforementioned video frame, and using the image block as a detection image;

[0018] Based on the trained twin network, feature extraction is performed on the template image and the detection image to obtain a template feature map corresponding to the template image and a detection feature map corresponding to the detection image;

[0019] Performing a convolution operation on the template feature map and the detection feature map to obtain a correlation map, wherein the correlation map includes a classification map and a regression map, and smoothing the correlation map based on a cosine window and a scale change penalty;

[0020] Selecting corresponding regression position information in the regression map based on the position with the highest score in the classification map to obtain a predicted target bounding box of the current video frame;

[0021] estimating the adaptive position confidence of the correlation graph, and comparing the adaptive position confidence of the correlation graph with a preset confidence threshold;

[0022] If the adaptive position confidence of the correlation graph is greater than a preset confidence threshold, the template image is updated based on the predicted target bounding box to obtain a new template image; otherwise, the template image is not updated;

[0023] Repeat the above operation to obtain the dynamic target bounding box in the scene video.

[0024] Preferably, the behavior detection model includes a feature extraction module, a feature fusion module and a feature classification module. The feature extraction module is used to perform feature extraction on the dynamic target bounding box to obtain an image feature set of the dynamic target bounding box. The feature fusion module is used to perform feature fusion on the image feature set to obtain semantic features of the dynamic target bounding box. The feature classification module is used to classify the semantic features to obtain the behavior label corresponding to the staff member.

[0025] Preferably, performing feature extraction on the dynamic object bounding box to obtain an image feature set of the dynamic object bounding box includes:

[0026] Based on the trained first feature extraction model, feature extraction is performed on the human body, object and merged area in the dynamic target bounding box to obtain human body features, object features and merged area features in the dynamic target bounding box;

[0027] Extracting key points of a human body in the dynamic target bounding box based on a trained second feature extraction model to obtain key point features of the human body in the dynamic target bounding box;

[0028] Constructing a character spatial structure graph based on human features, object features, merged region features, and human key point features, and extracting a feature vector of the character spatial structure graph based on a full connection operation to obtain spatial structure features within the dynamic target bounding box;

[0029] The human body features, the human body key point features and the object features are combined to obtain a person pair feature set, and the person pair feature set, the merged region features and the spatial structure features are combined to obtain an image feature set of the dynamic target bounding box.

[0030] Preferably, performing feature fusion on the image feature set to obtain the semantic features of the dynamic target bounding box includes:

[0031] Constructing a first scene graph based on key point features of the human body;

[0032] Performing feature optimization on the first scene graph based on a graph convolution operation, reducing the number of nodes in the first scene graph based on a graph attention pooling operation, and adding the spatial structure features and the human body features to construct a second scene graph;

[0033] Performing feature optimization on the second scene graph based on a graph convolution operation, and adding the object features and the merged region features to construct a third scene graph;

[0034] Based on the graph attention pooling operation, the number of nodes of the third scene graph is reduced, and the nodes of the third scene graph are fused based on the attention weights of the nodes to obtain the semantic features of the dynamic target bounding box.

[0035] Preferably, the anomaly prediction model adopts a graph neural network, and the anomaly prediction model consists of stacked spatiotemporal layers and an output layer. The spatiotemporal layers include a graph convolutional network and a gated temporal convolutional network. The stacked spatiotemporal layers are used to process spatial dependencies at different time levels. The output of each spatiotemporal layer is summed through a skip connection, and the output layer is used to process through two ReLU functions and a linear layer to obtain a weighted spatial topological relationship graph at the next preset moment.

[0036] Preferably, the loss function of the anomaly prediction model is as follows:

[0037] ;

[0038] in, represents the loss function of the anomaly prediction model, represents the time step of the preset time, Indicates the number of nodes, represents the feature dimension of the node, Indicates the default standard output, Indicates the actual output

[0039] The beneficial effects of the present invention are as follows: (1) Through video monitoring and dynamic target detection, the system of the present invention can monitor the dynamics of personnel and goods in the storage area in real time, improve the timeliness and accuracy of management, and can identify and classify the behavior of staff through the behavior detection unit, help identify potential abnormal behavior, and improve work efficiency and safety; (2) The present invention uses the regional evaluation unit to perform statistical analysis on the behavior labels, thereby evaluating the status of each sub-area, and timely issues an alarm through the regional early warning unit to reduce the possibility of accidents, and the dynamic update unit adjusts the weight of the spatial topological relationship diagram according to the abnormality coefficient, which helps to optimize the storage layout and improve space utilization; (3) The present invention uses the abnormality prediction unit to predict possible abnormal situations in the future based on historical data and models, and take preventive measures in advance, and the monitoring interface unit dynamically displays the weighted spatial topological relationship diagram, so that managers can intuitively understand the storage status and facilitate decision-making and adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of the system architecture of an overall system in an embodiment of the present invention.

[0041] Figure numerals: 1. spatial modeling unit; 2. dynamic update unit; 3. video monitoring unit; 4. target detection unit; 5. target tracking unit; 6. behavior detection unit; 7. regional evaluation unit; 8. regional warning unit; 9. abnormal prediction unit; 10. monitoring interface unit. DETAILED DESCRIPTION

[0042] Embodiment 1, as Figure 1 As shown, the logistics warehousing monitoring system based on artificial intelligence proposed by the present invention includes a dynamic update unit 2, a behavior detection unit 6, a regional evaluation unit 7, a regional early warning unit 8, an abnormality prediction unit 9 and a monitoring interface unit 10, and also includes:

[0043] The spatial modeling unit 1 is used to divide the target warehouse into multiple sub-areas based on the spatial connectivity of the target warehouse, and to construct a corresponding multiple digital grids based on the multiple sub-areas, and to construct a spatial topological relationship graph based on the multiple digital grids, wherein the spatial topological relationship graph includes multiple nodes and multiple connecting edges, the nodes represent the digital grids, and the connecting edges represent the spatial connectivity between the sub-areas, and the spatial topological relationship graph is transmitted to the dynamic updating unit 2;

[0044] The video monitoring unit 3 is used to set a video acquisition device in a plurality of sub-areas to acquire video of the sub-areas to obtain scene videos, and transmit the scene videos to the target detection unit 4;

[0045] The target detection unit 4 receives the scene video transmitted by the video monitoring unit 3, and performs target detection on the video frame of the scene video based on the trained target detection model to obtain a target bounding box of the video frame, wherein the target includes the human body of the staff and the cargo, and transmits the target bounding box of the video frame to the target tracking unit 5;

[0046] The target tracking unit 5 receives the target bounding box of the video frame transmitted by the target detection unit 4, and tracks the target in the scene video based on the target bounding box of the video frame and a preset template image to obtain a dynamic target bounding box in the scene video, and transmits the dynamic target bounding box in the scene video to the behavior detection unit 6.

[0047] In the present invention, the region partition divides the target space (such as a warehouse) into multiple sub-regions according to certain standards (such as spatial connectivity) for easy management and monitoring; the digital grid divides the space into several small grids, each of which can represent a specific area, which is convenient for digital processing and analysis; the spatial topology relationship diagram is a graphical structure that represents the connection relationship between different areas. The nodes represent the digital grids, and the edges represent the connectivity between the sub-regions; the target detection models include but are not limited to R-CNN, Fast R-CNN, SSD and EfficientDet.

[0048] In an optional embodiment, the behavior detection unit 6 receives the dynamic target bounding box in the scene video transmitted by the target tracking unit 5, and classifies the dynamic target bounding box based on the trained behavior detection model to obtain the behavior label corresponding to the dynamic target bounding box, that is, obtains the behavior label sequence corresponding to the staff, and transmits the behavior label sequence corresponding to the staff to the regional evaluation unit 7.

[0049] In an optional embodiment, the regional evaluation unit 7 receives the behavior label sequence corresponding to the staff transmitted by the behavior detection unit 6, and counts the behavior label sequences corresponding to all the staff in the sub-region to obtain a behavior label quantity sequence in the sub-region, and evaluates the sub-region based on the behavior label quantity sequence in the sub-region and the behavior label sequences corresponding to all the staff in the sub-region to obtain an abnormality coefficient of the sub-region, and transmits the abnormality coefficient of the sub-region to the dynamic update unit 2.

[0050] It should be noted that the behavior label sequence marks the behavior of the staff within a certain period of time to form a behavior record sequence, and each label represents a specific behavior; the behavior label quantity sequence refers to the number of times each behavior label appears within a certain period of time to form a quantity record. This sequence reflects the frequency and distribution of the behavior of the staff in the sub-area; regional evaluation refers to the analysis of the behavior labels and quantity sequences in the sub-area to determine indicators such as the working status, efficiency or safety of the sub-area; the abnormality coefficient is an indicator calculated by comparing the behavior label quantity sequence under normal circumstances with the actual observed sequence. It is used to evaluate whether there is any abnormality in the behavior within a sub-area. For example, when the number of a certain behavior far exceeds expectations, it may mean that there are safety hazards or management problems in the area.

[0051] In an optional embodiment, the dynamic update unit 2 receives the spatial topological relationship diagram transmitted by the spatial modeling unit 1 and the anomaly coefficient of the sub-region transmitted by the regional evaluation unit 7, and assigns weights to multiple nodes in the spatial topological relationship diagram based on the anomaly coefficient of the sub-region to obtain a weighted spatial topological relationship diagram, and transmits the weighted spatial topological relationship diagram to the regional early warning unit 8, the anomaly prediction unit 9 and the monitoring interface unit 10.

[0052] In an optional embodiment, the regional warning unit 8 receives the weighted spatial topological relationship graph transmitted by the dynamic update unit 2, and compares the weighted spatial topological relationship graph with a preset standard weighted spatial topological relationship graph to obtain difference nodes, obtains corresponding sub-regions based on the difference nodes, and performs regional warning on the sub-regions corresponding to the difference nodes based on the sound and light alarm equipment.

[0053] It should be noted that the preset standard weighted spatial topological relationship diagram is a benchmark diagram used to represent the spatial topological relationship under ideal conditions. This diagram may be based on historical data, design specifications or best practices and serve as a standard for subsequent comparisons; comparison refers to comparing the current weighted spatial topological relationship diagram with the standard diagram to identify the differences between them; difference nodes are nodes discovered during the comparison process that are significantly different from the standard diagram. These nodes may represent abnormal situations, changes or potential problems; the sound and light alarm device is an alarm system that uses sound and light signals to alert staff or managers to potential dangers or abnormal situations; regional warning refers to the activation of the sound and light alarm device for warning based on the identified difference nodes and corresponding sub-areas, so as to promptly deal with possible safety hazards or other problems.

[0054] In an optional embodiment, the anomaly prediction unit 9 receives the weighted spatial topological relationship graph transmitted by the dynamic update unit 2, and performs anomaly prediction on the weighted spatial topological relationship graph based on a trained anomaly prediction model to obtain the weighted spatial topological relationship graph at the next preset moment, and transmits the weighted spatial topological relationship graph at the next preset moment to the monitoring interface unit 10.

[0055] In an optional embodiment, the monitoring interface unit 10 receives the weighted spatial topological relationship diagram transmitted by the dynamic update unit 2 and the weighted spatial topological relationship diagram at the next preset moment transmitted by the abnormality prediction unit 9, and dynamically displays the weighted spatial topological relationship diagram and the weighted spatial topological relationship diagram at the next preset moment in the intelligent monitoring device.

[0056] It should be noted that dynamic display refers to the real-time display of changes in the weighted spatial topological relationship diagram in the monitoring device. This display can reflect the changes in spatial relationships and weights over time, helping managers to obtain important information in a timely manner.

[0057] Embodiment 2, a logistics warehousing monitoring system based on artificial intelligence proposed by the present invention, compared with embodiment 1, this embodiment further includes: tracking the target in the scene video based on the target bounding box of the video frame and the preset template image to obtain the dynamic target bounding box in the scene video, including:

[0058] Cropping an image block based on the center position of the target bounding box of the current video frame, and using the image block as a detection image;

[0059] Based on the trained twin network, feature extraction is performed on the template image and the detection image to obtain a template feature map corresponding to the template image and a detection feature map corresponding to the detection image;

[0060] The template feature map and the detection feature map are convolved to obtain a correlation map, wherein the correlation map includes a classification map and a regression map, and the correlation map is smoothed based on a cosine window and a scale change penalty;

[0061] Based on the position with the highest score in the classification map, the corresponding regression position information is selected in the regression map to obtain the predicted target bounding box of the current video frame;

[0062] estimating the adaptive position confidence of the correlation graph, and comparing the adaptive position confidence of the correlation graph with a preset confidence threshold;

[0063] If the adaptive position confidence of the correlation graph is greater than a preset confidence threshold, the template image is updated based on the predicted target bounding box to obtain a new template image, otherwise, the template image is not updated;

[0064] Repeat the above steps to obtain the dynamic target bounding box in the scene video.

[0065] In this embodiment, a video frame refers to a single static picture of a continuous image in a video, and a video is composed of a series of frames; a twin network is a neural network architecture consisting of two identical sub-networks, and is generally used to compare the similarity of two inputs or extract features; a convolution operation is a mathematical operation used to extract features, and processes an image through a sliding window, and is commonly used in image processing and deep learning; a correlation graph refers to the result obtained by a convolution operation, and is used to represent the similarity between a template and a detected image; a classification graph refers to the classification information contained in the correlation graph, and represents the probability of a target category at each position; a regression graph refers to the regression information contained in the correlation graph, and is used to output the specific bounding box position of the target; a cosine window is a smoothing function used to reduce the edge effect of a convolution calculation and improve the smoothness of a correlation graph; an adaptive position confidence represents the probability of a target existing at a certain position, and is dynamically calculated and adjusted; a template image is generally an image used as a reference in a target detection or tracking task, and this image can be an example of a specific object, and features are extracted from it and compared with images captured in real time. In this way, targets similar to the template image in a video stream can be identified, and their positions and states can be predicted.

[0066] In an optional embodiment, the behavior detection model includes a feature extraction module, a feature fusion module and a feature classification module. The feature extraction module is used to perform feature extraction on the dynamic target bounding box to obtain an image feature set of the dynamic target bounding box. The feature fusion module is used to perform feature fusion on the image feature set to obtain semantic features of the dynamic target bounding box. The feature classification module is used to classify the semantic features to obtain the behavior label corresponding to the staff member.

[0067] In an optional embodiment, feature extraction is performed on the dynamic object bounding box to obtain an image feature set of the dynamic object bounding box, including:

[0068] Based on the trained first feature extraction model, feature extraction is performed on the human body, object and merged region within the dynamic target bounding box to obtain human body features, object features and merged region features within the dynamic target bounding box;

[0069] Extracting key points of the human body in the dynamic target bounding box based on the trained second feature extraction model to obtain key point features of the human body in the dynamic target bounding box;

[0070] A character spatial structure graph is constructed based on human features, object features, merged region features, and human key point features. The feature vector of the character spatial structure graph is extracted based on a full connection operation to obtain the spatial structure features within the dynamic target bounding box.

[0071] The human body features, human body key point features and object features are combined to obtain a person pair feature set, and the person pair feature set, merged region features and spatial structure features are combined to obtain an image feature set of a dynamic target bounding box.

[0072] It should be noted that the feature extraction model is a machine learning model used to extract useful information from input data, usually a deep learning network; the dynamic target bounding box is a rectangular box drawn around a moving object (such as a person or object) in a video stream, which is used for positioning and tracking; the merged region feature represents the combined features of multiple targets (such as human bodies and objects) in space; the human key points refer to important joints in the human body structure, such as shoulders, elbows, and knees; the character spatial structure diagram represents a graphical structure of the relationship between the human body and its surrounding environment, reflecting the spatial layout; the fully connected operation is a way of connecting all nodes to each other in a neural network, usually used to integrate feature information; the character pair feature set refers to a feature set generated by combining different features (such as human body and object features).

[0073] In an optional embodiment, feature fusion is performed on the image feature set to obtain semantic features of the dynamic target bounding box, including:

[0074] Constructing a first scene graph based on key point features of the human body;

[0075] The first scene graph is optimized based on the graph convolution operation, the number of nodes in the first scene graph is reduced based on the graph attention pooling operation, and the spatial structure features and human body features are added to construct the second scene graph;

[0076] Based on the graph convolution operation, the second scene graph is feature optimized, and object features and merged region features are added to construct a third scene graph;

[0077] Based on the graph attention pooling operation, the number of nodes in the third scene graph is reduced, and the nodes of the third scene graph are fused based on the attention weights of the nodes to obtain the semantic features of the dynamic target bounding box.

[0078] It should be noted that the scene graph is a graph structure in which nodes represent objects or key points and edges represent the relationship between them, which is used to represent and analyze the semantic information of the scene; the graph convolution operation is a method of extracting features on the graph structure through convolution operations in graph neural networks, which can consider the neighbor information of the nodes; the graph attention pooling operation uses the attention mechanism to weight the nodes in the graph, thereby selecting important nodes and reducing the number of nodes; the spatial structure feature describes the relative position and layout information of objects and key points in space, and is usually used to understand the geometric structure of the scene; the node attention weight is the weight assigned to each node in the graph attention mechanism, reflecting the importance of the node in the feature fusion process.

[0079] In an optional embodiment, the anomaly prediction model adopts a graph neural network, which consists of stacked spatiotemporal layers and output layers. The spatiotemporal layers include a graph convolutional network and a gated temporal convolutional network. The stacked spatiotemporal layers are used to process spatial dependencies at different time levels. The output of each spatiotemporal layer is summed through skip connections, and the output layer is used to process through two ReLU functions and a linear layer to obtain a weighted spatial topological relationship graph at the next preset moment.

[0080] It should be noted that a graph neural network (GNN) refers to a neural network that processes graph-structured data, can learn the relationships and features between nodes, and is suitable for complex data structures; a spatiotemporal layer refers to a layer that combines temporal and spatial features, and can capture the spatial dependencies of data over time; a gated temporal convolutional network (GTCN) refers to a convolutional network combined with a gating mechanism, which is used to process time series data and can effectively capture temporal dependencies; a skip connection refers to a connection mechanism that promotes feature transmission and prevents information loss by directly adding the output of the previous layer to the output of the subsequent layer; the output layer refers to the final layer of the model, which is responsible for processing the features of the previous layer into the desired output form; a linear layer refers to a simple neural network layer that performs linear transformations and is used to map input features to output features.

[0081] In an optional embodiment, the loss function of the abnormal prediction model is as follows:

[0082] ;

[0083] in, represents the loss function of the anomaly prediction model, represents the time step of the preset time, Indicates the number of nodes, represents the feature dimension of the node, Indicates the default standard output, Indicates the actual output.

[0084] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A logistics warehousing monitoring system based on artificial intelligence, comprising a dynamic update unit (2), a behavior detection unit (6), a regional evaluation unit (7), a regional early warning unit (8), an abnormality prediction unit (9) and a monitoring interface unit (10), characterized in that: A spatial modeling unit (1), the spatial modeling unit (1) is used to divide the target warehouse into multiple sub-areas based on the spatial connectivity of the target warehouse, and to construct a corresponding multiple digital grids based on the multiple sub-areas, and to construct a spatial topological relationship graph based on the multiple digital grids, wherein the spatial topological relationship graph includes multiple nodes and multiple connecting edges, the nodes represent the digital grids, and the connecting edges represent the spatial connectivity between the sub-areas, and the spatial topological relationship graph is transmitted to the dynamic updating unit (2); A video monitoring unit (3), the video monitoring unit (3) being used to set a video acquisition device in the plurality of sub-areas to acquire video of the sub-areas to obtain scene videos, and transmit the scene videos to the target detection unit (4); A target detection unit (4), the target detection unit (4) receives the scene video transmitted by the video monitoring unit (3), and performs target detection on the video frame of the scene video based on a trained target detection model to obtain a target bounding box of the video frame, wherein the target includes a human body of a worker and cargo, and transmits the target bounding box of the video frame to a target tracking unit (5); A target tracking unit (5), the target tracking unit (5) receives the target bounding box of the video frame transmitted by the target detection unit (4), and tracks the target in the scene video based on the target bounding box of the video frame and a preset template image to obtain a dynamic target bounding box in the scene video, and transmits the dynamic target bounding box in the scene video to the behavior detection unit (6); The behavior detection unit (6) receives the dynamic target bounding box in the scene video transmitted by the target tracking unit (5), and classifies the dynamic target bounding box based on the trained behavior detection model to obtain the behavior label corresponding to the dynamic target bounding box, that is, to obtain the behavior label sequence corresponding to the staff member, and transmits the behavior label sequence corresponding to the staff member to the regional evaluation unit (7); The area evaluation unit (7) receives the behavior label sequence corresponding to the staff member transmitted by the behavior detection unit (6), and counts the behavior label sequences corresponding to all the staff members in the sub-area to obtain a behavior label quantity sequence in the sub-area, evaluates the sub-area based on the behavior label quantity sequence in the sub-area and the behavior label sequences corresponding to all the staff members in the sub-area to obtain an abnormality coefficient of the sub-area, and transmits the abnormality coefficient of the sub-area to the dynamic update unit (2); The dynamic updating unit (2) receives the spatial topological relationship diagram transmitted by the spatial modeling unit (1) and the abnormal coefficient of the sub-region transmitted by the regional evaluation unit (7), and assigns weights to multiple nodes in the spatial topological relationship diagram based on the abnormal coefficient of the sub-region to obtain a weighted spatial topological relationship diagram, and transmits the weighted spatial topological relationship diagram to the regional early warning unit (8), the abnormal prediction unit (9) and the monitoring interface unit (10); The regional warning unit (8) receives the weighted spatial topology relationship diagram transmitted by the dynamic update unit (2), and compares the weighted spatial topology relationship diagram with a preset standard weighted spatial topology relationship diagram to obtain a difference node, obtains a corresponding sub-region based on the difference node, and performs a regional warning on the sub-region corresponding to the difference node based on the sound and light alarm device; The abnormality prediction unit (9) receives the weighted spatial topological relationship graph transmitted by the dynamic update unit (2), and performs abnormality prediction on the weighted spatial topological relationship graph based on the trained abnormality prediction model to obtain the weighted spatial topological relationship graph at the next preset moment, and transmits the weighted spatial topological relationship graph at the next preset moment to the monitoring interface unit (10); The monitoring interface unit (10) receives the weighted spatial topology relationship diagram transmitted by the dynamic update unit (2) and the weighted spatial topology relationship diagram at the next preset moment transmitted by the abnormality prediction unit (9), and dynamically displays the weighted spatial topology relationship diagram and the weighted spatial topology relationship diagram at the next preset moment in the intelligent monitoring device.

2. According to claim 1, a logistics warehousing monitoring system based on artificial intelligence is characterized in that: Tracking a target in the scene video based on a target bounding box of the video frame and a preset template image to obtain a dynamic target bounding box in the scene video includes: Cropping an image block based on the center position of the target bounding box of the aforementioned video frame, and using the image block as a detection image; Based on the trained twin network, feature extraction is performed on the template image and the detection image to obtain a template feature map corresponding to the template image and a detection feature map corresponding to the detection image; Performing a convolution operation on the template feature map and the detection feature map to obtain a correlation map, wherein the correlation map includes a classification map and a regression map, and smoothing the correlation map based on a cosine window and a scale change penalty; Selecting corresponding regression position information in the regression map based on the position with the highest score in the classification map to obtain a predicted target bounding box of the current video frame; estimating the adaptive position confidence of the correlation graph, and comparing the adaptive position confidence of the correlation graph with a preset confidence threshold; If the adaptive position confidence of the correlation graph is greater than a preset confidence threshold, the template image is updated based on the predicted target bounding box to obtain a new template image; otherwise, the template image is not updated; Repeat the above operation to obtain the dynamic target bounding box in the scene video.

3. According to claim 2, a logistics warehousing monitoring system based on artificial intelligence is characterized in that: The behavior detection model includes a feature extraction module, a feature fusion module and a feature classification module. The feature extraction module is used to perform feature extraction on the dynamic target bounding box to obtain an image feature set of the dynamic target bounding box. The feature fusion module is used to perform feature fusion on the image feature set to obtain semantic features of the dynamic target bounding box. The feature classification module is used to classify the semantic features to obtain a behavior label corresponding to the staff member.

4. The artificial intelligence-based logistics warehousing monitoring system according to claim 3 is characterized in that: Performing feature extraction on the dynamic object bounding box to obtain an image feature set of the dynamic object bounding box includes: Based on the trained first feature extraction model, feature extraction is performed on the human body, object and merged area in the dynamic target bounding box to obtain human body features, object features and merged area features in the dynamic target bounding box; Extracting key points of a human body in the dynamic target bounding box based on a trained second feature extraction model to obtain key point features of the human body in the dynamic target bounding box; Constructing a character spatial structure graph based on human features, object features, merged region features, and human key point features, and extracting a feature vector of the character spatial structure graph based on a full connection operation to obtain spatial structure features within the dynamic target bounding box; The human body features, the human body key point features and the object features are combined to obtain a person pair feature set, and the person pair feature set, the merged region features and the spatial structure features are combined to obtain an image feature set of the dynamic target bounding box.

5. The artificial intelligence-based logistics warehousing monitoring system according to claim 4 is characterized in that: Performing feature fusion on the image feature set to obtain semantic features of the dynamic target bounding box includes: Constructing a first scene graph based on key point features of the human body; Performing feature optimization on the first scene graph based on a graph convolution operation, reducing the number of nodes in the first scene graph based on a graph attention pooling operation, and adding the spatial structure features and the human body features to construct a second scene graph; Performing feature optimization on the second scene graph based on a graph convolution operation, and adding the object features and the merged region features to construct a third scene graph; Based on the graph attention pooling operation, the number of nodes of the third scene graph is reduced, and the nodes of the third scene graph are fused based on the attention weights of the nodes to obtain the semantic features of the dynamic target bounding box.

6. The artificial intelligence-based logistics warehousing monitoring system according to claim 5 is characterized in that: The anomaly prediction model adopts a graph neural network, which consists of stacked spatiotemporal layers and an output layer. The spatiotemporal layers include a graph convolutional network and a gated temporal convolutional network. The stacked spatiotemporal layers are used to process spatial dependencies at different time levels. The output of each spatiotemporal layer is summed through a skip connection, and the output layer is used to process through two ReLU functions and a linear layer to obtain a weighted spatial topological relationship graph at the next preset moment.

Images