A fire safety risk analysis method and system for power sites

Abstract

The present invention relates to a method and system for analyzing fire safety risks in power places, the method comprising: collecting video images, image noise reduction processing, dynamic target detection, when the presence of dynamic targets is detected, the image with dynamic targets is input into a convolutional neural network semantic segmentation model, the dynamic targets and risk targets are segmented, the Euclidean distance between the pixel centroid of the dynamic target and the pixel centroid of the risk target is calculated, the topological structure of the dynamic targets is analyzed, and the fire safety risk level and alarm level are defined. The system comprises: an image acquisition module, a dynamic target detection module, a processing module and an alarm module.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a method and system for analyzing fire safety risks in power sites. Background Art

[0002] As an important place for energy supply, the quality and safety of energy supply in power plants are directly related to the quality of people's lives and work. Only when the continuous and reliable supply of electricity is guaranteed can the energy needed for people's lives be provided. However, power plants need to face many safety hazards, among which fire safety issues are the most prominent. Therefore, power plants need to control risk issues and reduce the occurrence rate of disasters through reasonable fire safety management.

[0003] After a comprehensive analysis of the safety accidents that occurred in my country's power plants, it was found that some of the accidents were caused by operators' illegal operations. In power companies, the management's professional literacy is not up to standard, so they cannot provide safety production training to operators and other grassroots personnel, resulting in low professional quality of employees, causing them to operate illegally in daily production processes, making it difficult to effectively carry out fire safety work, and laying hidden safety hazards for the future.

[0004] The traditional method is to install cameras for monitoring and have personnel perform remote monitoring in the monitoring room. This method requires a lot of manpower and financial resources and is easily interfered by other reasons.

[0005] The invention patent with patent publication number CN113763203A discloses a smart safety monitoring system for substations and a safety management method for on-site operations: a smart safety monitoring system for substations and a safety management method that integrates heterogeneous data of multi-source safety status are proposed, in which multiple sensors are used to increase costs, and risk objects are not classified.

[0006] The invention patent with patent publication number CN113486856A discloses a method for detecting irregular driver behavior based on semantic segmentation and convolutional neural network: it proposes to use semantic segmentation network and convolutional neural network to classify irregular behavior, and its algorithm has high complexity.

[0007] In terms of dynamic detection, the traditional method is to set several rectangular invisible detection areas in the video display area. When an abnormal situation occurs at the scene of video surveillance, the brightness, contrast and content of the image in the detection area (that is, the amplitude of the video signal changes). When this change exceeds the preset threshold (sensitivity), an alarm is generated. Or the video signal is directly sampled and compared with the normal data. When the comparison result exceeds the threshold, an alarm is generated. There are also dynamic detection technologies based on moving targets, such as the difference method. The disadvantage of this method is that the image subtraction obtains a partial outline or a thickened outline of the moving target, so the extracted moving target is distorted, and the effect is not ideal in the complex monitoring environment, especially in the real situation where the lighting conditions change during the day.

[0008] In terms of dangerous behavior recognition: the existing behavior recognition algorithm based on deep learning can automatically learn behavioral features, but does not take into account the direct motion information of continuous video frames. In terms of dangerous posture estimation: the existing method builds a personal posture library and makes a pre-judgment alarm through a series of posture and action recognition, but the amount of calculation is huge, the hardware equipment requirements are high, it is difficult to achieve real-time detection, and there is a lack of recognition and understanding of the environment in which people are located, resulting in low accuracy. In terms of safety scene understanding: the existing method classifies the input video image through a deep learning framework, and classifies and recognizes the human body and its environment through a trained image data set, showing the relationship between the human body and the surrounding environment, and setting a fall level alarm according to the environmental hazard level, but the training method is relatively cumbersome, lacks a fast and accurate image segmentation algorithm, and it is difficult to accurately extract the human body and the surrounding environment. Summary of the invention

[0009] In view of the shortcomings of existing fire risk warning methods for power sites, the present application provides a fire safety risk analysis method and system for power sites.

[0010] This application is implemented through the following technical solutions:

[0011] A method for analyzing fire safety risks in power places comprises: acquiring video images, performing image noise reduction processing, and performing dynamic target detection. When a dynamic target is detected, the image with the dynamic target is input into a convolutional neural network semantic segmentation model to segment the dynamic target and the risk target, calculate the Euclidean distance between the pixel centroid of the dynamic target and the pixel centroid of the risk target, perform topological structure analysis on the moving target, and define the fire safety risk level and alarm level.

[0012] Furthermore, the dynamic target detection method is: calculating the grayscale centroid of the dynamic target image, the quasi-template centroid offset and the quasi-template image grayscale change, and then determining whether there is a dynamic target.

[0013] Furthermore, the method for determining whether a dynamic target exists comprises the following steps:

[0014] Before detecting the appearance or movement of a moving target, obtain a quasi-template image by recording the grayscale changes and grayscale centroids of several consecutive frames of images;

[0015] By comparing the difference between the basic images of the previous and next frames, it is determined whether it is a quasi-template. If the conditions are met, the next frame image is determined to be a quasi-template.

[0016] The moving points in the video are detected by using the difference algorithm between the current frame and the quasi-template frame. The number of moving points and the grayscale centroid offset between the current frame image and the quasi-template frame image are used to determine whether there is a dynamic target in the current frame.

[0017] Preferably, for the detection image sequence (x, y, t), the quasi-template determination condition is:

[0018]

[0019] Where f(x,y,t1) and f(x,y,t2) are the grayscale value functions of the image sequence at time t1 and t2 respectively, ΔC is the grayscale centroid offset, T is the judgment threshold, is the illumination sensitivity term, a is the suppression coefficient, and N is the number of image sequences.

[0020] Furthermore, the grayscale difference of each point in the frame image is calculated by using a differential algorithm, and then the non-moving points and moving points are judged by threshold comparison. The judgment formula is:

[0021]

[0022] Where f(x,y,t) is the grayscale value of the image pixel at time t, F(x,y) is the grayscale value of the quasi-template image pixel, and T d is the motion threshold. If the M value is 1, the pixel is a motion point. If the M value is 0, the pixel is a non-motion point.

[0023] Preferably, continuous multi-frame sequence images and quasi-template frames are used for difference respectively, and the image grayscale of continuous multi-frame images in the sequence images and the absolute difference image of the quasi-template frame are calculated to obtain the number of moving points, and then the difference results are superimposed, and the grayscale centroid offset ΔC of each frame in the continuous multi-frame sequence images and the quasi-template frame is introduced to judge whether there is a dynamic target in the sequence image. The judgment formula is:

[0024]

[0025] Where n is the number of consecutive multi-frame images, N is the number of image pixels, ε is the offset coefficient, and T D is the judgment threshold, if MD Value ≥ T D value, there are dynamic targets in the image, if M D Value<T D Value, there is no dynamic object in the image.

[0026] Furthermore, the method for defining the fire safety risk level and the alarm level is to calculate the change in the Euclidean distance between the centroid of the dynamic target pixel and the centroid of the risk target pixel obtained by image segmentation and the change in the topological structure of the dynamic target pixel;

[0027] Optionally, a dynamic target pixel centroid position offset is also included.

[0028] Preferably, the calculation formula of the pixel centroid is:

[0029]

[0030]

[0031]

[0032] Where x0, y0 are the centroid coordinates, f(x,y) is the pixel value of the grayscale image at (x,y), if the pixel is a dynamic target or a risk target after semantic segmentation, the value of f(x,y) is 1, if the pixel is a non-dynamic target or a non-risk target after semantic segmentation, the value of f(x,y) is 0.

[0033] Furthermore, the fire safety risk level is represented by a risk index b, and the calculation formula of the risk index b is:

[0034]

[0035] or,

[0036]

[0037] Where α is the Euclidean distance coefficient, β is the topological structure transformation coefficient, T is the topological structure state quantity, and ΔD is the centroid position offset of the dynamic target pixel.

[0038] The present invention also provides a fire safety risk analysis system for power sites, comprising:

[0039] Image acquisition module, used for acquiring video images and image noise reduction processing;

[0040] A data transmission module, used for data transmission;

[0041] A dynamic target detection module, used for dynamic target detection;

[0042] The processing module is used to segment dynamic targets and risky objects, calculate the Euclidean distance between the pixel centroids of the dynamic targets and the pixel centroids of the risky objects, perform topological structure analysis on the moving targets, and define the fire safety risk level and alarm level;

[0043] The alarm module is used to issue an alarm according to the alarm level.

[0044] Compared with the prior art, this application has the following beneficial effects:

[0045] 1. The method and system of the present invention can quickly and accurately detect dynamic targets;

[0046] 2. Classify the risk levels of dynamic targets and conduct risk warning more effectively;

[0047] 3. Use semantic segmentation model to segment dynamic targets, analyze risk level through centroid distance and topological structure, which can improve accuracy and real-time performance;

[0048] 4. For different security risk objects and special areas, each video channel can independently set the dynamic detection area and sensitivity, which can be used in different areas and at different times;

[0049] 5. Semantic segmentation is performed only after the dynamic target is identified by multiple frame grayscale changes, which can greatly reduce the computing time and resource consumption of the device;

[0050] 6. Use continuous multi-frame sequence images and quasi-template frames to perform differences respectively, calculate the image grayscale of continuous multi-frame images in the image sequence and the absolute difference image of the quasi-template frame, then superimpose the difference results and introduce the grayscale centroid offset, which can further improve the reliability of dynamic target detection;

[0051] 7. The accuracy of behavior analysis is improved by combining the pixel centroid of dynamic targets and the pixel centroid of safety risk objects with the topological structure analysis of dynamic targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the embodiments of the present application, constitute a part of the present application, and do not constitute a limitation on the embodiments of the present invention.

[0053] Figure 1 It is a flowchart of the method of the present invention;

[0054] Figure 2 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and beneficial effects of the present application clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0056] The present invention discloses a method for analyzing fire safety risks in power places based on dynamic targets, comprising: acquiring video images, performing image noise reduction processing, and performing dynamic target detection. When the presence of a dynamic target is detected, the image with the dynamic target is input into a convolutional neural network semantic segmentation model to segment the dynamic target and the risk target, calculate the Euclidean distance between the pixel centroid of the dynamic target and the pixel centroid of the risk target, perform a topological structure analysis on the moving target, and define the fire safety risk level and the alarm level.

[0057] Among them, dynamic targets refer to moving targets that appear in power places; risk objects refer to fire protection objects in power places, such as meters, lightning arresters, switches, current and voltage transformers, insulators, distribution rooms, main transformers, main transformer terminal boxes, intelligent control cabinets, fire sand boxes, temperature and humidity sensors and busbars, etc.

[0058] Step 1. Use a camera to capture a stable video image.

[0059] Step 2: Process the video image obtained in step 1 to eliminate noise interference factors and obtain processed image information.

[0060] Step 3. Calculate the grayscale centroid of the image information obtained in step 2, determine it as a quasi-template and save it to the memory for future use.

[0061] The grayscale centroid calculation method of the image is:

[0062] Define the moment m of image block B pq for:

[0063]

[0064] Where x and y correspond to the horizontal and vertical coordinates of the pixel of image block B respectively, I(x,y) is the grayscale value of the corresponding pixel at (x,y), p corresponds to the order in the x direction, and the p value is 0 or 1, and q corresponds to the order in the y direction, and the q value is 0 or 1.

[0065] The centroid C of the image block B can be found by moment:

[0066]

[0067] In the formula, m 10 is the moment of image block B in the x direction, m01 is the moment of image block B in the y direction, m 00 is the image grayscale centroid and .

[0068] Before detecting motion (target appearance or movement), the quasi-template image is obtained by recording the grayscale changes and grayscale centroid of several consecutive frames of images for moving target detection when the moving target appears. First, the difference between the basic images of the previous and next frames is compared to determine whether it is a quasi-template. If the conditions are met, the next frame of the image is determined as a quasi-template and saved to the memory for standby. For the detection image sequence (x, y, t), the quasi-template determination conditions are:

[0069]

[0070] Where f(x,y,t1) and f(x,y,t2) are the grayscale value functions of the image sequence at time t1 and t2 respectively, ΔC is the grayscale centroid offset, T is the judgment threshold, is the illumination sensitive item, a is the suppression coefficient, and N is the number of image sequences. T can be set according to the on-site conditions. The influence of changes in illumination conditions is taken into account in the judgment conditions, so it has a certain adaptability to illumination changes.

[0071] Step 4. Calculate the image grayscale centroid.

[0072] Step 5. Calculate the moving pixels.

[0073] The motion points in the video are detected by using the difference algorithm between the current motion frame and the quasi-template frame. The main idea of ​​the algorithm is to calculate the video difference (grayscale difference) of each pixel in the frame image, and then determine which are non-motion points and which are motion points by setting a motion threshold. The judgment formula is:

[0074]

[0075] Where f(x,y,t) is the grayscale value of the image pixel at time t, F(x,y) is the grayscale value of the quasi-template image pixel, and T d is the motion threshold. If the absolute difference between the grayscale values ​​of the corresponding pixels in the two frames and the grayscale centroid offset are greater than the threshold, it indicates that the pixel is a moving point and the M value is 1; otherwise, it is a non-moving point and the M value is 0.

[0076] Step 6. Calculate the grayscale centroid offset between the current motion frame image and the quasi-template frame image, and then determine whether there is a dynamic target in the current motion frame.

[0077] This patent uses continuous multi-frame sequence images and quasi-template frames to perform differences respectively, calculates the image grayscale of continuous multi-frame images in the sequence image and the absolute difference image of the quasi-template frame, obtains the number of moving points, and then superimposes the difference results, and introduces the grayscale centroid offset ΔC of each frame in the continuous multi-frame sequence image and the quasi-template frame, so as to judge whether there is a dynamic target in the sequence image. The judgment formula is:

[0078]

[0079] Where n is the number of consecutive multi-frame images, N is the number of image pixels, ε is the offset coefficient, and T D is the judgment threshold. D Value ≥ T D value, there are dynamic targets in the image; if M D Value<T D Value, there is no dynamic object in the image.

[0080] Step 7. Send the image with dynamic targets into the network improved based on deeplabv3 for semantic segmentation, and segment the dynamic targets and security risk objects. The image segmentation network improved based on deeplabv3 has achieved very good segmentation results on the UCF101 and HMDB51 datasets.

[0081] Step 8: Calculate the pixel centroid of the segmented dynamic target and the pixel centroid of the segmented risk target.

[0082]

[0083]

[0084]

[0085] Where x0 and y0 are the centroid coordinates, f(x,y) is the pixel value of the grayscale image at (x,y). If the pixel is a dynamic target or a risk target after semantic segmentation, the value of f(x,y) is 1. If the pixel is a non-dynamic target or a non-risk target after semantic segmentation, the value of f(x,y) is 0.

[0086] Step 9. Calculate the topological structure state quantity T of the dynamic target segmented from several consecutive frames;

[0087] Step 10. Calculate the risk level by combining the values ​​of step 8 and step 9;

[0088] Step 11. Alert.

[0089] After detecting the presence of dynamic targets in the image, the semantic segmentation model is used to segment the dynamic targets and risk targets. DeepLabV3 is used to semantically segment the dynamic targets and risk targets, and then the pixel centroid is calculated. The Euclidean distance between the pixel centroid of the dynamic target and the pixel centroid of the risk target is calculated to determine the motion trend of the dynamic target, and the fire safety risk level is calculated by combining the topological structure changes of the dynamic target in consecutive frames.

[0090] The calculation formula of the pixel centroid Euclidean distance d is:

[0091]

[0092] In the formula, (x d ,y d ) is the pixel centroid coordinate of the risk object, (x f ,y f ) is the pixel centroid coordinate of the dynamic target.

[0093] The calculation method of the topological structure state quantity T of the dynamic target vertex and the dynamic target centroid is:

[0094] The vertex and centroid angle set {α n}Recorded as A:

[0095] A=[α1,α2,...,α n ]

[0096] The distance set between the vertex and the centroid {l n} is denoted as L:

[0097] L=[l1,l2,...,l n ]

[0098]

[0099] The fire safety risk level is expressed by the risk index b, and the calculation formula of the risk index b is:

[0100]

[0101] or,

[0102]

[0103] Where α is the Euclidean distance coefficient, β is the topological structure transformation coefficient, T is the topological structure state quantity, and ΔD is the centroid position offset of the dynamic target pixel.

[0104] ① Based on the change in the Euclidean distance between the pixel centroid of the dynamic target segmented by the image and the pixel centroid of the risk target, as well as the topological structure of the target in the image segmentation model, each identified object has a risk coefficient judgment, and the risk level will be determined according to the risk coefficient.

[0105] Specifically, according to the various specific situations of the segmented dynamic target topological structure changes and risky objects, the various specific situations are: if there is no moving object in the image, it is judged as a safe environment; if there is a dynamic target in the image, and the risk level determined by the topological analysis and the Euclidean distance change is low, it is judged as a first-level warning; if there is a dynamic target in the image, and the risk level determined by the topological analysis and the Euclidean distance change is medium, it is judged as a second-level warning; if there is a dynamic target in the image, and the risk level determined by the topological analysis and the Euclidean distance change is high, it is judged as a third-level warning.

[0106] ② The multi-frame image segmentation can also be used to obtain the centroid position offset of the dynamic target pixel, the change in the Euclidean distance between the multi-frame segmented dynamic target pixel centroid and the centroid of the risk target, and the change in the dynamic target pixel topological structure. The risk factor of each identified moving object can be determined, and the risk level can be determined according to the risk factor.

[0107] Specifically, according to the segmented topological structure changes of dynamic targets and various specific situations of risky objects, the various specific situations are: if there is no moving object in the image, it is judged as a safe environment; if there is a dynamic target in the image, and the topological structure analysis and the centroid position offset determine that the target is not close to the risky object and has no dangerous behavior, the risk level is low, then it is judged as a first-level warning; if there is a dynamic target in the image, and the topological structure analysis and the centroid position offset determine that the target is close to the risky object or has dangerous behavior, the risk level is medium, then it is judged as a second-level warning; if there is a dynamic target in the image, the topological structure analysis and the centroid position offset determine that the target is close to the risky object and has dangerous behavior, the risk level is high, then it is judged as a third-level warning.

[0108] For example, fire safety risk levels can be divided into:

[0109]

[0110] The present invention also discloses a fire safety risk analysis system for power sites, including:

[0111] Image acquisition module, used for acquiring video images and image noise reduction processing;

[0112] A data transmission module, used for data transmission;

[0113] A dynamic target detection module, used for dynamic target detection;

[0114] The processing module is used to segment dynamic targets and risky objects, calculate the Euclidean distance between the pixel centroids of the dynamic targets and the pixel centroids of the risky objects, perform topological structure analysis on the moving targets, and define the fire safety risk level and alarm level;

[0115] The alarm module is used to issue an alarm according to the alarm level.

[0116] The above specific implementation methods further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fire safety risk analysis method for power sites, comprising: Collect video images, perform image noise reduction, and perform dynamic target detection. The feature is that: when a dynamic target is detected, the image with the dynamic target is input into a convolutional neural network semantic segmentation model to segment the dynamic target and the risk target, calculate the Euclidean distance between the pixel centroid of the dynamic target and the pixel centroid of the risk target, perform topological structure analysis on the moving target, and define the fire safety risk level and alarm level; The fire safety risk level is expressed by the risk index b, and the calculation formula of the risk index b is: Where n is the number of consecutive multi-frame images, α is the Euclidean distance coefficient, β is the topological structure transformation coefficient, T is the topological structure state quantity, and d is the Euclidean distance of the pixel centroid; Among them, the calculation method of the topological structure state quantity T of the dynamic target vertex and the dynamic target centroid is: The vertex and centroid angle set {α n }Recorded as A: A＝[α1,α2,...,α n ] The distance set between the vertex and the centroid {l n } is denoted as L: L＝[l1,l2,...,l n ] 2. The method for analyzing fire safety risks in power places according to claim 1 is characterized by: The dynamic target detection method comprises: calculating the grayscale centroid of the dynamic target image, the offset of the quasi-template centroid and the grayscale variation of the quasi-template image, and then judging whether there is a dynamic target.

3. The method for analyzing fire safety risks in power places according to claim 2 is characterized by: The method for determining whether a dynamic target exists comprises the following steps: Before detecting the appearance or movement of a moving target, obtain a quasi-template image by recording the grayscale changes and grayscale centroids of several consecutive frames of images; By comparing the difference between the basic images of the previous and next frames, it is determined whether it is a quasi-template. If the conditions are met, the next frame image is determined to be a quasi-template. The moving points in the video are detected by using the difference algorithm between the current frame and the quasi-template frame. The number of moving points and the grayscale centroid offset between the current frame image and the quasi-template frame image are used to determine whether there is a dynamic target in the current frame.

4. The method for analyzing fire safety risks in power places according to claim 3 is characterized by: For the detection image sequence (x, y, t), the quasi-template determination condition is: Where f(x,y,t1) and f(x,y,t2) are the grayscale value functions of the image sequence at time t1 and t2 respectively, ΔC is the grayscale centroid offset, T is the judgment threshold, is the illumination sensitivity term, a is the suppression coefficient, and N is the number of image sequences.

5. The method for analyzing fire safety risks in power places according to claim 4 is characterized by: The difference algorithm is used to calculate the grayscale difference of each pixel in the current frame image, and then the non-moving points and moving points are judged by comparing with the motion threshold. The judgment formula is: Where f(x,y,t) is the grayscale value of the image pixel at time t, F(x,y) is the grayscale value of the quasi-template image pixel, and T d is the motion threshold. If the M value is 1, the pixel is a motion point. If the M value is 0, the pixel is a non-motion point.

6. The method for analyzing fire safety risks in power places according to claim 5 is characterized by: The continuous multi-frame sequence images and the quasi-template frame are used for difference respectively, and the image grayscale of the continuous multi-frame images in the sequence images and the absolute difference image of the quasi-template frame are calculated to obtain the number of moving points. Then the difference results are superimposed, and the grayscale centroid offset ΔC of each frame in the continuous multi-frame sequence images and the quasi-template frame is introduced to determine whether there is a dynamic target in the sequence image. The judgment formula is: Where n is the number of consecutive multi-frame images, N is the number of image pixels, ε is the offset coefficient, and T D is the judgment threshold, if M D Value ≥ T D value, there are dynamic targets in the image, if M D Value<T D Value, there is no dynamic object in the image.

7. The method for analyzing fire safety risks in power places according to claim 1 is characterized by: The calculation formula of the pixel centroid is: Where x0 and y0 are the centroid coordinates, f(x,y) is the pixel value of the grayscale image at (x,y). If the pixel is a dynamic target or a risk target after semantic segmentation, the value of f(x,y) is 1. If the pixel is a non-dynamic target or a non-risk target after semantic segmentation, the value of f(x,y) is 0.

8. The method for analyzing fire safety risks in power places according to claim 7 is characterized by: The fire safety risk level is represented by a risk index b, and the calculation formula of the risk index b can also be: Where n is the number of consecutive multi-frame images, α is the Euclidean distance coefficient, β is the topological structure transformation coefficient, T is the topological structure state quantity, ΔD is the centroid position offset of the dynamic target pixel, and d is the Euclidean distance of the pixel centroid.

9. A power site fire safety risk analysis system for implementing the power site fire safety risk analysis method according to any one of claims 1 to 8, characterized in that: include: Image acquisition module, used for acquiring video images and image noise reduction processing; A data transmission module, used for data transmission; A dynamic target detection module, used for dynamic target detection; The processing module is used to segment dynamic targets and risky objects, calculate the Euclidean distance between the pixel centroids of the dynamic targets and the pixel centroids of the risky objects, perform topological structure analysis on the moving targets, and define the fire safety risk level and alarm level; The alarm module is used to issue an alarm according to the alarm level.

Images