Chemical behavior safety analysis system based on spatial scene modeling and behavior safety

Abstract

The chemical behavior safety analysis system based on spatial scene modeling and behavior safety belongs to the technical field of safety analysis system. It includes a multi-eye space level division module, a behavior recognition online learning module and an immediate danger response system module. The multi-eye space level division module performs spatial modeling of the three-dimensional production space through a plurality of multi-eye cameras and divides the spatial depth information; the behavior recognition online learning module obtains the image of a specific area through the camera, and then recognizes and judges the image; the immediate danger response system module combines the output of the multi-eye space level division module and the behavior recognition online learning module, and sets a variety of different levels of immediate response strategies in a targeted manner. The system of the present invention can improve the recognition rate of safe behaviors, and at the same time can make reasonable immediate response measures, greatly reduce the probability of accidents in chemical production, improve the safety factor, and ensure the effective implementation of chemical production.

Description

Technical Field

[0001] The present invention belongs to the technical field of safety analysis systems, and in particular relates to a chemical industry behavior safety analysis system based on spatial scene modeling and behavior safety. Background Art

[0002] There are many potential safety hazards in the chemical industry. As the main unsafe factor, personnel bring many potential dangers. The existing solution generally adopts the method of manual monitoring. The camera in the construction scene is monitored by human eyes. The inspection is carried out according to manual experience. The disadvantages of the above solution are as follows:

[0003] 1. Uncertainty in surveillance video image acquisition: The image acquisition and imaging of video will change with the changes in the environment. For some specific dangerous behaviors, the imaging will interfere with the human eye.

[0004] 2. High redundancy of manual investigation: For each suspected dangerous behavior, manual investigation is required, which is inefficient, resulting in slow construction progress and affecting production;

[0005] 3. Response delay problem of existing solutions: Once unsafe behavior occurs, it needs to be manually checked before taking response measures. This solution will directly lead to serious safety accidents. Summary of the invention

[0006] In view of the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide a chemical behavior safety analysis system based on spatial scene modeling and behavioral safety, which can identify behaviors in production spaces with high precision and set immediate corresponding measures according to the levels of different hazardous areas.

[0007] The present invention provides the following technical solution: A chemical behavior safety analysis system based on spatial scene modeling and behavior safety, characterized by: comprising a multi-view spatial level division module, a behavior recognition online learning module and an immediate danger response system module;

[0008] The multi-camera space level division module performs spatial modeling on the three-dimensional production space through a plurality of pairs of multi-cameras to divide the spatial depth information;

[0009] The behavior recognition online learning module obtains images of a specific area through a camera, and then recognizes and judges the images;

[0010] The instant danger response system module combines the outputs of the multi-view spatial classification module and the behavior recognition online learning module to specifically set up multiple instant response strategies at different levels.

[0011] The chemical behavior safety analysis system based on spatial scene modeling and behavior safety is characterized in that the multi-eye space level division module adopts a multi-eye space modeling method to perform space modeling, four pairs of binocular cameras are established in the modeling space, and the four pairs of binocular cameras are used to perform depth modeling of the modeling space in four directions respectively. The specific modeling process is as follows:

[0012] 1.1. Checkerboard calibration of binocular cameras: Use the checkerboard to calibrate the four pairs of binocular cameras so that the presentation of the four pairs of binocular cameras remains consistent in the world coordinate system. Then, capture several groups of images of the production space, and take two images of the same position in different states as a group.

[0013] 1.2. Obtain a disparity map between two images in each group of images captured in step 1.1 through a stereo matching algorithm, and then obtain a depth map through the disparity map;

[0014] 1.3. According to the depth map obtained in step 1.2, the production space is divided into three spaces of A, B, and C with decreasing danger levels according to different depth ranges, and a danger level space model is established, wherein the depth refers to the depth value from the space point to the center of the space, and the depth range of the three danger space levels A, B, and C is determined according to the actual safety level requirements on site.

[0015] The chemical behavior safety analysis system based on spatial scene modeling and behavior safety is characterized in that the behavior recognition online learning module uses a method based on deep network learning to judge the images in the production space input into the module.

[0016] The chemical behavior safety analysis system based on spatial scene modeling and behavior safety is characterized in that the operation process of the behavior recognition online learning module is as follows:

[0017] 2.1. The camera obtains images of specific areas in the production space;

[0018] 2.2. Import the image obtained in step 2.1 into the trained deep network learning model for judgment, output the judgment result of non-dangerous behavior or non-dangerous behavior, and input the output result into the immediate danger response system module.

[0019] The chemical behavior safety analysis system based on spatial scene modeling and behavior safety is characterized in that the training process of the deep network learning model is as follows:

[0020] 2.2.1. Image preprocessing: Collect training sample images for edge detection to obtain edge features of sample images, and then normalize the sample image pixels to 0-255;

[0021] 2.2.2. Network model construction: Fine-tune training is performed on the vgg16 network model, keeping the parameters of the first three layers of the vgg16 network model unchanged, the network uses four-channel input, the number of categories in the last layer of the network is set to n+1, and the other network structure parameters remain unchanged;

[0022] 2.2.3. Network model training: Set the training framework and learning rate of the network model built in step 2.2.2, use the acquired RGB images as the training set to train the network model, and establish the judgment criteria for dangerous behavior or non-dangerous behavior based on the training set sample data.

[0023] The chemical behavior safety analysis system based on spatial scene modeling and behavior safety is characterized in that an online learning scheme is also introduced in the behavior recognition online learning module to deal with the problem of newly added categories, and the online learning scheme includes the following steps:

[0024] 3.1. Collect and calibrate sample images of intruding objects in the production space as training sets;

[0025] 3.2. Import the sample image obtained in step 2.1 into the trained network model for main category judgment to update the network model. When performing the main category judgment, if one of the output probabilities of the feature map of the last layer of the network model exceeds x%, the current sample image is added to the training set of the corresponding label. If the output probabilities of the feature map of the last layer of the network model are all lower than (100-x)%, the current sample image is regarded as a new category. If the output probability of the feature map of the last layer of the network model is between (100-x)% and x%, the current sample image is discarded as an interference sample, where x is determined according to actual needs.

[0026] 3.3. Data set difference judgment: compare the difference between the training set sample images and the network model updated in step 2.2. If the recognition rate difference between the training set sample images and the network model updated in step 2.2 exceeds 20%, start online learning.

[0027] The chemical behavior safety analysis system based on spatial scene modeling and behavior safety is characterized in that the response process of the immediate danger response system module is as follows:

[0028] When the behavior recognition online learning module outputs a dangerous behavior, confirm which danger level space the dangerous behavior is in. When the dangerous behavior is in the C space, give a monitoring reminder response. When the dangerous behavior is in the B space, give an alarm response. When the dangerous behavior is in the A space, terminate the previous production operation. When the online learning module outputs a non-dangerous behavior, the current production operation runs normally.

[0029] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1) The present invention proposes an intelligent identification and grading warning system based on chemical safety behavior recognition, which can improve the recognition rate of safety behaviors and make reasonable and immediate response measures, greatly reducing the probability of accidents in chemical production, improving the safety factor, and ensuring the effective implementation of chemical production;

[0031] 2) The present invention can accurately control the safety measures of the entire production space through the multi-view space level division module, avoiding erroneous responses caused by misoperation;

[0032] 3) The present invention can improve the accuracy of behavior recognition by setting up an online learning module for behavior recognition, and can perform online learning based on real-time category updates to improve the recognition robustness of the model;

[0033] 4) The instant danger response system module set up in the present invention can combine the outputs of the multi-view spatial classification module and the behavior recognition online learning module to specifically set up multiple instant response strategies at different levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a schematic diagram of the division of hazardous levels of production space;

[0035] Figure 2 It is a flow chart of the online learning solution of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments of the specification. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0037] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0038] See also Figure 1-2 , a chemical behavior safety analysis system based on spatial scene modeling and behavioral safety, including a multi-view spatial classification module, a behavior recognition online learning module and an immediate hazard response system module.

[0039] The multi-eye spatial level division module uses several pairs of multi-eye cameras to perform spatial modeling of the three-dimensional production space and divide the spatial depth information.

[0040] Specifically, the multi-eye space level division module adopts a multi-eye space modeling method, and uses four pairs of binocular cameras to perform deep modeling of the three-dimensional production space from four directions (excluding the upper and lower directions), and divides the production space into three dangerous spaces: A, B, and C. The levels of the three dangerous spaces A, B, and C are determined according to the actual safety level requirements, among which the danger levels of the three spaces A, B, and C decrease in sequence.

[0041] The process of producing spatial depth modeling is as follows:

[0042] 1.1. Checkerboard calibration of binocular cameras: Use the checkerboard to calibrate the four pairs of binocular cameras so that the presentation of the four pairs of binocular cameras remains consistent in the world coordinate system. Then, capture images of the production space and capture several groups of images. Take two images of the same position in different states as a group.

[0043] 1.2. Obtain a disparity map between two images in each group of images captured in step 1.1 through a stereo matching algorithm, and then obtain a depth map through the disparity map;

[0044] 1.3. Based on the depth map obtained in step 1.2, the production space is divided into three spaces with decreasing danger levels, A, B, and C, according to different depth ranges.

[0045] The behavior recognition online learning module obtains images of specific areas in the production space through cameras, uses a deep network learning method to judge the images of specific areas input into the module, and introduces an online learning solution to deal with the problem of newly added categories.

[0046] Specifically, the operation process of the behavior recognition online learning module is as follows:

[0047] 2.1. The camera obtains images of specific areas in the production space;

[0048] 2.2. Import the image obtained in step 2.1 into the trained deep network learning model for judgment and output the judgment result. If the judgment result is non-dangerous behavior, the production process in the production space runs normally. If the judgment result is dangerous behavior, the multi-eye spatial classification module will classify the production space into spatial danger levels.

[0049] Specifically, the training process of the deep network learning model is as follows:

[0050] 2.2.1. Image preprocessing: Collect training sample images and use the Canny edge detection algorithm to perform edge detection to obtain edge features of the sample images. Then normalize the sample image pixels to 0-255, perform sample image scale transformation to obtain a 256*256 RGB image. The label range corresponding to the obtained RGB image is 0 to N, where the Nth sample is an unknown sample and is temporarily set as a pure background image in the present invention.

[0051] 2.2.2. Network model construction: Fine-tune training is performed on the vgg16 network model, keeping the parameters of the first three layers of the vgg16 network model unchanged, changing the network input from the previous 3 channels to 4 channels, setting the number of categories in the last layer of the network to n+1, and keeping the other network structure parameters unchanged;

[0052] 2.2.3. Network model training: Set the training framework and learning rate of the network model built in step 2.2.2. The training framework is caffe, the learning rate is set to 0.001, and the network model is trained using the acquired RGB images.

[0053] Specifically, online learning solutions include Figure 2 As shown, the multiple “image captures” in the figure represent the same real space object images taken by different cameras. The specific process is as follows:

[0054] 3.1. Collect and calibrate sample images of intruding objects in the production space as training links. Assume that the number of samples is N.

[0055] 3.2. Import the sample image obtained in step 2.1 into the trained network model for main category judgment to update the network model. When judging the main category, if one of the output probabilities of the feature map of the last layer of the network model exceeds 70%, the current sample is added to the training set of the corresponding label. If the output probabilities of the feature map of the last layer of the network model are all lower than 30%, the current sample is regarded as a new category. If the output probability of the feature map of the last layer of the network model is between 30% and 70%, the current sample is discarded as an interference sample.

[0056] 3.3. Dataset difference judgment: Compare the difference between the training set sample images and the network model updated in step 2.2. Only when the difference between the training set sample images and the network model updated in step 2.2 exceeds 20%, start online learning to avoid redundant training.

[0057] Specifically, the response process of the immediate danger response system module is as follows:

[0058] When the behavior recognition online learning module outputs a dangerous behavior, it confirms which danger level space the dangerous behavior is in. When the dangerous behavior is in the C space, a monitoring reminder response is given. When the dangerous behavior is in the B space, an alarm response is given. When the dangerous behavior is in the A space, the previous production operation is terminated. When the online learning module outputs a non-dangerous behavior, the current production operation runs normally.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A chemical behavior safety analysis system based on spatial scene modeling and behavior safety, characterized by: It includes multi-view space classification module, behavior recognition online learning module and immediate danger response system module; The multi-camera space level division module performs spatial modeling on the three-dimensional production space through a plurality of pairs of multi-cameras to divide the spatial depth information; The behavior recognition online learning module obtains images of a specific area through a camera, and then recognizes and judges the images; The instant danger response system module combines the outputs of the multi-view space classification module and the behavior recognition online learning module to set up a variety of instant response strategies at different levels; The multi-eye space level division module adopts a multi-eye space modeling method to perform space modeling. Four pairs of binocular cameras are established in the modeling space. The four pairs of binocular cameras are used to perform depth modeling of the modeling space in four directions. The specific modeling process is as follows: 1.

1. Checkerboard calibration of binocular cameras: Use the checkerboard to calibrate the four pairs of binocular cameras so that the presentation of the four pairs of binocular cameras remains consistent in the world coordinate system. Then, capture several groups of images of the production space, and take two images of the same position in different states as a group. 1.

2. Obtain a disparity map between two images in each group of images captured in step 1.1 through a stereo matching algorithm, and then obtain a depth map through the disparity map; 1.

3. According to the depth map obtained in step 1.2, the production space is divided into three spaces of A, B, and C with decreasing danger levels according to different depth ranges, and a danger level space model is established, wherein the depth refers to the depth value from the space point to the center of the space, and the depth range of the three danger space levels A, B, and C is determined according to the actual safety level requirements on site.

2. The chemical behavior safety analysis system based on spatial scene modeling and behavior safety according to claim 1 is characterized in that The behavior recognition online learning module uses a method based on deep network learning to judge the images in the production space input into the module.

3. The chemical behavior safety analysis system based on spatial scene modeling and behavior safety according to claim 2 is characterized in that The operation process of the behavior recognition online learning module is as follows: 2.

1. The camera obtains images of specific areas in the production space; 2.

2. Import the image obtained in step 2.1 into the trained deep network learning model for judgment, output the judgment result of non-dangerous behavior or non-dangerous behavior, and input the output result into the immediate danger response system module.

4. The chemical behavior safety analysis system based on spatial scene modeling and behavior safety according to claim 3 is characterized in that The training process of the deep network learning model is as follows: 2.2.

1. Image preprocessing: Collect training sample images for edge detection to obtain edge features of sample images, and then normalize the sample image pixels to 0-255; 2.2.

2. Network model construction: Fine-tune training is performed on the vgg16 network model, keeping the parameters of the first three layers of the vgg16 network model unchanged, the network uses four-channel input, the number of categories in the last layer of the network is set to n+1, and the other network structure parameters remain unchanged; 2.2.

3. Network model training: Set the training framework and learning rate of the network model built in step 2.2.2, use the acquired RGB images as the training set to train the network model, and establish the judgment criteria for dangerous behavior or non-dangerous behavior based on the training set sample data.

5. The chemical behavior safety analysis system based on spatial scene modeling and behavior safety according to claim 4 is characterized in that The behavior recognition online learning module also introduces an online learning solution to deal with the problem of adding new categories. The online learning solution includes the following steps: 3.

1. Collect and calibrate sample images of intruding objects in the production space as training sets; 3.

2. Import the sample image obtained in step 2.1 into the trained network model for main category judgment to update the network model. When making the main category judgment, if one of the output probabilities of the feature map of the last layer of the network model exceeds x%, the current sample image is added to the training set of the corresponding label. If the output probabilities of the feature map of the last layer of the network model are all lower than (100-x)%, the current sample image is regarded as a new category. If the output probability of the feature map of the last layer of the network model is between (100-x)% and x%, the current sample image is discarded as an interference sample, where x is determined according to actual needs. 3.

3. Data set difference judgment: compare the difference between the training set sample images and the network model updated in step 2.

2. If the recognition rate difference between the training set sample images and the network model updated in step 2.2 exceeds 20%, start online learning.

6. The chemical behavior safety analysis system based on spatial scene modeling and behavior safety according to claim 4 is characterized in that The response process of the immediate danger response system module is as follows: When the behavior recognition online learning module outputs a dangerous behavior, confirm which dangerous level space the dangerous behavior is in. When the dangerous behavior is in the C space, give a monitoring reminder response. When the dangerous behavior is in the B space, give an alarm response. When the dangerous behavior is in the A space, terminate the previous production operation. When the online learning module output is identified as a non-dangerous behavior, the current production operation is running normally.

Images