Method for identifying and early warning unsafe behaviors of offshore petroleum operation

Abstract

The invention discloses an offshore oil operation unsafe behavior identification and early warning method, which comprises the following steps: determining the installation position and angle of camera equipment, and making an unsafe behavior data set; the scene of the data set is analyzed and labeled, a detection network framework needed to be used by each behavior is judged, and training and edge end deployment are completed; cutting and quantifying the model operator; linkage of the camera devices is realized; and video information acquired by the camera shooting equipment is displayed on a software end in real time. According to the method, hardware resources are reasonably utilized, different network frameworks and recognition algorithms are used according to specific scenes, and diversity and accuracy of detection effects are achieved. On limited equipment resources of an offshore platform, the video memory occupation and the calculation complexity of the model are reduced, and the calculation speed of an algorithm program is improved.

Description

Technical Field

[0001] This invention belongs to the field of offshore oil operation technology, and in particular relates to a method for identifying and warning of unsafe behaviors in offshore oil operations. Background Technology

[0002] Offshore oil operations are characterized by high risk, high investment, and high technical requirements. Due to the harsh working environment, complex construction methods, and confined space, numerous complex and diverse risks lurk during production. Management loopholes can easily lead to safety accidents, damaging equipment and personnel on the work platform. Effectively controlling these core risks has become a critical issue that urgently needs to be addressed in on-site safety management.

[0003] In existing technologies, video acquisition equipment is generally used to collect on-site conditions for production operations. The collected signals are then processed and analyzed to identify abnormal behaviors and conditions, triggering an alarm. However, existing models heavily rely on hardware resources, resulting in high resource consumption, slow processing speed, and poor accuracy and real-time performance in detecting unsafe behaviors during operations, which cannot meet the actual working needs of marine platforms. Summary of the Invention

[0004] The problem this invention aims to solve is to provide a method for identifying and warning of unsafe behaviors in offshore oil operations. This method makes reasonable use of hardware resources, employing different network frameworks and recognition algorithms according to specific scenarios to achieve diverse and accurate detection results. It reduces the model's memory usage and computational complexity on the limited equipment resources of offshore platforms, thereby improving the algorithm's computational speed.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for identifying and warning of unsafe behaviors in offshore oil operations, comprising the following steps: S1: Analyze historical statistical data along with actual scenarios to determine the installation location and angle of the camera equipment. Combine historical unsafe behavior image data to simulate unsafe behavior on-site, collect video, and create a dataset. S2: Analyze and label the scenarios in the dataset, determine the detection network framework to be used for each behavior, and complete the training and edge deployment; S3: Prune and quantize the model operators to reduce the model's memory usage and computational complexity; S4: For multiple detection scenarios, various network architectures are combined and multiple network models are integrated to achieve the linkage of the various camera devices; S5: Display the video information collected by the camera device in real time on the software. The software integrates a visualization detection algorithm module, a historical behavior storage and operation module, and an early warning alarm module. At the same time, alarm devices are set for each of the camera devices.

[0006] Furthermore, S1 includes the following steps: S11: Conduct on-site surveys of the offshore platform environment, statistically record scenarios where unsafe behaviors may occur, establish a normal distribution function for the types of unsafe behaviors that may occur in each scenario and the number of times unsafe behaviors have occurred in the past in the scenario, and install the camera equipment in scenarios where the values ​​are distributed in (μ-3σ,μ+3σ). S12: Determine the installation position and angle of the camera equipment, and connect the images captured by the camera equipment in the same scene; S13: Simulate and save videos of various scenarios and possible unsafe behaviors.

[0007] Furthermore, S2 includes the following steps: S21: Extract frames from the collected video data into images, and analyze and label the data on unsafe behaviors in each scenario; S22: Train a control group based on the labeled datasets for each scene; S23: Improve the YOLOv8 framework and use it as a framework for training models; S24: Modify and migrate the improved yolov8 network framework to the MindSpore framework, adjust the arrangement of images and annotation files in the dataset, install the MindSpore virtual environment and the third-party libraries required in the code. After the environment and dataset are ready, train different models for different scenarios on the AI ​​chip Ascend910 to obtain different models. S25: Building a C++ model inference application based on Ascend CANN.

[0008] Furthermore, S23 includes the following steps: S231: The YOLOv8n model network structure consists of three parts: Backbone, Neck, and Head. The input end of the model performs preprocessing operations on the input image. The backbone network uses the CSP module to divide the feature map into two branches, process them separately, and then merge them to enhance the learning effect and reduce the amount of computation. The C2f module is used to perform feature transformation, branch processing, and feature fusion on the input data. The Neck part uses a PAN-FPN structure, which combines the feature pyramid constructed by FPN with the upsampling, downsampling, and lateral connections of PAN to achieve multi-scale object detection. The loss function of YOLOv8 is VFL Loss for classification and DFL Loss + CIOU Loss for regression. S232: The backbone network adds an attention mechanism module, replacing the combination of Focal Loss and GIoU Loss with the DIoU loss function; S233: Introducing the DIoU loss function. During training, when the center point of the insulator dataset is at the same location as the predicted bounding box, the DIoU regression value remains the same despite the different shapes of the predicted bounding boxes. The formula for the DIoU loss function is: In the formula: The loss is Distance-IoU; IoU is the intersection-union ratio, representing the degree of overlap between the predicted and ground truth boxes; ρ is the Euclidean distance; c is the diagonal distance between the predicted and ground truth boxes and the smallest bounding rectangle within the predicted box; b is the coordinate of the center point of the predicted box; b gt The coordinates of the center point of the true bounding box; S234: The YOLOv8-ER algorithm is proposed for detecting the trampling winch scene. The efficient channel attention mechanism ECA is integrated in the backbone region to enhance the model's efficiency in extracting image features; the REP technology is added in the neck region.

[0009] Furthermore, step S3 includes the following steps: S31: Accelerate inference through quantization training, reduce quantization error through fine-tuning training, insert pseudo-quantization nodes to simulate the error introduced by quantization, fold the attributes of pseudo-quantization nodes into tensors during edge inference, and use parameters with quantization attributes during inference. S32: Reduce the number of bits required for weight representation / activation by parameter pruning, remove redundant non-critical weights for model compression, calculate the sum of the absolute values ​​of each column for each convolutional kernel, sort the summation results, prune m columns corresponding to the kernel matrix and the corresponding output, remove the rows corresponding to the kernel matrix of the corresponding input, and retain the remaining kernel weights.

[0010] Furthermore, in S4, after an unsafe behavior is detected by a general detection model for all unsafe behaviors, the corresponding sub-model is called. The sub-model includes a segmentation model, a human posture recognition model, and a personnel intrusion detection model. The camera device splices and archives the unsafe behavior videos collected each time. The program background records the time, personnel, and location of each unsafe behavior in the database and performs quantitative analysis curves on the number and frequency of each behavior.

[0011] Furthermore, S5 includes the following steps: S51: Design an unsafe behavior recognition video analysis system platform, and build modules including data source, basic management, task management, data management, external interface, and middleware to realize the display and recognition of various types of unsafe behaviors from multiple data sources and data management, and open external interfaces; S52: When the camera device detects an unsafe act, draw an outer box for the act on the software interface and write the corresponding name of the unsafe act. At the same time, detect the specific person who committed the act and archive the location, person, name of the unsafe act, and corresponding saved video in the background database for the software administrator to query. S53: When the camera detects that a person is engaging in unsafe behavior, the alarm device will broadcast the detection result and provide a reminder.

[0012] Furthermore, the present invention also provides an apparatus for performing the above-described data processing method.

[0013] Furthermore, the present invention also provides an apparatus including a memory, a processor, and an algorithm stored in the memory and executable on the processor, wherein the processor implements the above-described data processing method when executing the computer program.

[0014] Furthermore, the present invention also provides a computer-readable storage medium storing a computer algorithm, which, when executed by a processor, performs the above-described data processing.

[0015] The advantages and positive effects of this invention are: This invention makes efficient use of hardware resources while comprehensively capturing scenes where unsafe behavior may occur. It employs different network frameworks and recognition algorithms for specific scenarios, deploying and integrating the recognition results of multiple models, and achieving intelligent linkage across different camera devices, thus achieving diverse and accurate detection results. Compression techniques such as pruning and quantization of model operators reduce memory usage and computational complexity on the limited equipment resources of offshore platforms, improving the algorithm's computational speed. Furthermore, it can use alarm devices to alert and deter individuals from engaging in unsafe behavior. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the structure of the improved training model according to an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the architecture of the yolov8-ER algorithm model in an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the quantization training steps in an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of the parameter pruning steps in an embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram of the software architecture of an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The embodiments of the present invention will be further described below with reference to the accompanying drawings: like Figure 1-6 As shown, the method for identifying and issuing early warnings of unsafe behaviors in offshore oil operations includes the following steps: S1: Analyze historical statistical data along with actual scenarios to determine the installation location and angle of the camera equipment. Combine this with historical unsafe behavior image data to simulate unsafe behaviors on-site, collect video footage, and create a dataset. Specifically, S1 includes the following steps: S11: Conduct on-site surveys of the offshore platform environment, statistically record scenarios where unsafe behaviors may occur, establish a normal distribution function for the types of unsafe behaviors that may occur in each scenario and the number of times unsafe behaviors have occurred in the past in the scenario, and install the camera equipment in scenarios where the values ​​are distributed in (μ-3σ,μ+3σ).

[0024] S12: Determine the installation position and angle of the camera equipment to ensure seamless transitions between images captured by cameras within the same scene. Preferably, in this embodiment, the installation position and angle of the camera equipment are determined based on the location of each scene on the ship, the field of view of the scene area, and the relative position of each scene where a camera needs to be installed, thus ensuring seamless transitions between images captured by cameras within the same scene.

[0025] S13: Simulate and save videos for various scenarios and possible unsafe behaviors. Preferably, in this embodiment, since unsafe behaviors on offshore platforms are not frequent occurrences and there is a lack of datasets for specific scenarios, the simulation is carried out on the offshore platform for each scenario and each possible unsafe behavior with the assistance of on-site seafarer colleagues, by combining network data, past recorded images from different platforms, and filming in a simulated environment built by the company.

[0026] S2: Analyze and label the scenarios in the dataset, determine the detection network framework required for each behavior, and complete training and edge deployment. Specifically, S2 includes the following steps: S21: Extract frames from the video data collected in S1 into images, and analyze and label the data on unsafe behaviors in each scenario. Preferably, this embodiment analyzes the data on unsafe behaviors in different scenarios and designs different labeling standards for labeling based on the actual situation. For example, for static unsafe behaviors such as whether or not a safety helmet or safety clothing is worn, target detection bounding boxes are used for labeling; for dynamic motion behaviors such as the position of the drill operator, human posture key points are used to identify the position of the hands; for behaviors such as whether or not a lifting equipment is used to push or pull a lever, polygon segmentation is used for labeling, and so on, using different labeling methods in combination with specific scenarios.

[0027] S22: Train the model on the labeled datasets for each scene to form a control group. Preferably, this embodiment analyzes the model architecture with the best recognition performance for different scenes based on past project experience. Different models are trained on datasets with the same annotation method for different scenes to form a control experimental group, from which the model with the highest actual recognition accuracy and fastest recognition speed is selected, such as... Figure 2 As shown.

[0028] S23: Improve the YOLOv8 framework and use it as the framework for training models, such as... Figure 3 As shown. Specifically, S23 includes the following steps: S231: The YOLOv8n model network structure consists of three parts: Backbone, Neck, and Head. The input end of the model performs preprocessing operations on the input image. The backbone network uses the CSP module to divide the feature map into two branches, process them separately, and then merge them to enhance the learning effect and reduce the amount of computation. The C2f module is used to perform feature transformation, branch processing, and feature fusion on the input data. In the Neck part, a PAN-FPN structure is used, which combines the feature pyramid constructed by FPN with the upsampling, downsampling, and lateral connections of PAN to achieve multi-scale object detection. The loss function of YOLOv8 is VFL Loss for classification and DFL Loss + CIOU Loss for regression.

[0029] S232: An attention mechanism module is added to the backbone network, replacing the combination of Focal Loss and GIoU Loss with the DIoU loss function. Specifically, the improved ED-YOLOv8n uses hooks, lifted goods, and workers in the image as extracted features to detect whether workers are wearing safety helmets. Replacing the combination of Focal Loss and GIoU Loss with the DIoU loss function speeds up the convergence.

[0030] S233: Introducing the DIoU loss function accelerates bounding box regression, improves localization accuracy, and speeds up target detection. It minimizes and normalizes the center point distance, accelerating the convergence process. Furthermore, the DIoU loss function is an optimization of the GIoU loss function, resulting in even faster convergence. During training, when the center point of the insulator dataset is the same as the predicted bounding box, the DIoU regression value remains the same despite different predicted box shapes. The formula for the DIoU loss function is: In the formula: The loss is Distance-IoU; IoU is the intersection-union ratio, representing the degree of overlap between the predicted and ground truth boxes; ρ is the Euclidean distance; c is the diagonal distance between the predicted and ground truth boxes and the smallest bounding rectangle within the predicted box; b is the coordinate of the center point of the predicted box; b gt The coordinates are the center points of the ground truth bounding boxes. DIoU considers both the overlap area and the center point distance between the predicted and ground truth bounding boxes. When a predicted bounding box is inside a ground truth bounding box and the two overlap, referencing the center point distance can accelerate the bounding box regression rate, thereby speeding up the convergence of the DIoU loss function.

[0031] S234: The YOLOv8-ER algorithm is proposed for detection of the winch scene. ① An efficient channel attention mechanism (ECA) is integrated into the backbone region to enhance the model's efficiency in extracting image features, effectively reduce the number of parameters, and improve the coal gangue feature recognition ability and generalization performance in complex backgrounds. ② REP technology is added to the neck region to reduce the model size, improve the running speed, and ensure that the model can run efficiently even under limited hardware resources.

[0032] S24: Modify and migrate the improved YOLOv8 network framework to the MindSpore framework, adjust the arrangement of images and labeled files in the dataset, install the MindSpore virtual environment and the third-party libraries required in the code. After the environment and dataset are ready, train different models for different scenarios on the Ascend AI chip Ascend910 to obtain different models.

[0033] S25: Build a C++ model inference application based on Ascend CANN. Preferably, due to the limited inference speed of Python, in order to improve the accuracy and real-time performance of unsafe behavior detection during program execution, a C++ model inference application based on AscendCANN (Compute Architecture for Neural Networks) is written. This improves inference speed while ensuring the domestic production of the model training and inference parts, further reducing the risk of being constrained by foreign hardware and algorithms.

[0034] S3: Prune and quantize the model operators to reduce the model's memory usage and computational complexity. Specifically, S3 includes the following steps: S31: Accelerates inference through quantization training, overcoming the hardware resource constraints of offshore platforms. Fine-tuning reduces quantization error by inserting pseudo-quantization nodes to simulate the error introduced by quantization. During edge inference, the attributes of the pseudo-quantization nodes are folded into the tensor, and parameters with quantization attributes are used during inference, such as... Figure 4 As shown.

[0035] S32: Reduce the number of bits required for weight representation / activation through parameter pruning, remove redundant and non-critical weights for model compression, calculate the sum of the absolute values ​​of each column for each convolutional kernel, sort the summation results, prune m columns corresponding to the kernel matrix and their corresponding outputs, remove the rows corresponding to the inputs in the kernel matrix, and retain the remaining kernel weights, such as... Figure 5 As shown.

[0036] S4: For multiple detection scenarios, various network architectures are combined and multiple network models are integrated to achieve the linkage of the various camera devices. Specifically, after an unsafe behavior is detected by a general detection model for all unsafe behaviors, the corresponding sub-models are further called. The sub-models include a segmentation model, a human pose recognition model, and a personnel intrusion detection model to achieve more accurate specific task requirements, such as segmenting the position of a hand. In practical applications, this mode has been found to save on device memory and performance while achieving faster computing speed.

[0037] For example, after the general detection model detects unsafe behaviors during drilling operations, the segmentation model is called to segment the drill rig gap and the bottom fixing slot, as well as the human's hands and feet. The ROI overlap area is calculated with the segmented areas of the drill rig gap and the bottom fixing slot to further confirm whether the person has indeed performed an unsafe operation.

[0038] For example, when detecting whether someone has entered a hoisting area, if cameras around the camera in that area detect someone passing by and heading towards the hoisting area, the program automatically starts calling the intrusion detection model. This method is also applicable to behaviors such as smoking. For example, if the current camera detects someone smoking, the program will automatically detect the direction the person is walking and estimate their location. If the person has not extinguished their cigarette before leaving the camera, the program will intelligently link with other cameras in the area to continuously record the person's behavior and location, making it easier for inspectors to locate the person in a timely manner and stop unsafe behaviors that pose a safety hazard.

[0039] The camera equipment will stitch together and archive the videos of unsafe behaviors collected each time they are linked together. The program background will record the time, people and locations of each unsafe behavior in the database and make quantitative analysis curves for the number of times and frequency of each behavior.

[0040] S5: Display the video information captured by the camera equipment in real time on the software. The software integrates a visualization detection algorithm module, a historical behavior storage and operation module, and an early warning alarm module. Alarm devices are also set up for each camera equipment. Specifically, S5 includes the following steps: S51: Design an unsafe behavior recognition video analysis system platform, and build modules including data source, basic management, task management, data management, external interface, and middleware to realize the display and recognition of various types of unsafe behaviors from multiple data sources and data management, and open external interfaces.

[0041] S52: When an unsafe act is detected by the camera device, an outer box is drawn for the act on the software interface and the corresponding name of the unsafe act is written. At the same time, the specific person who committed the act is detected, and the location, person, name of the unsafe act, and corresponding saved video are archived in the background database for the software administrator to query.

[0042] S53: When the camera detects unsafe behavior, the alarm device broadcasts and alerts the user. This effectively alerts to unsafe behavior and, to some extent, can interrupt and stop it.

[0043] The advantages and positive effects of this invention are: This invention makes efficient use of hardware resources while comprehensively capturing scenes where unsafe behavior may occur. It employs different network frameworks and recognition algorithms for specific scenarios, deploying and integrating the recognition results of multiple models, and achieving intelligent linkage across different camera devices, thus achieving diverse and accurate detection results. Compression techniques such as pruning and quantization of model operators reduce memory usage and computational complexity on the limited equipment resources of offshore platforms, improving the algorithm's computational speed. Furthermore, it can use alarm devices to alert and deter individuals from engaging in unsafe behavior.

[0044] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for identifying and issuing early warnings of unsafe behaviors in offshore oil operations, characterized by: Includes the following steps, S1: Analyze historical statistical data along with actual scenarios to determine the installation location and angle of the camera equipment. Combine historical unsafe behavior image data to simulate unsafe behavior on-site, collect video, and create a dataset. S2: Analyze and label the scenarios in the dataset, determine the detection network framework to be used for each behavior, and complete the training and edge deployment; S3: Prune and quantize the model operators to reduce the model's memory usage and computational complexity; S4: For multiple detection scenarios, various network architectures are combined and multiple network models are integrated to achieve the linkage of the various camera devices; S5: Display the video information collected by the camera device in real time on the software. The software integrates a visualization detection algorithm module, a historical behavior storage and operation module, and an early warning alarm module. At the same time, alarm devices are set for each of the camera devices.

2. The method for identifying and warning of unsafe behaviors in offshore oil operations according to claim 1, characterized in that: S1 includes the following steps: S11: Conduct on-site surveys of the offshore platform environment, statistically record scenarios where unsafe behaviors may occur, establish a normal distribution function for the types of unsafe behaviors that may occur in each scenario and the number of times unsafe behaviors have occurred in the past in the scenario, and install the camera equipment in scenarios where the values ​​are distributed in (μ-3σ,μ+3σ). S12: Determine the installation position and angle of the camera equipment, and connect the images captured by the camera equipment in the same scene; S13: Simulate and save videos of various scenarios and possible unsafe behaviors.

3. The method for identifying and warning of unsafe behaviors in offshore oil operations according to claim 1 or 2, characterized in that: S2 includes the following steps: S21: Extract frames from the collected video data into images, and analyze and label the data on unsafe behaviors in each scenario; S22: Train a control group based on the labeled datasets for each scene; S23: Improve the YOLOv8 framework and use it as a framework for training models; S24: Modify and migrate the improved yolov8 network framework to the MindSpore framework, adjust the arrangement of images and annotation files in the dataset, install the MindSpore virtual environment and the third-party libraries required in the code. After the environment and dataset are ready, train different models for different scenarios on the AI ​​chip Ascend910 to obtain different models. S25: Building a C++ model inference application based on Ascend CANN.

4. The method for identifying and warning of unsafe behaviors in offshore oil operations according to claim 3, characterized in that: S23 includes the following steps: S231: The YOLOv8n model network structure consists of three parts: Backbone, Neck, and Head. The input end of the model performs preprocessing operations on the input image. The backbone network uses the CSP module to divide the feature map into two branches, process them separately, and then merge them to enhance the learning effect and reduce the amount of computation. The C2f module is used to perform feature transformation, branch processing, and feature fusion on the input data. The Neck part uses a PAN-FPN structure, which combines the feature pyramid constructed by FPN with the upsampling, downsampling, and lateral connections of PAN to achieve multi-scale object detection. The loss function of YOLOv8 is VFL Loss for classification and DFL Loss + CIOU Loss for regression. S232: The backbone network adds an attention mechanism module, replacing the combination of Focal Loss and GIoU Loss with the DIoU loss function; S233: Introducing the DIoU loss function. During training, when the center point of the insulator dataset is at the same location as the predicted bounding box, the DIoU regression value remains the same despite the different shapes of the predicted bounding boxes. The formula for the DIoU loss function is: In the formula: The loss is Distance-IoU; IoU is the intersection-union ratio, representing the degree of overlap between the predicted and ground truth boxes; ρ is the Euclidean distance; c is the diagonal distance between the predicted and ground truth boxes and the smallest bounding rectangle within the predicted box; b is the coordinate of the center point of the predicted box; b gt The coordinates of the center point of the true bounding box; S234: The YOLOv8-ER algorithm is proposed for detecting the trampling winch scene. The efficient channel attention mechanism ECA is integrated in the backbone region to enhance the model's efficiency in extracting image features; the REP technology is added in the neck region.

5. The method for identifying and warning of unsafe behaviors in offshore oil operations according to claim 1 or 2, characterized in that: S3 includes the following steps: S31: Accelerate inference through quantization training, reduce quantization error through fine-tuning training, insert pseudo-quantization nodes to simulate the error introduced by quantization, fold the attributes of pseudo-quantization nodes into tensors during edge inference, and use parameters with quantization attributes during inference. S32: Reduce the number of bits required for weight representation / activation by parameter pruning, remove redundant non-critical weights for model compression, calculate the sum of the absolute values ​​of each column for each convolutional kernel, sort the summation results, prune m columns corresponding to the kernel matrix and the corresponding output, remove the rows corresponding to the kernel matrix of the corresponding input, and retain the remaining kernel weights.

6. The method for identifying and warning of unsafe behaviors in offshore oil operations according to claim 1 or 2, characterized in that: In step S4, after an unsafe behavior is detected by a general detection model for all unsafe behaviors, the corresponding sub-model is called. The sub-model includes a segmentation model, a human posture recognition model, and a personnel intrusion detection model. The camera device splices and archives the unsafe behavior videos collected each time. The program background records the time, personnel, and location of each unsafe behavior in the database and performs quantitative analysis curves on the number and frequency of each behavior.

7. The method for identifying and warning of unsafe behaviors in offshore oil operations according to claim 1 or 2, characterized in that: S5 includes the following steps: S51: Design an unsafe behavior recognition video analysis system platform, and build modules including data source, basic management, task management, data management, external interface, and middleware to realize the display and recognition of various types of unsafe behaviors from multiple data sources and data management, and open external interfaces; S52: When the camera device detects an unsafe act, draw an outer box for the act on the software interface and write the corresponding name of the unsafe act. At the same time, detect the specific person who committed the act and archive the location, person, name of the unsafe act, and corresponding saved video in the background database for the software administrator to query. S53: When the camera detects that a person is engaging in unsafe behavior, the alarm device will broadcast the detection result and provide a reminder.

8. An apparatus, characterized in that: The data processing method described in any one of claims 1 to 7 is executed.

9. An apparatus comprising a memory, a processor, and an algorithm stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the data processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer algorithm, characterized in that, When the computer algorithm is executed by the processor, it performs the data processing as described in any one of claims 1 to 7.

Images