A periscope dome surveillance system

Abstract

The present invention discloses a periscope-type vault monitoring system, comprising: a monitoring device, an image acquisition module, and an image recognition module connected thereto; relating to the technical field of vault monitoring, and solving the technical problem that the current method of monitoring the vault cannot directly and quickly understand the casting situation of the vault; by using the monitoring device to obtain the monitoring video in real time during the casting of the tunnel vault, and extracting the depth-of-field detection image and the corresponding monitoring time from the monitoring video by means of frame extraction, it is possible to reduce the processing of redundant data volume, so as to be able to perform real-time monitoring, avoid the problem that it is more troublesome to repair the area with problems in the casting after the casting is formed, and at the same time, by analyzing the detection information, the accuracy of the monitoring result can be ensured.

Description

Technical Field

[0001] The present invention belongs to the field of construction, relates to the technology of vault monitoring, and specifically is a periscope-type vault monitoring system. Background Art

[0002] Tunnel projects account for a large proportion in the construction in various places. Since the main function of the tunnel vault is to bear the pressure from the top of the tunnel and play a role in maintaining the overall stability of the tunnel, in the field of tunnel construction such as building construction and mine exploitation, the construction quality of the vault is directly related to the safety and stability of the tunnel. Therefore, the concrete placement operation of the vault is a key process to ensure the structural safety and construction quality. And the uneven pouring of the tunnel vault will lead to the appearance of cavities, which will pose a serious threat to the overall structural safety of the tunnel.

[0003] Currently, the common method for monitoring the pouring of the tunnel vault is to calculate the required amount of concrete for pouring, and then compare it with the actual usage amount of concrete through statistics to monitor the pouring of the vault. However, this method cannot directly understand whether the internal pouring of the tunnel vault is uniform. Although pressure sensors or displacement gauges are installed on the formwork or support structure, and the concrete filling situation is inferred through mechanical data, this method can only detect the local state at the installation position of the sensor, and the data can only stabilize after the concrete solidifies. Therefore, the pouring situation of the vault cannot be obtained in a timely manner. Summary of the Invention

[0004] The present invention aims to at least solve one of the technical problems existing in the prior art; for this purpose, the present invention provides a periscope-type vault monitoring system, which is used to solve the technical problem that the current method for monitoring the vault cannot directly and quickly understand the pouring situation of the vault. The present invention can obtain the monitoring video during the pouring of the tunnel vault in real time through the monitoring device, and extract the depth-of-field detection image and the corresponding monitoring time from the monitoring video by means of frame extraction of the picture. By analyzing the detection information, the accuracy of the monitoring result can be guaranteed, thus solving the above problems.

[0005] To achieve the above object, the first aspect of the present invention provides a periscope-type vault monitoring system, including: a monitoring device, an image acquisition module, and an image recognition module connected thereto;

[0006] The image acquisition module: is used to extract the monitoring video during the pouring of the vault captured by the monitoring device, extract a number of monitoring information from the monitoring video, and send the number of monitoring information to the image recognition module; wherein, the monitoring information includes the depth-of-field monitoring image and the monitoring time;

[0007] The image recognition module: is used to classify the detection images in the monitoring information, then extract and analyze the image features to obtain the pouring information of the roof.

[0008] It should be noted that: the position of the monitoring device is moved according to the distance from the casting end of the vault.

[0009] Preferably, the depth of field monitoring image is the picture in the monitoring video during the monitoring time, and the monitoring time is the shooting time of the depth of field monitoring image.

[0010] In order to further ensure the accuracy, multiple frames of pictures can be extracted when casting at the same position of the casting vault. In this way, it can not only ensure the reduction of redundant data volume and improve the processing efficiency, but also ensure its accuracy to a certain extent.

[0011] Preferably, after classifying the detection images in the monitoring information, image feature extraction and analysis are carried out, including:

[0012] Classify several vault monitoring images through an image classification model to obtain several groups of monitoring image sets; the casting positions of each group of the monitoring image sets are the same;

[0013] Generate depth maps for all the vault monitoring images in each group of detection image sets through a monocular depth estimation model; obtain the shooting distance between the vault monitoring image and the monitoring device, and extract the casting images from the vault monitoring images according to the shooting distance and the depth maps;

[0014] Obtain the casting information according to the casting images and the monitoring time corresponding to the vault monitoring images in several groups of monitoring image sets; wherein, the casting information includes the uniformity of the casting volume and the casting speed.

[0015] It should be noted that when the monitoring is just started, since the amount of data is small, historical data can be introduced when analyzing and comparing the vault monitoring images in each group of monitoring image sets. However, it should be noted that the introduced historical data needs to ensure the consistency of the data of the tunnel vault. If they are consistent, manual correction is required. And the "same casting position" mentioned above refers to the area where the cross-sectional positions of the tunnel vault at different depths are the same.

[0016] Preferably, the classification of several vault monitoring images includes:

[0017] Construct an image classification model: construct an image classification model based on a convolutional neural network and introducing an attention mechanism. Obtain several vault monitoring images and the corresponding classification labels through a database; integrate several vault monitoring images and the corresponding classification labels into a training data group and a test data group; train the image classification model through the training data group, and test the trained image classification model through the test data group; adjust the image classification model according to the test accuracy, and finally obtain an image classification model with the vault monitoring image as the input and the classification label as the output.

[0018] Image classification: Integrate the vault monitoring images with the same output classification labels to obtain several groups of monitoring image sets.

[0019] It should be noted that the classification label corresponding to the vault monitoring image is determined according to the position of the pouring end of the vault, that is, the classification labels corresponding to different pouring ends are different.

[0020] Preferably, the extracting the pouring image from the vault monitoring image according to the shooting distance and the depth map includes:

[0021] Extract the vault monitoring image and the corresponding depth map; obtain the gray value of each pixel point in the depth map, and calculate the monitoring distance of each pixel point from the monitoring device according to the linear mapping relationship between the gray value and the distance;

[0022] Compare the monitoring distance with the preset pouring distance interval; when the monitoring distance is within the pouring distance interval, mark the pixel corresponding to the monitoring distance as the target pixel; when the monitoring distance is outside the pouring distance area, mark the pixel corresponding to the monitoring distance as the irrelevant pixel, and correct the irrelevant pixel based on the positional relationship of the target pixel;

[0023] Extract and count the target pixels of each vault monitoring image in the monitoring image set to obtain the pouring image.

[0024] Among them, the pouring distance interval is determined based on the shooting distance and the maximum distance of the free fall of concrete during vault pouring; and the linear mapping relationship between the gray value and the distance is , where d is the monitoring distance, g is the gray value of the pixel point, d max is the distance of the farthest object from the monitoring device in the monitoring image, d min is the distance of the nearest object from the monitoring device in the monitoring image.

[0025] Preferably, the correcting the irrelevant pixel based on the positional relationship of the target pixel includes:

[0026] Extract the area covered and surrounded by the target pixel and mark it as the target area; determine whether all around the irrelevant pixel in the target area are target pixels; if so, mark the corresponding irrelevant pixel as the target pixel; if not, keep the mark of the irrelevant pixel.

[0027] Preferably, the obtaining method of the pouring information includes:

[0028] Extract the number of target pixels of the pouring image corresponding to each vault monitoring image in several groups of detection image sets and the time difference between pouring two identical vault areas;

[0029] Obtain measurement data for measuring the number of target pixels and the degree of dispersion of each time difference in each set of detected image sets; among them, the measurement data can be standard deviation, variance, range, etc.;

[0030] Compare the measurement data corresponding to the number of target pixels and each time difference with a preset threshold respectively to obtain pouring information.

[0031] Specifically, the uniformity of the pouring volume includes uniformity and differences; the uniformity of the pouring speed also includes uniformity and differences, and the differences can also be classified according to the difference between the measurement data and the corresponding preset threshold. The time difference can reflect whether the pouring time between two vault pouring areas is too different.

[0032] Preferably, the obtaining method of the pouring information includes:

[0033] Extract the number of target pixels of the pouring image corresponding to each vault monitoring image in several sets of detected image sets and the time difference between each vault monitoring image and the adjacent vault monitoring image;

[0034] Input the vault monitoring image and the corresponding target pixel data and time difference into a prediction model to obtain an output label, and match the corresponding pouring information according to the output label. Among them, the prediction model is constructed based on a recurrent neural network model.

[0035] Preferably, the construction method of the prediction model includes:

[0036] Train a BP neural network model with standard training data, and mark the trained BP neural network model as a prediction model; among them, the standard training data consists of training input data and training output data. The training input data is the vault monitoring image and the corresponding target pixel data and time difference, and the training output data is the output label corresponding to the training input data.

[0037] It should be noted that the training input data is historical data, and the training output data is manually labeled according to the input data.

[0038] Preferably, the image recognition module is also used to: judge the availability of the depth of field monitoring image; when the monitoring information is unavailable, adjust the monitoring parameters of the monitoring device and re-obtain the monitoring information;

[0039] The method for judging the availability of the depth of field monitoring image can perform a Fourier transform on the image, extract the high-frequency components, and judge whether the depth of field monitoring image is clear by calculating the proportion of the high-frequency components and comparing it with a preset threshold. The lower the proportion of the high-frequency components, the blurrier the image.

[0040] Preferably, the monitoring device includes a periscope, an image acquisition and transmission device, a slewing bearing, a guide rod and a servo motor;

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows: By using the monitoring device to obtain the monitoring video during the casting of the tunnel vault in real time, and extracting the depth-of-field detection image and the corresponding monitoring time from the monitoring video by means of frame extraction, the processing of redundant data volume can be reduced, so that real-time monitoring can be carried out, avoiding the problem that it is more troublesome to repair the area with problems after casting. At the same time, by analyzing the detection information, the accuracy of the monitoring results can be ensured; And in order to obtain the uniformity of each casting area of the tunnel vault, the casting image is extracted from the concrete image in the depth-of-field monitoring image through gray-distance mapping, and the casting images corresponding to the same-position casting are compared, so that comparison can be carried out in the areas of each same casting position at the depth of the tunnel vault to measure whether there are large differences in the casting volume of the concrete in the vault. Through this method, independent analysis of different casting areas is realized, which can avoid the mixing of data at different positions and further improve the accuracy of monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0043] Figure 1 It is a schematic flow chart of Embodiment 1 of the present invention;

[0044] Figure 2 It is a schematic flow chart of Embodiment 3 of the present invention;

[0045] Figure 3 It is a schematic diagram of the vault monitoring device of the present invention.

[0046] In the figure: 1, periscope device; 2, image acquisition and transmission device; 3, slewing bearing; 4, guide rod; 5, servo motor. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0048] Please refer toFigures 1-3 , the present invention provides the following embodiments:

[0049] Embodiment 1

[0050] A periscope-type vault monitoring system, comprising: a monitoring device, an image acquisition module, and an image recognition module connected thereto;

[0051] The image acquisition module: extracts the monitoring video during the pouring of the vault captured by the image acquisition and transmission device 2 in the monitoring device through 4G transmission, extracts a number of monitoring information from the monitoring video, and sends the number of monitoring information to the image recognition module; wherein, the monitoring information includes depth of field monitoring images and monitoring time;

[0052] The image recognition module: classifies the detection images in the monitoring information, then extracts and analyzes the image features to obtain the pouring information of the vault;

[0053] The monitoring device includes a periscope 1, an image acquisition and transmission device 2, a slewing bearing 3, a guide rod 4, and a servo motor 5; the design and performance of the periscope 1 play a key role in the image acquisition ability of the system. Therefore, both the objective lens and the eyepiece of the periscope 1 adopt high-resolution infrared cameras to ensure that clear and delicate image information can be obtained. The high-resolution infrared cameras can capture more details. In addition, the optical properties of the objective lens and the eyepiece also have good light transmittance, low distortion and other characteristics to ensure the quality and authenticity of the images. At the same time, the design of the prism should consider minimizing the loss of light to improve the light efficiency of the system. The image acquisition and transmission device 2 is a camera device with a 4G module. The slewing bearing 3 is used to control its overall rotation, while the role of the guide rod 4 is mainly to guide the movement direction to ensure that the monitoring device can accurately move and adjust in the vault area of the lining trolley, realizing the comprehensive monitoring of the vault, and the servo motor 5 drives the movement of the monitoring device.

[0054] It should be noted that the above-mentioned depth of field monitoring images are the pictures in the monitoring video during the monitoring time, and the monitoring time is the shooting time of the depth of field monitoring images. In order to further ensure the accuracy, in another preferred embodiment, multiple frames of pictures can be extracted when pouring at the same position of the pouring vault.

[0055] Furthermore, after classifying the detection images in the monitoring information, extracting and analyzing the image features includes:

[0056] Using an image classification model to classify a number of vault monitoring images according to the position during the pouring of the vault to obtain a number of groups of monitoring image sets; the pouring positions of each group of the monitoring image sets are the same;

[0057] Specifically, the construction method of the image classification model includes:

[0058] Building an Image Classification Model: Build an image classification model based on a convolutional neural network and introducing an attention mechanism. Obtain a number of vault monitoring images and their corresponding classification labels through a database; integrate the number of vault monitoring images and their corresponding classification labels into a training data set and a test data set; train the image classification model with the training data set and test the trained image classification model with the test data set; adjust the image classification model according to the test accuracy, and finally obtain an image classification model with the input being the vault monitoring image and the output being the classification label;

[0059] Image Classification: Integrate the vault monitoring images with the same output classification labels to obtain several groups of monitoring image sets.

[0060] Then generate monocular depth maps for all the vault monitoring images in each group of detection image sets through the MiDaS v3.1 pre-trained model;

[0061] The specific implementation code is as follows:

[0062] import torch

[0063] from midas.model_loader import load_model

[0064] device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

[0065] model, transform = load_model("MiDaS_small", device)

[0066] image = load_image("vault_monitor_001.jpg") # Load the original image

[0067] input_tensor = transform(image).unsqueeze(0).to(device)

[0068] with torch.no_grad():

[0069] depth_map = model(input_tensor) # Output the relative depth values of the depth map

[0070] depth_map = depth_map.squeeze().cpu().numpy()

[0071] Obtain the shooting distance between the vault monitoring image and the monitoring device, and extract the pouring image from the vault monitoring image according to the shooting distance and the depth map;

[0072] Specifically, extracting the pouring image from the vault monitoring image according to the shooting distance and the depth map includes:

[0073] Extract the vault monitoring image and the corresponding depth map; obtain the gray value of each pixel point in the depth map, and calculate the monitoring distance of each pixel point from the monitoring device according to the linear mapping relationship between the gray value and the distance;

[0074] Compare the monitoring distance with the preset pouring distance interval; when the monitoring distance is within the pouring distance interval, mark the pixel corresponding to the monitoring distance as the target pixel; when the monitoring distance is outside the pouring distance area, mark the pixel corresponding to the monitoring distance as the irrelevant pixel, and correct the irrelevant pixel based on the positional relationship of the target pixel;

[0075] Statistically extract the target pixels of each vault monitoring image in the monitoring image set to obtain the pouring image.

[0076] Among them, the pouring distance interval is determined based on the shooting distance and the farthest distance of the free fall of concrete during vault pouring; and the linear mapping relationship between the gray value and the distance is , where d is the monitoring distance, g is the gray value of the pixel point, d max is the distance of the farthest object from the monitoring device in the monitoring image, d min is the distance of the nearest object from the monitoring device in the monitoring image.

[0077] Before construction, measure the farthest point d max = 15m (such as the top of the vault) and the nearest point d min = 2m (such as the position of the reference object) in the scene by a laser rangefinder. In this embodiment, a monocular depth estimation model is used, so the gray value of the pixel 0 (black) → the farthest distance, 255 (white) → the nearest distance.

[0078] Obtain the pouring information according to the pouring images and the monitoring time corresponding to the vault monitoring images in several groups of monitoring image sets; among them, the pouring information includes the uniformity of the pouring volume and the pouring speed.

[0079] The code implemented by the specific pixel classification logic is as follows:

[0080] def classify_pixels(depth_map, D, d_min = 2.0, d_max = 15.0):

[0081] # Calculate the actual distance of each pixel

[0082] normalized_depth = (depth_map - np.min(depth_map)) / (np.max(depth_map) - np.min(depth_map))

[0083] d_pixels = d_max - normalized_depth (d_max - d_min)

[0084] # Set the pouring distance range

[0085] lower_bound = D - 0.5 # Unit: meter

[0086] upper_bound = D + 0.3

[0087] # Generate a binary mask

[0088] mask = np.where((d_pixels >= lower_bound) & (d_pixels <= upper_bound), 1, 0)

[0089] # Fill holes with morphological closing operation

[0090] kernel = np.ones((5,5), np.uint8)

[0091] mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel)

[0092] return mask

[0093] In this implementation, the correction of irrelevant pixels based on the positional relationship of target pixels includes:

[0094] Extract the area covered and surrounded by the target pixel and mark it as the target area; judge whether all pixels in one week of the irrelevant pixels in the target area are target pixels; if so, mark the corresponding irrelevant pixel as a target pixel; if not, keep the mark of the irrelevant pixel; in this embodiment, one week of the irrelevant pixel is eight pixels around this irrelevant pixel.

[0095] Exemplarily, if a certain pixel is marked as an "irrelevant pixel" (mask = 0), but there are more than 5 "target pixels" (mask = 1) in its 8-neighborhood, then it is corrected to a target pixel.

[0096] Specifically, the method for obtaining pouring information includes:

[0097] Extracting the target pixel quantity of the pouring image corresponding to each vault monitoring image in several groups of detection image sets and the time difference between pouring two identical vault areas;

[0098] Obtaining the variance that measures the dispersion degree of each target pixel quantity and each time difference in each group of detection image sets;

[0099] Comparing the variances corresponding to each target pixel quantity and each time difference with a preset threshold respectively to obtain the pouring information.

[0100] Embodiment 2

[0101] Different from the above embodiment, the method for obtaining pouring information in this application includes:

[0102] Extracting the target pixel quantity of the pouring image corresponding to each vault monitoring image in several groups of detection image sets and the time difference between each vault monitoring image and its adjacent vault monitoring image;

[0103] Inputting the vault monitoring image and the corresponding target pixel data and time difference into a prediction model to obtain an output label, and matching the corresponding pouring information according to the output label. Among them, the prediction model is constructed based on a recurrent neural network model.

[0104] Specifically, the construction method of the prediction model includes:

[0105] Training a BP neural network model with standard training data, and then testing the trained model with standard test data. When the accuracy rate of the test reaches the preset threshold, marking the trained BP neural network model as the prediction model; among them, the standard training data consists of training input data and training output data. The training input data is the vault monitoring image and the corresponding target pixel data and time difference, and the training output data is the output label corresponding to the training input data; while the standard test data has the same composition as the standard training data.

[0106] The pouring information obtained through the artificial intelligence model can be iteratively optimized with new data, gradually improving the adaptability to different project scenarios.

[0107] Embodiment 3

[0108] Different from the above embodiment, in order to further improve the accuracy of its pouring information.

[0109] The image recognition module is further used for: judging the availability of the depth of field monitoring image; when the monitoring information is unavailable, adjusting the monitoring parameters of the monitoring device and re-obtaining the monitoring information.

[0110] The method for determining the availability of the depth-of-field monitoring image may perform a Fourier transform on the image, extract the high-frequency components, and determine whether the depth-of-field monitoring image is clear by calculating the proportion of the high-frequency components and comparing it with a preset threshold. The lower the proportion of the high-frequency components, the blurrier the image is.

[0111] Of course, rapid screening can also be carried out by means of manual evaluation.

[0112] Some of the data in the above formula are taken as numerical values after removing the dimension. The formula is obtained by software simulation of a large amount of collected data to get a formula closest to the actual situation; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0113] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A periscope-type vault monitoring system, characterized in that Including: A monitoring device, an image acquisition module, and an image recognition module connected thereto; The image acquisition module: used to extract the monitoring video during the pouring of the vault captured by the monitoring device, extract a number of monitoring information from the monitoring video, and send the number of monitoring information to the image recognition module; wherein, the monitoring information includes a depth of field monitoring image and a monitoring time; The image recognition module: used to classify the detected images in the monitoring information, extract and analyze image features, and obtain the pouring information of the vault; The classification of the detected images in the monitoring information, followed by image feature extraction and analysis, includes: Classify a number of vault monitoring images according to the position during the pouring of the vault to obtain a number of groups of monitoring image sets; the pouring positions of each group of the monitoring image sets are the same; Generate a depth map for all the vault monitoring images in each group of detected image sets through a monocular depth estimation model; obtain the shooting distance between the vault monitoring image and the monitoring device, and extract the pouring image from the vault monitoring image according to the shooting distance and the depth map; Obtain the pouring information according to the pouring images and the monitoring time corresponding to the vault monitoring images in a number of groups of monitoring image sets; wherein, the pouring information includes the uniformity of the pouring volume and the pouring speed; The obtaining method of the pouring information includes: Extract the target pixel number of the pouring image corresponding to each vault monitoring image in a number of groups of detected image sets and the time difference between pouring two identical vault areas; Obtain the measurement data for measuring the dispersion degree of each target pixel number and each time difference in each group of detected image sets; Compare the measurement data corresponding to each target pixel number and each time difference with a preset threshold respectively to obtain the pouring information.

2. The periscope type vault monitoring system according to claim 1, characterized in that, The depth of field monitoring image is the picture in the monitoring video during the monitoring time, and the monitoring time is the shooting time of the depth of field monitoring image.

3. The periscope type vault monitoring system according to claim 1, characterized in that, The method of classifying a number of the vault monitoring images is to classify them through an image classification model; The construction method of the image classification model includes: Construct an image classification model: construct an image classification model based on a convolutional neural network and introduce an attention mechanism. Obtain a number of vault monitoring images and the corresponding classification labels through a database; integrate the number of vault monitoring images and the corresponding classification labels into a training data set and a test data set; train the image classification model through the training data set, and test the trained image classification model through the test data set; adjust the image classification model according to the test accuracy, and finally obtain an image classification model with the input being the vault monitoring image and the output being the classification label; Image classification: Integrate the vault monitoring images with the same output classification label to obtain a number of groups of monitoring image sets.

4. The periscope type vault monitoring system according to claim 1, wherein The extraction of the pouring image from the vault monitoring image according to the shooting distance and the depth map includes: Extract the vault monitoring image and the corresponding depth map; obtain the gray value of each pixel point in the depth map, and calculate the monitoring distance of each pixel point from the monitoring device according to the linear mapping relationship between the gray value and the distance; Compare the monitoring distance with a preset pouring distance range; when the monitoring distance is within the pouring distance range, mark the pixels corresponding to the monitoring distance as target pixels; when the monitoring distance is outside the pouring distance area, mark the pixels corresponding to the monitoring distance as irrelevant pixels, and correct the irrelevant pixels based on the positional relationship of the target pixels. Extract and count the target pixels of each vault monitoring image in the monitoring image set to obtain a pouring image.

5. A periscope-type vault monitoring system according to claim 4, characterized in that, The correction of the irrelevant pixels based on the positional relationship of the target pixels includes: Extract the area covered and surrounded by the target pixels and mark it as the target area; determine whether all pixels in the target area are target pixels; if so, mark the corresponding irrelevant pixels as target pixels; otherwise, keep the marking of the irrelevant pixels.

6. A periscope type vault monitoring system according to claim 1, characterized in that, The method for obtaining the pouring information includes: Extract the number of target pixels of each pouring image corresponding to each vault monitoring image in several groups of detection image sets and the time difference between each vault monitoring image and the adjacent vault monitoring image. Input the vault monitoring image, the corresponding target pixel data and time difference into a prediction model to obtain an output label, and match the corresponding pouring information according to the output label; among them, the prediction model is constructed based on a recurrent neural network model.

7. The periscope dome monitoring system according to claim 6, characterized in that, The construction method of the prediction model includes: Train a BP neural network model with standard training data, and mark the trained BP neural network model as a prediction model; among them, the standard training data consists of training input data and training output data, the training input data is the vault monitoring image, the corresponding target pixel data and time difference, and the training output data is the output label corresponding to the training input data.

8. The periscope type vault monitoring system according to claim 1, characterized in that, The image recognition module is also used to: judge the availability of the depth-of-field monitoring image; when the monitoring information is unavailable, adjust the monitoring parameters of the monitoring device and re-obtain the monitoring information.

Images