Tipping paper production safety monitoring method and system

Abstract

The invention relates to the field of production safety management, in particular to a tipping paper production safety monitoring method and system, and the method comprises the steps: calculating the abnormal score of a person behavior in each frame of real-time image; and generating an abnormal behavior judgment result of the personnel according to the abnormal score, and for the personnel judged to be abnormal, obtaining identity information of the personnel so as to supervise safety behaviors of all personnel in the production place. The method and the device have the effect of improving the smoking behavior recognition accuracy in the production process of personnel.

Description

Technical Field

[0001] The present application relates to the field of production safety management, and in particular to a method and system for monitoring the production safety of tipping paper. Background Art

[0002] Flammability. Stacking finished or semi-finished products is an unavoidable step in the production of tipping paper. During this stacking process, tipping paper is often stored densely, potentially with poor air circulation. Once exposed to open flames or hot objects (such as unextinguished cigarette butts), it can easily cause a fire. If production staff smoke, the open flame or hot butt generated by a lit cigarette can directly ignite the tipping paper. If smokers carelessly discard partially extinguished cigarette butts, the embers can slowly ignite nearby combustible materials. Tipping paper is a paper material primarily composed of plant fibers and is inherently highly flammable. Therefore, it is necessary to manage and supervise personnel's smoking behavior.

[0003] The technology used to identify smokers primarily relies on image recognition, leveraging deep learning algorithms to analyze surveillance video streams in real time. These algorithms can identify features such as smoke, the presence of a cigarette in a hand, and mouth movements to determine if someone is smoking. Once smoking is detected, the system automatically triggers an alarm, ensuring that relevant personnel are informed and can take timely action.

[0004] In the above-mentioned methods, cigarettes are relatively small in production site images, making them easily misidentified or missed in surveillance videos. This is especially true in long-distance monitoring scenarios, where the pixel ratio of cigarettes is extremely low, making it difficult for deep learning models to accurately identify them. Workers also frequently make hand movements, such as those involving holding a cigarette, which can easily lead to misidentification. Furthermore, smoke is relatively thin in its early stages, with unclear color and texture features, making it easily confused with the background. Complex production environment conditions (such as lighting and dust) can also interfere with smoke detection. In summary, existing technologies suffer from unclear features and low recognition accuracy. Summary of the Invention

[0005] In order to improve the accuracy of identifying smoking behavior of personnel during the production process, the present application provides a tipping paper production safety monitoring method and system.

[0006] In a first aspect, the present application provides a method for monitoring the safety of tipping paper production, which adopts the following technical solution: A tipping paper production safety monitoring method includes the following steps: calculating anomaly scores of personnel behavior in each frame of real-time image; generating abnormal behavior determination results of personnel based on the anomaly scores; obtaining identity information of personnel determined to be abnormal, so as to monitor the safety behavior of all personnel in the production site; The anomaly score is calculated as follows: a window is set and the motion features of the person in the real-time image within the window are calculated; a dynamic curve of time and motion features is constructed, and the abnormal behavior sequence in the dynamic curve is divided, with the dynamic feature with the minimum value within the window as the starting point of the abnormal behavior sequence, and the length of the abnormal sequence is a preset fixed length; the abnormal behavior segment is segmented according to the maximum and minimum values of the motion feature to obtain multiple subsequences; similarly, the preset standard abnormal behavior sequence is segmented to obtain multiple standard subsequences; if the length of the first subsequence is not equal to the length of the first standard subsequence, the first subsequence is compressed or the first standard subsequence is expanded to make the first subsequence and the first standard subsequence equal in length, and similarly, each subsequence is made equal in length to the corresponding standard subsequence; the similarity between the subsequence and the standard subsequence is calculated, and the cumulative mean of the similarities is used as the abnormality degree of the abnormal behavior segment; a first weight and a second weight are set, and the anomaly score is calculated as the product of the first weight and the abnormality degree and the second weight and the total number of abnormal behavior segments.

[0007] This application quantifies a person's behavior as an anomaly score. The higher the anomaly score, the greater the possibility of smoking behavior. By analyzing the changes in the distance between the key points of the fingers and the mouth, when a person's hands frequently and repeatedly approach and move away from the person's mouth, it may indicate that the person is smoking.

[0008] In smoking behavior, the minimum distance between the finger and the mouth key point may correspond to the moment the cigarette is brought to the mouth, while the maximum value may correspond to the moment the cigarette is removed. Using these key points as the basis for segmentation can more accurately extract the characteristic patterns of the behavior. The dynamic feature with the minimum value within the window is used as the starting point of the abnormal behavior sequence. This abnormal behavior sequence corresponds to the dynamic feature of the person removing the cigarette from the mouth. Compared with the dynamic feature of bringing the cigarette into the mouth, the dynamic feature of removing the cigarette from the mouth is easier to quantify, has less external interference, and provides accurate recognition results.

[0009] Optionally, the expression for the anomaly score is: Where, Indicates the The anomaly score of each person, Indicates the The degree of abnormality of an individual, represents the first hyperparameter, represents the second hyperparameter, Indicates the The total number of abnormal behavior segments corresponding to each person.

[0010] The abnormality score is quantified by combining the degree of abnormality and the number of repetitions of the abnormal behavior segment. Smoking is a repetitive behavior, such as repeatedly moving the hand toward and away from the mouth. By counting the number of repetitions of abnormal behavior segments, we can effectively distinguish between accidental face touching and smoking.

[0011] Optional, Where, Indicates the subsequences With the Standard subsequence The similarity of represents the Euclidean distance, represents the L2 norm.

[0012] It provides a method for quantifying similarity, which is suitable for situations where the absolute difference between sequences needs to be accurately measured. It can effectively identify abnormal behavior, but is more sensitive to noise.

[0013] Optional, Where, Indicates the subsequences With the Standard subsequence The similarity of express and Pearson correlation coefficient.

[0014] It provides another method for quantifying similarity, which is suitable for scenarios where it is necessary to measure the consistency of dynamic trends between sequences. It is robust to amplitude changes, but does not consider absolute differences and is insensitive to noise.

[0015] Optionally, the action feature is the Euclidean distance from a finger key point to a mouth key point or the Euclidean distance from a palm key point to a mouth key point.

[0016] When detecting smoking behavior, the change in the distance between the finger and the mouth is a key feature, as smoking is often accompanied by frequent movements of the hand towards and away from the mouth. By calculating the distance changes between these key points, the behavioral characteristics can be quantified and converted into data points on a dynamic curve.

[0017] Optionally, finger keypoints, palm keypoints, and mouth keypoints are obtained through OpenPose.

[0018] Optionally, the compression method is: dividing the subsequence into several segments evenly by segmented sampling, and extracting feature points from each segment.

[0019] Optionally, the expansion method is linear interpolation or spline interpolation.

[0020] Optionally, generating a person's abnormal behavior determination result based on the abnormality score includes: presetting a threshold of the abnormality score, and determining it as abnormal behavior if the abnormality score is greater than the threshold, otherwise, determining it as normal behavior.

[0021] In a second aspect, the present application provides a tipping paper production safety monitoring system, which adopts the following technical solutions: A tipping paper production safety monitoring system comprises: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the tipping paper production safety monitoring method described above is implemented.

[0022] The above-mentioned tipping paper production safety monitoring method is generated into a computer program and stored in a memory so as to be loaded and executed by a processor. Thus, a system is manufactured based on the memory and the processor for easy use.

[0023] This application has the following technical effects: 1. This application quantifies personnel behavior into an anomaly score. The higher the anomaly score, the greater the likelihood of smoking. By analyzing the distance changes between the key points of the finger and the mouth, when the person's hand frequently approaches and moves away from the mouth, it may indicate that the person is smoking. At the same time, it can distinguish normal production behavior from the person to reduce the possibility of misjudgment.

[0024] 2. During smoking behavior, the minimum distance between the finger and the mouth key point may correspond to the moment the cigarette is brought to the mouth, while the maximum value may correspond to the moment the cigarette is removed. Using these key points as the basis for segmentation can more accurately extract the characteristic patterns of the behavior. The dynamic feature with the minimum value within the window is used as the starting point of the abnormal behavior sequence. This abnormal behavior sequence corresponds to the dynamic feature of the person removing the cigarette from the mouth. Compared with the dynamic feature of bringing the cigarette into the mouth, the dynamic feature of removing the cigarette from the mouth is easier to quantify, has less external interference, and provides accurate recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a method flow chart of a method for safety monitoring of tipping paper production according to an embodiment of the present application.

[0026] Figure 2 This is a flow chart of a method for calculating anomaly scores in a method for monitoring tipping paper production safety in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The present application is further described in detail below with reference to the accompanying drawings.

[0028] The present application discloses a method for monitoring the production safety of tipping paper, which is used to detect the smoking behavior of people in the production site. Figure 1 , including steps S1 and S2, specifically as follows: S1: Calculate the abnormal score of human behavior in each frame of real-time image.

[0029] By analyzing collected videos and images, we can identify smoking behavior. To reduce computational complexity and improve processing efficiency, we can use frame extraction technology to preprocess the video. This extraction method can be based on fixed frame intervals (e.g., 10 frames per second) or fixed frame counts (e.g., 1 frame every 10 frames).

[0030] After extracting frames from the video, we generate multiple real-time images. We then use OpenPose to extract key points of the human body, including those on the fingers, palms, and mouth. OpenPose is an existing technology that can detect human postures in real time and provide key point coordinates.

[0031] This application analyzes the distance changes between a person's finger and key points of the mouth. When a person's hand frequently moves closer to and further away from the mouth, it may indicate that the person is smoking. This application quantifies the person's behavior into an anomaly score. A higher anomaly score indicates a greater likelihood of smoking. The probability of abnormal behavior (smoking) is determined by analyzing the anomaly score.

[0032] Reference Figure 2 The method for calculating the abnormal score of human behavior in the real-time image includes steps S10 to S13, which are as follows: S10: Set a window and calculate the motion features of the person in the real-time image within the window; construct a dynamic curve about time and motion features, divide the abnormal behavior sequence in the dynamic curve, and use the dynamic feature with the minimum value in the window as the starting point of the abnormal behavior sequence. The length of the abnormal sequence is a preset fixed length.

[0033] The user can set the window length based on the specific application scenario, such as 1 minute. This window divides continuous image frames into multiple segments, allowing key information to be extracted and analyzed within each window. Windows can define a time range, and multiple frames within a time range can reflect dynamic features, which better reflect the essence of behavior than static features in a single frame.

[0034] Extreme values (such as minimum values) of dynamic features often correspond to key moments in a behavior. For example, the minimum distance between the finger and the mouth during smoking may correspond to the moment when the cigarette is brought to the mouth. Using this minimum value as the starting point of an abnormal behavior sequence can more clearly pinpoint the starting point of the behavior and reduce false positives. The length of an abnormal sequence is fixed and can consist of action feature values within 10 seconds or 30 feature values. This is not detailed here.

[0035] In one embodiment, the motion feature can be the Euclidean distance from the finger key point to the mouth key point, or the Euclidean distance from the palm key point to the mouth key point, or other combinations of human key points that can reflect the change in the distance from the human hand to the face. These key points are selected based on their importance in abnormal behavior detection. For example, when detecting smoking behavior, the change in the distance between the finger and the mouth is a key feature because smoking is usually accompanied by frequent movements of the hand towards and away from the mouth. By calculating the change in distance between these key points, the behavioral characteristics can be quantified and converted into data points on a dynamic curve.

[0036] During the act of smoking, the distance between the key points of the two processes—bringing the cigarette to the mouth and removing it—changes in opposite but similar patterns. Specifically, during the process of approaching the mouth, the distance between the key points of the finger and the mouth gradually decreases until the hand brings the cigarette to the mouth. During the process of moving away from the mouth, the distance between the key points of the finger and the mouth gradually increases until the hand removes the cigarette from the mouth. These two processes exhibit symmetry in the distance change: the distance decreases as the hand approaches and increases as the hand moves away.

[0037] To reduce computational complexity, this application decomposes the complete smoking action into two phases. The first phase (approaching the mouth) starts from the moment the distance between the finger and the key point of the mouth decreases to the moment the distance reaches its minimum. The second phase (moving away from the mouth) starts from the moment the distance between the finger and the key point of the mouth increases to the moment the distance returns to or is close to the initial state.

[0038] In real-world surveillance scenarios, the first segment of motion (near the mouth) is likely to be subject to more interference from external factors, such as other hand movements or occlusions. In contrast, the second segment of motion is typically clearer and less likely to be subject to interference, so this application analyzes the second segment of motion to improve detection reliability.

[0039] Furthermore, the starting point of smoking in production environments is difficult to quantify. The moment the distance between the finger and the mouth key point begins to decrease may not be the start of smoking, but rather the noise generated by other production activities. These behaviors may have similarities with smoking in the changes in key point distance.

[0040] Furthermore, the starting point of the first movement can vary significantly due to individual differences, movement speed, and environmental factors, while the starting point of the second movement usually has clearer characteristics and is easier to identify on the dynamic curve. The distance between the key points in this stage changes more smoothly and regularly.

[0041] S11: Segment the abnormal behavior segment according to the maximum and minimum values of the action features to obtain multiple subsequences; similarly, segment the preset standard abnormal behavior sequence to obtain multiple standard subsequences; if the length of the first subsequence is not equal to the length of the first standard subsequence, compress the first subsequence or expand the first standard subsequence to make the first subsequence and the first standard subsequence equal in length. Similarly, make each subsequence equal in length to the corresponding standard subsequence.

[0042] In time series analysis, abnormal behavior often manifests as subsequences that differ significantly from normal behavior patterns. By segmenting detected abnormal behavior segments based on extreme values of motion features, key behavioral changes can be more accurately captured. For example, in smoking behavior, the minimum distance between the finger and the mouth key point may correspond to the moment the cigarette is brought to the mouth, while the maximum value may correspond to the moment the cigarette is removed. Using these key points as the basis for segmentation can more accurately extract the characteristic patterns of the behavior, rather than simply decomposing it into equal-length subsequences. This approach is more consistent with the dynamic changes in behavioral characteristics. Compared to equal-length subsequences, this segmentation method reduces the problem of truncation or splicing of behavioral features caused by fixed-length segmentation. For example, in smoking behavior, the lengths of movements near and away from the mouth may be different, thereby improving the accuracy of anomaly detection.

[0043] Comparing the detected subsequence with the preset standard subsequence can provide a clear reference template, thereby reducing misjudgments caused by individual differences or different movement speeds.

[0044] In one embodiment, the compression method is: the subsequence is evenly divided into several segments by segmented sampling, and feature points are extracted from each segment. For example, assuming that the length of the subsequence is 100 time points and the length of the standard subsequence is 50 time points, the subsequence can be divided into 10 segments, each segment contains 10 time points; key feature points are extracted from each segment. For example, the mean, median or extreme point of the segment can be selected as the feature point. If the mean is selected, the 10 time points of each segment will be compressed into a representative value. The feature points of all segments are combined into a new subsequence. Taking the mean as an example, the mean of 10 segments will generate a subsequence of length 10, which is aligned with the standard subsequence of length 50.

[0045] In one embodiment, the expansion method is a linear interpolation method or a spline interpolation method. Both the linear interpolation method and the spline interpolation method are existing technologies and will not be described in detail here.

[0046] S12: Calculate the similarity between the subsequence and the standard subsequence, and use the accumulated mean of the similarities as the abnormality degree of the abnormal behavior segment.

[0047] In one embodiment, the similarity calculation formula is: Where, Indicates the subsequences With the Standard subsequence The similarity of represents the Euclidean distance, represents the L2 norm.

[0048] The calculation of Euclidean distance and L2 norm are both existing technologies and will not be described in detail here. measure and The smaller the distance, the more similar the two sequences are. Standardization is performed to eliminate the deviation caused by the different amplitudes of the eigenvalues. The value range is between 0 and 1. and The more similar, The closer it is to 0, the closer the similarity is. The closer the value of is to 1, the smaller the similarity is. The closer the value of is to 0. This embodiment is applicable to situations where the absolute difference between sequences needs to be accurately measured and can effectively identify abnormal behaviors, but is more sensitive to noise.

[0049] In one embodiment, the similarity calculation formula is: Where, Indicates the subsequences With the Standard subsequence The similarity of express and The calculation of the Pearson correlation coefficient is a prior art and will not be described in detail here. This method measures the linear correlation between two sequences and can reflect whether their dynamic trends are consistent. This embodiment is suitable for scenarios where dynamic trend consistency between sequences needs to be measured. It is robust to amplitude changes, but does not consider absolute differences and is insensitive to noise. In practical applications, the appropriate similarity calculation method can be selected based on specific needs, or the advantages of both methods can be combined to improve the accuracy and robustness of abnormal behavior detection.

[0050] The expression of abnormality degree is: Where, Indicates the The degree of abnormality of an individual, Indicates the subsequences With the Standard subsequence similarity; Indicates the total number of subsequences or the total number of standard subsequences. The total number of subsequences is equal to the total number of standard subsequences.

[0051] The similarity of all subsequences to the standard subsequence Accumulate and merge multiple local similarity information into a global similarity metric by dividing the accumulated similarity by the total number of subsequences N , and get the average similarity , even if some subsequences are affected by noise, resulting in similarity Low, and will not affect the overall abnormality Have too much impact.

[0052] S13: Setting a first weight and a second weight, and calculating the product of the first weight and the abnormality degree and the sum of the second weight and the total number of abnormal behavior segments as the abnormality score.

[0053] In one embodiment, the expression for the anomaly score is: Where, Indicates the The anomaly score of each person, Indicates the The degree of abnormality of an individual, represents the first hyperparameter, represents the second hyperparameter, Indicates the The total number of abnormal behavior segments corresponding to each person; represents the hyperbolic tangent function.

[0054] Abnormality The distance between the finger and the key point of the mouth can be analyzed for calculation. However, judging smoking behavior by distance change alone will result in large errors. For example, touching the face with the hand may also be considered as smoking behavior. Therefore, this application combines the number of repetitions of abnormal behavior segments, because smoking actions are repetitive, such as the hand repeatedly approaching and moving away from the mouth. By counting the number of repetitions of abnormal behavior segments, the number of repetitions of abnormal behavior segments is calculated. , which can effectively distinguish between accidental facial touching and smoking behavior.

[0055] represents the first weight, The first and second weights are dynamically adjusted based on the number of abnormal behavior segments. The greater the total number of abnormal behavior segments, the smaller the first weight and the larger the second weight. This indicates that more abnormal behavior segments occur, indicating that the person repeatedly brings their hand close to their mouth, and the greater the likelihood that they are smoking. Therefore, when a person repeatedly brings their hand close to their mouth, the more abnormal behavior segments appear, and the more the anomaly score tends to reflect the frequency of the behavior.

[0056] The more abnormal behavior segments appear, the The closer the value of is to 1, the less The value of is closer to 0. along with The increase gradually tends to 1, indicating that the more times the abnormal behavior segment appears, the greater the possibility that the person is smoking, and it can effectively capture the cumulative effect of repeated smoking behavior of the person.

[0057] represents the first hyperparameter, represents the second hyperparameter, the first hyperparameter Controls the adjustment rate of the first and second weights. Second hyperparameter control , thereby affecting the contribution of the number of abnormal behavior segments to the anomaly score.

[0058] By quantifying human behavior into anomaly scores, the degree of abnormality of human behavior can be intuitively assessed. The closer the value is to 1, the more similar the person's behavior is to the standard abnormal behavior, and the higher the possibility of smoking. Conversely, the lower the possibility of smoking.

[0059] Exemplarily, the value of the first hyperparameter is 2, and the value of the second hyperparameter is 0.5. The values of the first hyperparameter and the second hyperparameter can be set according to actual application scenarios.

[0060] In one embodiment, the user may also set a fixed first weight and a fixed second weight according to a specific application scenario, which will not be described in detail here.

[0061] S2: Generate the abnormal behavior judgment result of the personnel based on the abnormality score. For the personnel judged to be abnormal, obtain the personnel's identity information to monitor the safety behavior of all personnel in the production site.

[0062] In one embodiment, when the abnormality score of any person is greater than a threshold (e.g., 0.5), the person is determined to be smoking; the person's identity information is obtained through facial recognition, and the abnormal behavior is recorded in a preset safety management system to simultaneously monitor and manage the safety behavior of all personnel in the production site.

[0063] An embodiment of the present application further discloses a tipping paper production safety monitoring system, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the tipping paper production safety monitoring method according to the present application is implemented.

[0064] The above system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. The configuration and functions of these components are known in the art and will not be described in detail here.

[0065] In this application, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, the computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), etc.

[0066] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for monitoring the production safety of tipping paper, characterized in that: The method includes the following steps: calculating anomaly scores of personnel behaviors in each frame of real-time image; generating abnormal behavior determination results of personnel based on the anomaly scores; obtaining identity information of personnel determined to be abnormal, so as to monitor the safety behaviors of all personnel in the production site; The anomaly score is calculated as follows: a window is set and the motion features of people in the real-time image within the window are calculated; a dynamic curve of time and motion features is constructed and the abnormal behavior sequence in the dynamic curve is divided. The dynamic feature with the minimum value within the window is used as the starting point of the abnormal behavior sequence, and the length of the abnormal sequence is a preset fixed length; the abnormal behavior segment is segmented according to the maximum and minimum values of the motion features to obtain multiple subsequences; similarly, the preset standard abnormal behavior sequence is segmented to obtain multiple standard subsequences; If the length of the first subsequence is not equal to that of the first standard subsequence, compress the first subsequence or expand the first standard subsequence to make the first subsequence equal to the first standard subsequence. Similarly, make each subsequence equal to the corresponding standard subsequence. Calculate the similarity between the subsequence and the standard subsequence, and use the cumulative mean of the similarities as the abnormality degree of the abnormal behavior segment; A first weight and a second weight are set, and the product of the first weight and the abnormality degree and the second weight and the total number of abnormal behavior segments are calculated as the abnormality score.

2. The tipping paper production safety monitoring method according to claim 1, characterized in that: The expression of anomaly score is: Where, Indicates the The anomaly score of each person, Indicates the The degree of abnormality of an individual, represents the first hyperparameter, represents the second hyperparameter, Indicates the The total number of abnormal behavior segments corresponding to each person; represents the hyperbolic tangent function.

3. The tipping paper production safety monitoring method according to claim 1, characterized in that: The formula for calculating similarity is: Where, Indicates the subsequences With the Standard subsequence The similarity of represents the Euclidean distance, represents the L2 norm.

4. The tipping paper production safety monitoring method according to claim 1, characterized in that: The formula for calculating similarity is: Where, Indicates the subsequences With the Standard subsequence The similarity of express and Pearson correlation coefficient.

5. The tipping paper production safety monitoring method according to claim 1, characterized in that: The motion feature is the Euclidean distance from the finger key point to the mouth key point or the Euclidean distance from the palm key point to the mouth key point.

6. The tipping paper production safety monitoring method according to claim 5, characterized in that: Finger key points, palm key points and mouth key points are obtained through OpenPose.

7. The tipping paper production safety monitoring method according to claim 1, characterized in that: The compression method is to evenly divide the subsequence into several segments through segmented sampling and extract feature points from each segment.

8. The tipping paper production safety monitoring method according to claim 1, characterized in that: The expansion method is linear interpolation or spline interpolation.

9. The tipping paper production safety monitoring method according to claim 1, characterized in that: The abnormal behavior determination result of the person is generated according to the abnormality score, including: a preset abnormality score threshold, and in response to the abnormality score being greater than the threshold, it is determined to be abnormal behavior, otherwise it is determined to be normal behavior.

10. A tipping paper production safety monitoring system, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the method for monitoring the safety of production of tipping paper according to any one of claims 1 to 9 is implemented.

Images