A pumping unit non-stop personnel approach detection method based on deep learning

Abstract

The present application belongs to the technical field of pumping unit non-stop target detection and target comparison, and particularly relates to a detection method for pumping unit non-stop personnel approach based on deep learning. The detection method for pumping unit non-stop personnel approach based on deep learning is used for real-time detection of safety hazards of personnel approach injury or crush injury when a pumping unit is not stopped (the pumping unit is in motion or stops motion without pulling the hand brake), and can not only meet the management requirements of oilfield enterprises, but also timely discover safety hazards of technical personnel and issue early warnings to avoid accidents. The detection method for pumping unit non-stop personnel approach based on deep learning comprises the steps of constructing a YOLOv5 network model structure, training a neural network model, establishing the association between the pumping unit state and the pumping unit image to be detected, judging whether the detection target result of the pumping unit image to be detected exists, whether the pumping unit is in a motion state, whether there is personnel, and calculating the intersection and union ratio of the personnel target and the fence target.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas production safety risk target detection technology, and in particular relates to a deep learning-based method for detecting personnel approaching an oil pumping unit that has not been shut down. Background Technology

[0002] To ensure the safety of operations at the well site, technicians need to monitor the pumping units for any instances of them not braking in real time. In particular, if someone enters the pumping unit enclosure or concrete area, timely detection of the potential hazard of personnel entering while the pumping unit is running can effectively protect the lives and property of personnel and prevent accidents from occurring.

[0003] With the rapid development of deep learning technology, object detection has become an important branch of computer vision and has been widely used in many fields such as security, industry, and autonomous driving. Among them, the patent document titled "A Method for Detecting the Start-up and Stoppage of a Video Pumping Unit" (application number: CN202011531759.4) describes a method for detecting the start-up and stoppage of a video pumping unit. This method mainly includes: first, extracting frames from a given video of a pumping unit to prepare for target detection of the pumping unit in the images; labeling the dataset of the pumping unit to train a Yolov5 target detection model; using the trained Yolov5 model to detect the pumping unit in the images, obtaining the pumping unit's bounding box, including the upper left corner (x1, y1) and lower right corner (x2, y2); in the current frame, cropping out the bounding box region of the pumping unit as image 2; locating the image T seconds before or after the current frame, cropping out the corresponding region in that frame as image 1; using a sliding window to take M×N windows on the two images, calculating the similarity between each pair of corresponding windows, and averaging to obtain the total similarity between the two images.

[0004] However, further research by technicians revealed that existing technologies, including the aforementioned patent documents, still present numerous technical challenges. For example, the detection process still relies on manually defined operating procedures. Furthermore, determining whether the pumping unit is operating during this process presents a significant challenge. Additionally, if the operating speed of the pumping unit and the interval between video frame cuts are not properly matched, it may result in the pumping unit remaining stationary during video detection (a false impression of shutdown). This introduces new uncertainties into the detection results and hinders further improvements in detection efficiency. Summary of the Invention

[0005] This invention provides a deep learning-based method for detecting personnel approaching an unstopped oil pumping unit. This method is used to detect in real time the safety hazards of personnel approaching an unstopped oil pumping unit (when the pumping unit is in motion or stopped and the handbrake is not engaged), which may result in injury from being hit or squeezed. It not only meets the management requirements of oilfield enterprises, but also can promptly detect the safety hazards of personnel intrusion and issue early warnings to avoid accidents.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A deep learning-based method for detecting personnel approaching an oil pumping unit that has not been shut down includes the following steps:

[0008] Step S1: Obtain the monitoring video of the pumping unit operation at the work site, and use the group frame extraction method to extract images containing the pumping unit from the monitoring video;

[0009] Step S2: Annotate the oil pumping unit image obtained in step S1 and generate sample data;

[0010] Step S3: Construct the YOLOv5 network model structure; train the YOLOv5 network model using the sample data obtained in step S2 to obtain a trained neural network model;

[0011] Step S4: Acquire real-time monitoring video of the work site, and acquire multiple consecutive frames of images of the pumping unit under test at intervals as a group of test units;

[0012] Step S5: Obtain the pumping unit status from the image of the pumping unit under test, and establish the association between the pumping unit status and the image of the pumping unit under test;

[0013] Step S6: Input the image of the pumping unit to be tested into the trained neural network model obtained in step S3 to obtain the detection target result of the image of the pumping unit to be tested;

[0014] Step S7: Determine whether an oil pumping unit exists in the detection target result of the image of the oil pumping unit to be tested;

[0015] If an oil pump is present, edge detection is used to extract the shape of the oil pump; the shape of the oil pump is compared in three consecutive frames of the oil pump image to be tested; if the shapes of the oil pump in the three consecutive frames are different, it is determined that the oil pump is in motion; otherwise, it is determined that the oil pump is in a non-motion state.

[0016] Step S8: When the pumping unit is not in operation, determine whether the handbrake has been engaged based on the angle between the handbrake and the pumping unit;

[0017] When the pumping unit is in a non-moving state and the handbrake is not engaged, or when the pumping unit is in a moving state, the trained neural network model is used to traverse the detection target results of the detected image of the pumping unit to be tested, and to determine whether there are personnel present; if there are personnel, the coordinates of the personnel are compared with the coordinates of the pumping unit.

[0018] Step S9: Calculate the intersection-union ratio (IUGR) of the personnel target and the fence target in the same frame of the image to be tested; if the IUGR of the personnel target and the fence target are both greater than the preset value, it is determined that the personnel entered the working area of ​​the pumping unit while the pumping unit was not stopped.

[0019] More preferably, step S2 can be specifically described as follows: labeling the pumping unit, personnel, and fence in the pumping unit image obtained in step S1; using 75% of the labeled sample data as training samples and 25% of the labeled sample data as test samples.

[0020] More preferably, in step S4, the number of frames of the pumping unit image to be tested in the continuously acquired multi-frame images of the pumping unit to be tested is not less than 45 frames.

[0021] More preferably, the edge detection process in step S7 can be specifically described as follows:

[0022] The image of the oil pumping unit under test containing the oil pumping unit is converted to grayscale.

[0023] Gaussian filtering is applied to the image of the oil pumping unit under test that contains an oil pumping unit.

[0024] The gradient intensity and direction of each pixel in the image of the oil pumping unit under test, where an oil pumping unit is present, are calculated.

[0025] Non-maximum suppression processing is performed on each pixel in the image of the oil pumping unit under test that contains the pumping unit;

[0026] Double threshold detection is performed on the remaining pixels after nonmaximum suppression processing.

[0027] Further preferred, the grayscale processing can be specifically described as: Gray = 0.299R + 0.587G + 0.114B; where R, G, and B are the color values ​​of the image of the oil pumping unit under test containing the oil pumping unit.

[0028] A further preferred embodiment of the Gaussian filtering process can be specifically described as follows:

[0029] The image of the oil pumping unit under test is convolved with a Gaussian filter and the image of the oil pumping unit containing the oil pumping unit is used.

[0030] Discrete Gaussian convolution kernel H: (2k+1)×(2k+1) dimensional, its calculation method satisfies: .

[0031] A further preferred method for calculating the gradient intensity and direction of each pixel in the image of the pumping unit to be tested, where the pumping unit is present, can be specifically described as follows:

[0032] .

[0033] More preferably, the preset value for the intersection-union ratio of the personnel target and the fence target in step S9 is 0.3.

[0034] This invention provides a deep learning-based method for detecting personnel approaching an unshutted oil pumping unit. The method includes steps such as constructing a YOLOv5 network model, training the neural network model, establishing a correlation between the pumping unit's state and an image of the pumping unit under test, determining whether the pumping unit exists in the detection target results of the image, determining whether the pumping unit is in motion, determining whether personnel are present within the pumping unit, and calculating the intersection-union ratio (IUU) of the personnel target and the fence target. This deep learning-based method, with the above-described steps, can be used in real-time to detect safety hazards such as personnel approaching an unshutted pumping unit (where the handbrake is not engaged when the pumping unit is in motion or stopped), and preventing injuries from falling or being crushed. It not only meets the management requirements of oilfield enterprises but also promptly identifies safety hazards for technical personnel and issues early warnings, preventing accidents. Attached Figure Description

[0035] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the following drawings:

[0036] Figure 1 A flowchart illustrating the deep learning-based detection method for detecting personnel approaching an oil pumping unit that has not been shut down, as provided by this invention.

[0037] Figure 2 One of the schematic diagrams of sample data provided by the present invention;

[0038] Figure 3 This is the second schematic diagram of the sample data provided by the present invention. Detailed Implementation

[0039] This invention provides a deep learning-based method for detecting personnel approaching an unstopped oil pumping unit. This method is used to detect in real time the safety hazards of personnel approaching an unstopped oil pumping unit (when the pumping unit is in motion or stopped and the handbrake is not engaged), which may result in injury from being hit or squeezed. It not only meets the management requirements of oilfield enterprises, but also can promptly detect the safety hazards of personnel intrusion and issue early warnings to avoid accidents.

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0041] Example 1

[0042] This invention provides a deep learning-based method for detecting personnel approaching an oil pumping unit that has not been shut down, such as... Figure 1 As shown, it includes the following steps:

[0043] Step S1: Obtain the monitoring video of the pumping unit operation at the work site, and use the group frame extraction method to extract images containing the pumping unit from the monitoring video;

[0044] Specifically, monitoring videos of pumping unit operations at the well site can be obtained through monitoring equipment configured at the pumping unit well site; alternatively, monitoring video data of pumping unit operations stored in the central control center can be referenced.

[0045] Step S2: Annotate the oil pumping unit image obtained in step S1 and generate sample data;

[0046] In a preferred embodiment of the present invention, the annotation of the pumping unit image obtained in step S1 can be further described as follows: the three types of objects in the pumping unit image obtained in step S1, namely the pumping unit, personnel, and fence, are labeled separately. In the generated sample data, 75% is selected as training samples, and the remaining 25% is used as test samples.

[0047] Step S3: Construct the YOLOv5 network model structure; train the YOLOv5 network model using the sample data obtained in step S2 to obtain a trained neural network model;

[0048] It should be noted that when constructing the YOLOv5 network model structure, sampling feature maps at magnification ratios of 8x, 16x, and 32x are preferred. For specific examples of these sampling feature maps, please refer to [link / reference needed]. Figure 2 , Figure 3During the training of the YOLOv5 network model, based on the preferred sample data classification method in step S2, 75% of the sample data is used as training samples and the remaining 25% is used as test samples for training, thus obtaining a trained neural network model.

[0049] Step S4: Acquire real-time monitoring video of the work site, and acquire multiple consecutive frames of images of the pumping unit under test at intervals as a group of test units;

[0050] In a preferred embodiment of the present invention, the number of frames of the oil pumping unit image to be tested in the continuous multi-frame image of the oil pumping unit to be tested acquired at intervals in step S4 is preferably not less than 45 frames.

[0051] Step S5: Obtain the pumping unit status from the image of the pumping unit under test, and establish the association between the pumping unit status and the image of the pumping unit under test;

[0052] Step S6: Input the image of the pumping unit to be tested into the trained neural network model obtained in step S3 to obtain the detection target result of the image of the pumping unit to be tested;

[0053] Step S7: Determine whether there is a pumping unit in the detection target result of the image of the pumping unit to be tested; if there is a pumping unit, use edge detection to extract the shape of the pumping unit; compare the shape of the pumping unit in three consecutive frames of the image of the pumping unit to be tested; if the shape of the pumping unit in the three consecutive frames is different, it is determined that the pumping unit is in motion.

[0054] Step S8: When the pumping unit is not in operation, determine whether the handbrake has been engaged based on the angle between the handbrake and the pumping unit;

[0055] When the pumping unit is in a non-moving state and the handbrake is not engaged, or when the pumping unit is in a moving state, the trained neural network model is used to traverse the detection target results of the detected image of the pumping unit to be tested, and to determine whether there are personnel present; if there are personnel, the coordinates of the personnel are compared with the coordinates of the pumping unit.

[0056] It should be further noted that, undeniably, there is a safety hazard if personnel enter the pumping unit while it is in operation; similarly, if the pumping unit is not in operation and its handbrake is not engaged, there is also a safety hazard if personnel enter the unit (i.e., the pumping unit may restart). Therefore, steps S7 and S8 record both of these pumping unit scenarios.

[0057] Step S9: Calculate the intersection-union ratio (IUGR) of the personnel target and the fence target in the same frame of the image to be tested; if the IUGR of the personnel target and the fence target are both greater than the preset value, it is determined that the personnel entered the working area of ​​the pumping unit while the pumping unit was not stopped.

[0058] It should be added that the preset value of the intersection-union ratio of the personnel target and the fence target in step S9 above is preferably 0.3. In addition, it should be noted that the monitoring equipment mentioned in the above example preferably refers to monitoring equipment installed on utility poles or streetlights in the construction site, which can acquire images around the oil pumping unit during construction operations; the horizontal distance between the monitoring equipment and the location of the oil pumping unit is controlled within 100 meters.

[0059] At this point, the testing of the group to be tested has been completed.

[0060] Example 2

[0061] Example 2 includes all the technical features of Example 1; specifically, this invention provides a deep learning-based method for detecting the approach of personnel when an oil pumping unit is not shut down, such as... Figure 1 As shown, it includes the following steps:

[0062] Step S1: Obtain the monitoring video of the pumping unit operation at the work site, and use the group frame extraction method to extract images containing the pumping unit from the monitoring video;

[0063] Specifically, monitoring videos of pumping unit operations at the well site can be obtained through monitoring equipment configured at the pumping unit well site; alternatively, monitoring video data of pumping unit operations stored in the central control center can be referenced.

[0064] Step S2: Annotate the oil pumping unit image obtained in step S1 and generate sample data;

[0065] In a preferred embodiment of the present invention, the annotation of the pumping unit image obtained in step S1 can be further described as follows: the three types of objects in the pumping unit image obtained in step S1, namely the pumping unit, personnel, and fence, are labeled separately. In the generated sample data, 75% is selected as training samples, and the remaining 25% is used as test samples.

[0066] Step S3: Construct the YOLOv5 network model structure; train the YOLOv5 network model using the sample data obtained in step S2 to obtain a trained neural network model;

[0067] It should be noted that when constructing the YOLOv5 network model structure, sampling feature maps at magnification ratios of 8x, 16x, and 32x are preferred. For specific examples of these sampling feature maps, please refer to [link / reference needed]. Figure 2 , Figure 3 During the training of the YOLOv5 network model, based on the preferred sample data classification method in step S2, 75% of the sample data is used as training samples and the remaining 25% is used as test samples for training, thus obtaining a trained neural network model.

[0068] Step S4: Acquire real-time monitoring video of the work site, and acquire multiple consecutive frames of images of the pumping unit under test at intervals as a group of test units;

[0069] In a preferred embodiment of the present invention, the number of frames of the oil pumping unit image to be tested in the continuous multi-frame image of the oil pumping unit to be tested acquired at intervals in step S4 is preferably not less than 45 frames.

[0070] Step S5: Obtain the pumping unit status from the image of the pumping unit under test, and establish the association between the pumping unit status and the image of the pumping unit under test;

[0071] Step S6: Input the image of the pumping unit to be tested into the trained neural network model obtained in step S3 to obtain the detection target result of the image of the pumping unit to be tested;

[0072] Step S7: Determine whether an oil pumping unit exists in the detection target result of the image of the oil pumping unit to be tested;

[0073] If an oil pump is present, edge detection is used to extract the shape of the oil pump; the shape of the oil pump is compared in three consecutive frames of the oil pump image to be tested; if the shapes of the oil pump in the three consecutive frames are different, it is determined that the oil pump is in motion; otherwise, it is determined that the oil pump is in a non-motion state.

[0074] Step S8: When the pumping unit is not in operation, determine whether the handbrake has been engaged based on the angle between the handbrake and the pumping unit;

[0075] When the pumping unit is in a non-moving state and the handbrake is not engaged, or when the pumping unit is in a moving state, the trained neural network model is used to traverse the detection target results of the detected image of the pumping unit to be tested, and to determine whether there are personnel present; if there are personnel, the coordinates of the personnel are compared with the coordinates of the pumping unit.

[0076] Step S9: Calculate the intersection-union ratio (IUGR) of the personnel target and the fence target in the same frame of the image to be tested; if the IUGR of the personnel target and the fence target are both greater than the preset value, it is determined that the personnel entered the working area of ​​the pumping unit while the pumping unit was not stopped.

[0077] It should be added that the preset value of the intersection-union ratio of the personnel target and the fence target in step S9 above is preferably 0.3. In addition, it should be noted that the monitoring equipment mentioned in the above example preferably refers to monitoring equipment installed on utility poles or streetlights in the construction site, which can acquire images around the oil pumping unit during construction operations; the horizontal distance between the monitoring equipment and the location of the oil pumping unit is controlled within 100 meters.

[0078] At this point, the testing of the group to be tested has been completed.

[0079] Furthermore, Example 2 further describes the edge detection process described in step S7 as follows:

[0080] Specifically, the first step is to perform grayscale processing on the image of the oil pumping unit to be tested, which contains the pumping unit.

[0081] Then, Gaussian filtering was applied to the images of the pumping unit under test that contained the pumping unit.

[0082] It is worth noting that in order to minimize the impact of noise on the edge detection results, noise must be filtered out to prevent false detections caused by noise. Gaussian filtering is used in this step to smooth the image and reduce the significant noise impact on detection.

[0083] Then, the gradient intensity and direction of each pixel in the image of the oil pumping unit under test containing the oil pumping unit are calculated.

[0084] Then, non-maximum suppression processing is performed on each pixel in the image of the oil pumping unit under test that contains the oil pumping unit.

[0085] It should be noted that non-maximum suppression can help suppress all gradient values ​​outside of local maxima to zero. The specific basis for non-maximum suppression is as follows:

[0086] 1. Compare the gradient strength of the current pixel with that of two pixels along the positive and negative gradient directions;

[0087] 2. If the gradient strength of the current pixel is the largest compared to the other two pixels, then the pixel is retained as an edge point; otherwise, the pixel will be suppressed.

[0088] Additionally, double threshold detection is performed on the remaining pixels after non-maximum suppression processing.

[0089] After applying nonmaximum suppression, the remaining pixels can more accurately represent the actual edges in the image. However, in more extreme cases, some edge pixels still exist due to noise and color variations. To address these stray responses, edge pixels must be filtered using weak gradient values, retaining those with high gradient values; here, a high and low threshold approach is chosen. Furthermore, as a supplement, strong edge points are generally considered true edges; while weak edge points may be true edges, or they may be caused by noise or color variations. For the most accurate results, weak edge points caused by the latter should be removed. Typically, weak edge points caused by true edges are connected to strong edge points, while weak edge points caused by noise are not. Therefore, a hysteresis boundary tracking algorithm is used to check the 8-connected neighborhood pixels of a weak edge point; if a strong edge point truly exists, then this weak edge point is considered a true edge and retained.

[0090] Example 3

[0091] Embodiment 3 includes all the technical features of Embodiment 1, and due to space limitations, they will not be elaborated upon here.

[0092] Furthermore, the algorithm formulas used in this third embodiment to perform grayscale processing, Gaussian filtering processing, and to calculate the gradient intensity and direction of each pixel on the image of the oil pumping unit under test containing the oil pumping unit are described below.

[0093] The grayscale processing can be specifically described as follows: Gray = 0.299R + 0.587G + 0.114B; where R, G, and B are the color values ​​of the image of the oil pumping unit under test, which contains the oil pumping unit.

[0094] The process of Gaussian filtering can be specifically described as follows:

[0095] The image of the oil pumping unit under test is convolved with a Gaussian filter and the image of the oil pumping unit containing the oil pumping unit is used.

[0096] Discrete Gaussian convolution kernel H: (2k+1)×(2k+1) dimensional, its calculation method satisfies: .

[0097] The process of calculating the gradient intensity and direction of each pixel in an image of a pumping unit containing a pumping unit can be specifically described as follows:

[0098] .

[0099] Of course, the above implementation is for reference only and is not intended to further limit the present invention.

[0100] This invention provides a deep learning-based method for detecting personnel approaching an unshutted oil pumping unit. The method includes steps such as constructing a YOLOv5 network model, training the neural network model, establishing a correlation between the pumping unit's state and an image of the pumping unit under test, determining whether the pumping unit exists in the detection target results of the image, determining whether the pumping unit is in motion, determining whether personnel are present within the pumping unit, and calculating the intersection-union ratio (IUU) of the personnel target and the fence target. This deep learning-based method, with the above-described steps, can be used in real-time to detect safety hazards such as personnel approaching an unshutted pumping unit (where the handbrake is not engaged when the pumping unit is in motion or stopped), and preventing injuries from falling or being crushed. It not only meets the management requirements of oilfield enterprises but also promptly identifies safety hazards for technical personnel and issues early warnings, preventing accidents.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting personnel approaching an oil pumping unit while it is not shut down, based on deep learning, characterized in that, The steps include the following: Step S1: Obtain the monitoring video of the pumping unit operation at the work site, and use the group frame extraction method to extract images containing the pumping unit from the monitoring video; Step S2: Annotate the oil pumping unit image obtained in step S1 and generate sample data; Step S3: Construct the YOLOv5 network model structure; train the YOLOv5 network model using the sample data obtained in step S2 to obtain a trained neural network model; Step S4: Acquire real-time monitoring video of the work site, and acquire multiple consecutive frames of images of the pumping unit under test at intervals as a group of test units; Step S5: Obtain the pumping unit status from the image of the pumping unit under test, and establish the association between the pumping unit status and the image of the pumping unit under test; Step S6: Input the image of the pumping unit to be tested into the trained neural network model obtained in step S3 to obtain the detection target result of the image of the pumping unit to be tested; Step S7: Determine whether an oil pumping unit exists in the detection target result of the image of the oil pumping unit to be tested; If an oil pump is present, edge detection is used to extract the shape of the oil pump; Compare the morphology of the pumping unit in three consecutive frames of images of the pumping unit under test; If the shape of the pumping unit is different in three consecutive frames, the pumping unit is determined to be in motion; otherwise, the pumping unit is determined to be in a non-motion state. Step S8: When the pumping unit is not in operation, determine whether the handbrake has been engaged based on the angle between the handbrake and the pumping unit; When the pumping unit is in a non-moving state and the handbrake is not engaged, or when the pumping unit is in a moving state, the trained neural network model is used to traverse the detection target results of the detected images of the pumping unit to be tested, and to determine whether there are any people present. If personnel are present, their coordinates are compared with the coordinates of the pumping unit. Step S9: Calculate the intersection-union ratio (IUGR) of the personnel target and the fence target in the same frame of the image to be tested; if the IUGR of the personnel target and the fence target are both greater than the preset value, it is determined that the personnel entered the working area of ​​the pumping unit while the pumping unit was not stopped.

2. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 1, is characterized in that... Step S2 can be specifically described as follows: label the oil pumping unit, personnel, and fence in the oil pumping unit image obtained in step S1; use 75% of the labeled sample data as training samples and 25% of the labeled sample data as test samples.

3. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 1, is characterized in that... In step S4, the number of frames of the pumping unit image to be tested in the continuous multi-frame images acquired at intervals shall not be less than 45 frames.

4. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 1, is characterized in that... The edge detection process in step S7 can be specifically described as follows: The image of the oil pumping unit under test containing the oil pumping unit is converted to grayscale. Gaussian filtering is applied to the image of the oil pumping unit under test that contains an oil pumping unit. The gradient intensity and direction of each pixel in the image of the oil pumping unit under test, where an oil pumping unit is present, are calculated. Non-maximum suppression processing is performed on each pixel in the image of the oil pumping unit under test that contains the pumping unit; Double threshold detection is performed on the remaining pixels after nonmaximum suppression processing.

5. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 4, is characterized in that... The grayscale processing can be specifically described as: Gray = 0.299R + 0.587G + 0.114B; where R, G, and B are the color values ​​of the image of the oil pumping unit under test, which contains the oil pumping unit.

6. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 4, is characterized in that... The process of Gaussian filtering can be specifically described as follows: The image of the oil pumping unit under test is convolved with a Gaussian filter and the image of the oil pumping unit containing the oil pumping unit is used. Discrete Gaussian convolution kernel H: (2k+1)×(2k+1) dimensional, its calculation method satisfies: 。 7. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 4, is characterized in that... The process of calculating the gradient intensity and direction of each pixel in an image of a pumping unit containing a pumping unit can be specifically described as follows: 。 8. The method for detecting personnel approaching an oil pumping unit without it being shut down, based on deep learning, as described in claim 1, is characterized in that... The default value for the intersection-union ratio of the personnel target and the fence target in step S9 is 0.3.

Images