Intelligent identification method and system for personnel status based on independent visual unit

Abstract

The present invention relates to a method and system for intelligent identification of personnel status based on an independent visual unit. The method detects the behavior status of targets in a monitoring area and identifies the behavior status of a predetermined target by constructing a lightweight neural network model; the lightweight neural network model is set in an integrated circuit board; the circuit board is deployed in an area to be monitored; result information is obtained by identification and processing of the lightweight neural network model in the integrated circuit board; the result information is output; the method provided by the present invention can achieve data confidentiality during the detection process, and the improved YOLO V7 model is integrated in a single-chip microcomputer, made into an integrated circuit board, and deployed locally in a confidential factory to ensure the confidentiality of the detection data. At the same time, due to the lightweight model, in order to facilitate the deployment of the detection model on the single-chip microcomputer, the improved YOLO V7 model is lightweight processed to ensure the feasibility of local deployment.

Description

Technical Field

[0001] The present invention relates to the technical field of safety detection, and in particular to a method and system for intelligently identifying personnel status based on an independent visual unit. Background Art

[0002] Factory personnel status monitoring primarily involves monitoring the activities of personnel at production and experimental sites to ensure they are in a safe position. Furthermore, it detects whether personnel at these sites are operating in accordance with standard specifications, ensuring that all personnel at the company's production and experimental sites are operating as directed. Monitoring personnel status at production and experimental sites is crucial for improving productivity, ensuring processing progress, and ensuring personnel safety. Therefore, this invention primarily addresses personnel status monitoring at production and experimental sites.

[0003] At present, on-site safety detection of personnel is mainly based on two methods: calibration object detection and machine vision. For the calibration five detection method, the main focus is to construct a working range for personnel restricted areas, and install corresponding infrared, distance and other sensors, and use signal sensing technology to sense whether personnel have entered the relevant area, thereby realizing alarm functions, etc. Based on this method, Hanwha Corporation has disclosed an invention patent for a worker detection device. The patent mentions the use of irradiation light to receive the light emitted from the light-emitting part, detect the danger of the worker movement area in the factory and judge the worker's work danger status. Sun Xiaowei and others from Hongyue Shenmei Jiangsu Information Technology Development Co., Ltd. proposed a method for detecting personnel flow in smart factories based on big data. The method proposes to use real-time clocking information, detection equipment, etc. to obtain the location status of personnel.

[0004] While these methods can capture the on-site activity status of personnel to a certain extent, they are characterized by strong customization and low flexibility. This makes it difficult to dynamically configure inspection content and detect defects based on the dynamic changes in production needs. Consequently, these methods are mostly applied to processing tasks with fixed production behavior patterns and are difficult to apply on a large scale in other fields.

[0005] Detection methods based on machine vision have the advantages of high flexibility and low cost. There are a large number of patents related to machine vision, such as the invention patent applied for by Zhang Wen and others from Beijing University of Chemical Technology, which uses a pixel-level annotation technology to construct a method for locating indoor workers at hazardous sources, in order to achieve personnel safety behavior detection. Situ Xuan and others from Qingyang Ruihua Energy Co., Ltd. announced a production safety personnel detection method based on deep learning. Using machine vision and other technologies, and deploying an IoT configuration server, they have built factory-level personnel safety detection technology. The above-mentioned machine vision-based personnel behavior status technology usually uses a camera to capture images of workers, which are then transmitted to the server using the Industrial Internet of Things. The acquired images are then calculated using the established visual model to achieve judgment of the behavior status of on-site personnel.

[0006] However, these visual inspection technologies are often difficult to apply to applications where data confidentiality is paramount. In these scenarios, the data collected by image sensors is often difficult to transmit remotely, requiring data processing to be performed locally on the image acquisition unit to detect a person's behavior.

[0007] Therefore, there are two typical problems when applying existing technologies in such scenarios.

[0008] Problem 1: The computational power of independent visual units is limited, making it difficult to support large-scale complex models. Therefore, it is necessary to develop a lightweight model that can be deployed and applied on small computing units and has sufficiently high detection accuracy.

[0009] Question 2: Traditional methods cannot integrate the image acquisition unit with the visual detection hardware unit, making it difficult to achieve on-machine data detection.

[0010] Therefore, it is necessary to develop a new hardware structure that can realize the acquisition and detection of visual images on the same machine without distance.

[0011] Based on the above requirements, it is very necessary to develop a personnel behavior recognition and detection device and detection technology based on independent visual units. Summary of the Invention

[0012] In view of this, the purpose of the present invention is to provide a method and system for intelligent identification of personnel status based on independent visual units, which can achieve a lightweight algorithm model and high-precision detection effect while ensuring data security.

[0013] In order to achieve the above object, the present invention provides the following technical solutions:

[0014] The method for intelligently identifying a person's status based on an independent visual unit provided by the present invention comprises the following steps:

[0015] S1: Constructing a lightweight neural network model; the lightweight neural network model is used to detect the behavior state of the target in the monitoring area and identify the behavior state of the target;

[0016] S2: Set the lightweight neural network model on an integrated circuit board and deploy the circuit board in the area that needs to be monitored;

[0017] S3: Obtain result information through identification processing of lightweight neural network model in integrated circuit board;

[0018] S4: Output result information.

[0019] Furthermore, the method further comprises the following steps:

[0020] S5: Upload the result information to the visual monitoring system for display.

[0021] Furthermore, the lightweight neural network model adopts an improved YOLO V7 model, and the improved YOLO V7 model includes an input end, a backbone network, a detection head, a feature fusion module, and an output end;

[0022] The input terminal is used to input production site data collected in real time;

[0023] The backbone network is used to perform feature processing on the collected data to obtain the characteristics of personnel behavior status;

[0024] The detection head is used to detect and identify the acquired field data;

[0025] The feature fusion module is used to fuse the features output from the detection head;

[0026] The output end is used to output the judgment result.

[0027] Furthermore, the integrated circuit board adopts a double-layer circuit board architecture, including a first layer board and a second layer board; the first layer board is provided with a camera module, a lightweight neural network deployment module, and a shared power supply; the second layer board is provided with a buzzer, an alarm light, and an infrared sensor;

[0028] The shared power supply is respectively connected to the camera module, the lightweight neural network deployment module, the alarm module, and the infrared sensor to provide power;

[0029] The camera module inputs the collected data into the lightweight neural network deployment module, and the lightweight neural network deployment module sends the processed signal to the alarm module;

[0030] The infrared sensor is connected to the lightweight neural network deployment module, and the infrared sensor is used to detect whether there is a target object in the monitoring area;

[0031] Furthermore, the backbone network in the improved YOLO V7 model adopts the YOLO V5 pruning algorithm, and the calculation formula of the lightweight convolution kernel is as follows:

[0032]

[0033] Where: c is the number of channels of the input feature map, which is mainly used for inputting scene environment image features;

[0034] n is the number of output feature channels, k is the size of ordinary convolution, which is used for dimensionality reduction;

[0035] s is the size of the depthwise convolution of the linear transformation set; d represents the step size; and rc represents the lightweight convolution result.

[0036] Furthermore, the infrared sensor is a passive infrared group. After a person enters the infrared detection area, the infrared group will deliver the detection signal to the power network; when there is no one, the camera module and the lightweight neural network deployment module are in the off state; when someone comes, the infrared ray is first identified to obtain the signal of whether there is anyone, the power is turned on, and the camera module and the lightweight neural network deployment module start working.

[0037] Furthermore, the method further includes evaluating the predicted data, wherein the evaluation is performed according to the following formula:

[0038]

[0039] Among them, TP represents a true prediction pair; FP represents a false prediction correct; FN represents a false prediction error;

[0040] P represents the precision rate; R represents the recall rate; F1 represents the correct evaluation index; IoU represents the precision rate.

[0041] The present invention provides an intelligent personnel status recognition method based on independent visual units, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The method is implemented when the processor executes the program.

[0042] The recognition system provided by the present invention, which is implemented according to the above-mentioned personnel status intelligent recognition method based on an independent visual unit, includes a housing, an integrated circuit board disposed in the housing for identifying the wearing status of a helmet, and a visual monitoring system for displaying the results processed by the integrated circuit board;

[0043] The integrated circuit board adopts a double-layer circuit board architecture, including a first layer board and a second layer board; the first layer board is provided with a camera module, a lightweight neural network deployment module, and a shared power supply; the second layer board is provided with a buzzer, an alarm light, and an infrared sensor;

[0044] The shared power supply is respectively connected to the camera module, the lightweight neural network deployment module, the alarm module, and the infrared sensor to provide power;

[0045] The camera module inputs the collected data into the lightweight neural network deployment module, and the lightweight neural network deployment module sends the processed signal to the alarm module;

[0046] The infrared sensor is connected to the lightweight neural network deployment module, and the infrared sensor is used to detect whether there is a target object in the monitoring area.

[0047] Furthermore, the lightweight neural network model adopts an improved YOLO V7 model, and the improved YOLO V7 model includes an input end, a backbone network, a detection head, a feature fusion module, and an output end;

[0048] The input terminal is used to input production site data collected in real time;

[0049] The backbone network is used to perform feature processing on the collected data to obtain the characteristics of personnel behavior status;

[0050] The detection head is used to detect and identify the acquired field data;

[0051] The feature fusion module is used to fuse the features output from the detection head;

[0052] The output end is used to output the judgment result.

[0053] The beneficial effects of the present invention are:

[0054] The present invention provides a method and system for intelligently identifying personnel status based on independent visual units. This method uses independent visual units to identify the wearing of safety helmets in confidential factories, thereby identifying personnel status. This method does not rely on the Internet platform, which solves the data confidentiality risk problem caused by the need to upload data; at the same time, a lightweight neural network model is used for the recognition algorithm, which solves the problem that traditional algorithms are difficult to effectively deploy in independent hosts due to large computational complexity. Since conventional visual detection algorithms have low accuracy in complex environments, they are difficult to apply to the problem of safe identification of multiple types of personnel behavior in complex industrial sites. The method uses the identification of the wearing of safety helmets in confidential factories to identify personnel status, which reduces the complexity of the recognition calculation. The model can run in embedded systems, reducing the dependence of traditional visual models on computing resources. This model is a high-precision machine vision calculation model that can accurately identify the behavioral status of established personnel and detect the behavioral status of factory personnel;

[0055] The visual acquisition unit hardware circuit platform on the integrated circuit board not only enables independent acquisition of visual images, but also enables on-machine processing of given visual images, thus avoiding the disadvantage of traditional methods that require network transmission to the server for image processing.

[0056] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0058] Figure 1 The figure is a flow chart of the intelligent identification method of personnel status based on independent visual units.

[0059] Figure 2 Schematic diagram of the hardware circuit board of the independent vision unit.

[0060] Figure 3 It is a lightweight neural network algorithm process.

[0061] Figure 4 Schematic diagram of algorithm training based on attention mechanism.

[0062] Figure 5 Schematic diagram of the algorithm's pruning behavior.

[0063] Figure 6 This is a schematic diagram of the program interface effect of this method.

[0064] Figure 7 This is the improved YOLO V7 model structure diagram.

[0065] Figure 8 This is the model prediction result graph. DETAILED DESCRIPTION

[0066] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0067] Example 1

[0068] like Figure 1 As shown, Figure 1 This is a flow chart of a method for intelligently identifying personnel status based on independent visual units. This embodiment provides a method and system for intelligently identifying personnel status based on independent visual units. This method uses independent visual units to identify the activity range of personnel in confidential factories and the helmet wearing behavior to realize the detection process.

[0069] The following steps are involved:

[0070] S1: Constructing a lightweight neural network model; the lightweight neural network model is used to detect the behavior of people in the monitoring area and identify the status of the target wearing a helmet, work clothes, etc.;

[0071] S2: Set the lightweight neural network model on an integrated circuit board and deploy the circuit board in the area that needs to be monitored;

[0072] S3: Obtain result information through identification processing of lightweight neural network model in integrated circuit board;

[0073] S4: output result information;

[0074] S5: uploading the result information to a visual monitoring system for display;

[0075] The lightweight neural network model in this embodiment adopts the lightweight YOLO V7 model, which is an improved YOLO V7 model. The lightweight YOLO V7 model includes: a backbone layer, a neck layer, and a head layer; wherein the backbone layer is mainly responsible for extracting features from the input image to obtain features such as the behavior and activity status of the personnel and the wearing status of the helmet; Figure 3 As shown in the figure, the specific structure is as follows: the Neck layer is responsible for multi-scale feature fusion of the extracted personnel behavior status feature map and passing these features to the prediction layer; the Head layer is responsible for the final personnel behavior status warning operation;

[0076] The meanings of the letters in the box are: cons-dsc is a separable convolution operation; mb is the sensitivity and activation threshold; SPPF is for achieving rapid fusion of local features and global features at the featherMap level, namely the features of personnel behavior status and helmet wearing features; GSConv is the operation of fusing ordinary convolution downsampling with DWconv deep convolution; Concat is a convolution cascade; Upsample is upsampling; LE is Lenet, which is used to recognize the captured image; Head is a prediction module, which is mainly used to characterize and recognize the behavior status of personnel.

[0077] like Figure 2 As shown, Figure 2 (a) shows the physical structure diagram of the integrated circuit board, and (b) shows the principle block diagram of the integrated circuit board. Figure 2 The specific meanings of circles 1 to 5 in (b) are as follows: Circle 1 is the captured image data transmitted to the neural network deployment module, Circle 2 is the judgment operation performed by the neural network, Circle 3 is the identification of personnel behavior status based on the neural network, and HC-SR501 performs operations based on the results of the aforementioned transmission, Circle 4 is the indicator light operation, using different LED combinations to identify personnel behavior status, and Circle 5 is the abnormal alarm circuit. When the personnel behavior status is abnormal, the buzzer sounds an alarm.

[0078] The integrated circuit board in this embodiment adopts a double-layer circuit board architecture, including a first layer board and a second layer board; the first layer board is provided with a camera module, a lightweight neural network deployment module, and a shared power supply; the second layer board is provided with a buzzer, an alarm light, and an infrared sensor;

[0079] The shared power supply is respectively connected to the camera module, the lightweight neural network deployment module, the alarm module, and the infrared sensor to provide power;

[0080] The camera module inputs the collected data into the lightweight neural network deployment module, and the lightweight neural network deployment module sends the processed signal to the alarm module;

[0081] The infrared sensor is connected to the lightweight neural network deployment module, and the infrared sensor is used to detect whether there is a target object in the monitoring area.

[0082] Example 2

[0083] like Figure 1 As shown, Figure 1This is a flow chart of a method for intelligently identifying personnel status based on an independent visual unit. This embodiment provides a method and system for intelligently identifying personnel status based on an independent visual unit. This method uses an independent visual unit to implement a detection process for identifying safety helmets worn in confidential factories. First, a designated industrial camera is used to acquire video stream data of personnel behavior status. Based on the acquired video stream data, the trained personnel behavior status is used to identify the status. On this basis, a lightweight neural network model is constructed and deployed on an integrated circuit board. Finally, a visual interface system is used to provide real-time prompts of personnel safety behavior status.

[0084] like Figure 2 As shown, Figure 2 This is a schematic diagram of the hardware circuit board of the independent vision unit. The integrated circuit board of this embodiment integrates the independent vision unit, a lightweight neural network model, and an on-site monitoring unit. The independent vision unit is used for visual acquisition and is embedded in the industrial camera, data acquisition board, and independent computing unit.

[0085] The integrated circuit board includes an Arduino Nano 33BLE Sense board, an OV7675 camera module, a lightweight neural network deployment module, two LEDs, a buzzer, and an HC-SR501 infrared sensor.

[0086] The integrated circuit board adopts a double-layer circuit board architecture. The camera module OV7675, lightweight neural network deployment module and shared power supply are arranged on the first layer of the board; the buzzer, alarm light LED and infrared sensor HC-SR501 are arranged on the second layer of the board.

[0087] Among them, the first layer board is connected to the opening and closing part of the second layer board on the right side through a cable, and is connected to the right LED circuit and the corresponding right buzzer through another cable.

[0088] S3: The HC-SR501 sensor is a passive infrared sensor. When a person enters the infrared detection area, the infrared sensor transmits a detection signal to the power network. When no one is present, the camera and algorithm modules are disabled to reduce power consumption. When a person approaches, the infrared sensor first identifies the person, detects the presence signal, and then powers on. Once powered on, the camera image and neural network algorithm begin to work.

[0089] The camera and algorithm deployment modules complete the startup process.

[0090] The OV7675 camera module automatically obtains video stream data on the behavior status of on-site personnel. The obtained data will be automatically uploaded to the lightweight neural network module for judgment through the neural network. After the judgment is completed, the visual inspection results will be further alarmed or improved by the honey device.

[0091] The lightweight neural network model in this embodiment is mainly based on the lightweight YOLO V7. The overall algorithm architecture is as follows Figure 3 As shown, Figure 3 It is a lightweight neural network algorithm process, and the lightweight neural network includes a Backbone module, a Neck module, and a Head module;

[0092] The Backbone module is responsible for extracting features from the input image to obtain features such as the behavior and activity status of the person and the helmet wearing status;

[0093] The Neck module is responsible for performing multi-scale feature fusion on the extracted personnel behavior state feature map and passing these features to the prediction layer;

[0094] The Head module is responsible for performing the final personnel behavior status warning operation;

[0095] Among them, the Backbone module (backbone network) adopts the Yolo v5 pruning algorithm, and the calculation formula of the lightweight convolution kernel is as follows:

[0096]

[0097] Where: c is the number of channels of the input feature map, which is mainly used for inputting scene environment image features;

[0098] n is the number of output feature channels, k is the size of ordinary convolution, which is used for dimensionality reduction;

[0099] s is the size of the depthwise convolution of the linear transformation set; d is the step size; r c Represents the lightweight convolution result;

[0100] In this embodiment, d and k are respectively set to 3. By using a lightweight convolution kernel, the number of neural network parameters can be effectively reduced, and the overall number of parameters can be reduced by about 50%.

[0101] After the previous convolution step, in order to adapt to factory conditions, an algorithm that combines the CBAM attention mechanism with the parameterless attention mechanism is introduced to increase the recognition accuracy under complex lighting conditions in the factory. The CAM attention mechanism modules involved are as follows: Figure 4 As shown, Figure 4 This is a schematic diagram of algorithm training based on the attention mechanism. The algorithm training process of the attention mechanism is as follows:

[0102] Based on this module, the main attention expressions involved are:

[0103]

[0104] Among them, MC (F) represents; σ represents activation function; α1 represents Lagrange multiplier 1; α2 represents Lagrange multiplier 1; represents the average value of input features; Indicates the maximum value of the feature map;

[0105] Use average pooling and maximum pooling to perform pooling operations on the output features of the lightweight convolution in S4 to obtain the pooled feature map information Mc(F);

[0106] On this basis, the MLP shared network is used to construct the attention feature map through the activation function σ. In this embodiment, the activation function is the Selu activation function, and its calculation formula is as follows:

[0107]

[0108] Where x represents the input parameter of the previous layer; λ represents the multiplier coefficient; α represents the learning Lagrangian logarithm coefficient;

[0109] λ≈1.0507,α≈1.6733, which can be adjusted and optimized according to the training fit.

[0110] After the above feature maps and activation functions, a parameter-free attention mechanism is introduced at the end of the neural network. By using this mechanism, the number of parameters can be effectively reduced while further improving the model performance.

[0111] like Figure 4 As shown, the attention mechanism in this embodiment is mainly introduced at the end of the neural network. Figure 4 The channel attention feature map in . Its symbol is SimAM, and its calculation formula is as follows

[0112]

[0113] In the formula, w, b represent the weight and bias of the model training process, y represents the output, and x i represents the input at the i-th time point, that is, the photo input at the i-th time point, t is the time series point, M is the energy function, y t is the output at time t, Represent the maximum likelihood estimates of the input and t function respectively.

[0114] like Figure 5 As shown, Figure 5 Schematic diagram of the pruning behavior algorithm. The specific process of the pruning behavior algorithm is as follows:

[0115] The specific calculation formula is as follows:

[0116]

[0117] Where E represents grouping across channels and spatial dimensions. In the parameterless attention mechanism, Sigmoid is used for calculation, and X represents the number of input parameters. The calculation formula of Sigmoid is as follows:

[0118]

[0119] Where S(x) represents the simgmoid function;

[0120] After constructing the above-mentioned convolution, pooling, and attention mechanisms, the algorithm further introduces the anchoring technology framework system given to K-means to ensure that the model algorithm reaches the optimal level while reducing the amount of training.

[0121]

[0122] Where P1 represents the k-means probability; D(x) represents the distance from the sampling point x to the target sample.

[0123] After building the neural network based on the above, to ensure the adaptability of the model, an adaptive prior box algorithm model is used. When the model is trained for different downstream tasks, the maximum polarization algorithm is used to generate a prior box based on the area with the highest bounding box coverage frequency in the training results.

[0124] A random mosaic data augmentation algorithm model was added to introduce noise into the data input, reducing the likelihood of overfitting and improving model robustness. This prior box was also used during model prediction, allowing the model to focus on areas with higher information concentration.

[0125] At the same time, the method provided in this embodiment also proposes an evaluation index system based on the following:

[0126]

[0127] Among them, TP represents a true prediction pair; FP represents a false prediction correct; FN represents a false prediction error;

[0128] P represents the precision rate; R represents the recall rate; F1 represents the correct evaluation index; IoU represents the precision rate;

[0129] After the model development is completed, in order to ensure on-site readability, a field monitoring system as shown in the figure is further constructed based on C# and pyqt. Figure 6 As shown, Figure 6 This is a schematic diagram of the program interface effect of this method. Figure 6 The middle figure is a comparison chart before and after training, which is used to characterize the behavioral status of personnel for case identification.

[0130] like Figure 7 As shown, Figure 7 This is the improved YOLO V7 model structure diagram. Figure 3 is a lightweight neural network algorithm flow chart, and Figure 3 For a simple neural network, Figure 7 The architecture behind the neural network deployment on the circuit board;

[0131] The improved YOLO V7 model includes an input end, a backbone network, a detection head, a feature fusion module, and an output end;

[0132] The input end is used to input the production site data collected in real time; the input feature in this embodiment uses a feature line extracted from multiple lines for feature extraction, and the feature extraction capability of this method is stronger; compared with the original yolo v7 input feature extraction of only a single line, the input of multiple feature lines makes the feature extraction capability stronger;

[0133] The backbone network is used to perform feature processing on the collected data to obtain the characteristics of personnel behavior status;

[0134] The detection head is mainly a combination of hardware and software, used to detect and identify the acquired field data;

[0135] The feature fusion module is used to fuse the features output from the detection head and finally output the fused features to reduce the feature output amount;

[0136] The output terminal is used to output the judgment result for subsequent operation;

[0137] Figure 7 The CBS in the image is divided into two branches according to the input feature map. One branch is used to generate a binary code for the obtained personnel behavior state, and the other branch is used to extract further personnel behavior state features. ELAN represents efficient segmentation of the established super-resolution network SwinIR, which improves network speed and reduces network calculation amount; MP-1 represents the first type of McCulloch and Pitts neurons in the present invention; MP-2 represents the second type of McCulloch and Pitts neurons in the present invention; ELAN-W represents the SwinIR operation after segmentation of w; cat represents the connection of the neural network model; UPSample represents upsampling; SPPCSPC represents the pyramid pooling module; REP represents a network module that divides the backbone of personnel behavior state; CBM represents the convolution operation on the personnel behavior state image, which is mainly composed of three modules: conv+batchnorm+mish; FUSE represents fully connected fusion;

[0138] like Figure 8As shown, Figure 8 This is the model prediction result graph, Figure 8 The video stream data displayed in the image can be identified with an accuracy of more than 90%. Figure 8 The blue frame of the person in the middle indicates that he has entered the danger zone, and the red frame indicates that it is normal. The probability of the head floating above the head indicates the status parameter of the person wearing a safety helmet.

[0139] The method provided in this embodiment can achieve data confidentiality during the detection process. The improved YOLO V7 model is integrated into a single-chip microcomputer, made into an integrated circuit board, and deployed locally in a confidential factory to ensure the confidentiality of the detection data.

[0140] At the same time, due to the lightweight model, in order to facilitate the deployment of the detection model on the single-chip computer, the improved YOLO V7 model was lightweighted to ensure the feasibility of local deployment.

[0141] Due to its high information concentration and low computational cost, the model can learn key areas in the dataset (such as workstations and corridors where people gather in the workshop) during training and use these as prior boxes for predictions. This allows the model to effectively focus on important parts of the input data compared to ordinary models, further enabling faster computational speeds when making predictions, especially when processing data types such as video streaming data that require high-frequency predictions.

[0142] An on-site monitoring system has been built to enable visual monitoring of on-site personnel; while accurately identifying personnel behavior, it ensures the reliability and consistency of the results.

[0143] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An intelligent identification method for personnel status based on independent visual units, characterized in that: The following steps are involved: S1: constructing a lightweight neural network model; the lightweight neural network model is used to detect the behavior state of the target in the monitoring area and identify the behavior state of the target; S2: Setting the lightweight neural network model in an integrated circuit board; and deploying the circuit board in the area that needs to be monitored; S3: Obtain result information through identification processing of lightweight neural network model in integrated circuit board; S4: output result information; The lightweight neural network model adopts an improved YOLO V7 model, and the improved YOLO V7 model includes an input end, a backbone network, a detection head, a feature fusion module, and an output end; The input end is used to input the production site data collected in real time; The backbone network is used to perform feature processing on the collected data to obtain the characteristics of the personnel behavior status; The detection head is used to detect and identify the acquired field data; The feature fusion module is used for fusion processing of features output from the detection head; The output terminal is used to output the judgment result; The integrated circuit board adopts a double-layer circuit board architecture, including a first layer board and a second layer board; the first layer board is provided with a camera module, a lightweight neural network deployment module, and a shared power supply; the second layer board is provided with a buzzer, an alarm light, and an infrared sensor; The common power supply is respectively connected to the camera module, the lightweight neural network deployment module, the alarm module, and the infrared sensor to provide power; The camera module inputs the collected data into the lightweight neural network deployment module, and the lightweight neural network deployment module sends the processed signal to the alarm module; The infrared sensor is connected to the lightweight neural network deployment module, and the infrared sensor is used to detect whether there is a target object in the monitoring area; The backbone network in the improved YOLO V7 model adopts the YOLO V5 pruning algorithm, and the calculation formula of the lightweight convolution kernel is as follows: Where: c is the number of channels of the input feature map, which is mainly used for the input scene environment image features; n is the number of output feature channels, k is the size of the ordinary convolution, which is used for dimensionality reduction; s is the size of the depthwise convolution of the linear transformation set; d is the step size; r c Represents the lightweight convolution result; It also includes evaluating the predicted data, which is performed according to the following formula: 、 、 、 ； in, represents the true prediction pair; indicates that the wrong prediction is correct; indicates a misprediction error; represents the precision rate; represents the recall rate; Indicates correct evaluation index; represents the precision of the wrong prediction errors.

2. The method for intelligently identifying personnel status based on independent visual units according to claim 1, characterized in that: The following steps are also included: S5: Upload the result information to the visual monitoring system for display.

3. The method for intelligently identifying personnel status based on independent visual units according to claim 1, characterized in that: The infrared sensor is a passive infrared device group. After a person enters the infrared detection area, the infrared device group delivers the detection signal to the power network. When there is no one around, the camera module and the lightweight neural network deployment module are in a closed state. When someone comes, the infrared ray is first used for identification to obtain the signal of whether there is anyone, turn on the power, and the camera module and the lightweight neural network deployment module start working.

4. A personnel status intelligent recognition system based on an independent visual unit, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 3 is implemented.

5. A recognition system implemented by the method for intelligently recognizing personnel status based on an independent visual unit according to any one of claims 1 to 3, comprising a housing, characterized in that: Also included is an integrated circuit board disposed in the housing for identifying the wearing status of the helmet, and a visual monitoring system for displaying the results processed by the integrated circuit board; The integrated circuit board adopts a double-layer circuit board architecture, including a first layer board and a second layer board; the first layer board is provided with a camera module, a lightweight neural network deployment module, and a shared power supply; the second layer board is provided with a buzzer, an alarm light, and an infrared sensor; The common power supply is respectively connected to the camera module, the lightweight neural network deployment module, the alarm module, and the infrared sensor to provide power; The camera module inputs the collected data into the lightweight neural network deployment module, and the lightweight neural network deployment module sends the processed signal to the alarm module; The infrared sensor is connected to the lightweight neural network deployment module, and the infrared sensor is used to detect whether there is a target object in the monitoring area.

Images