Personnel off-post detection method and system

Abstract

The invention discloses a personnel off-post detection method and system, and relates to the technical field of image detection, and the method comprises the steps: collecting a video image, carrying out the preprocessing of a video, and extracting the feature information in a video frame image; respectively training a behavior recognition model and a face recognition model by using the feature information to obtain behavior recognition information and face recognition information; obtaining a recognition result according to the behavior recognition information and the face recognition information; and judging the personnel departure condition according to the identification result. Outputting a first recognition result through the behavior recognition model; if the first identification result is non-human body, warning information is sent out; if the first recognition result is a human body, outputting a face recognition result through a face recognition model; comparing the face recognition result with preset personnel information, and recording personnel on-duty conditions; and comparing the personnel on-duty condition with a preset rule and outputting warning information. According to the working environment requirement, whether the monitoring personnel leave the post or not is intelligently judged, and correct alarming or reminding can be carried out.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection, and more particularly to a method and system for detecting personnel leaving their posts. Background Art

[0002] With the increasing popularity of monitoring electronic devices in various fields, a large number of monitoring videos and images are in urgent need of being effectively processed. At the same time, due to business needs, companies (such as banks, stock exchanges, etc.) and business halls in society need to ensure the attendance rate of employees. In order to monitor whether employees arrive on time and leave their posts at will, they need to use monitoring to understand in real time so as to have an accurate attendance assessment of employees. The traditional monitoring method is to use manual monitoring, but the manual monitoring video method is inefficient and difficult to guarantee accuracy. Therefore, there is an urgent need for a technology that can intelligently determine whether the monitored personnel are off duty and can correctly alarm or remind.

[0003] In the field of computer vision, there are a large number of methods that can be used for personnel absence detection, such as location detection, target detection and tracking. However, in many cases, the real-time and accuracy of background modeling, foreground target detection and tracking are difficult to meet the requirements. At the same time, pre-training models based on large-scale datasets are very important. However, there is currently no dedicated database suitable for working environments, which prevents neural networks from obtaining rich low-level features for image recognition.

[0004] Therefore, how to propose a method and system for detecting personnel leaving their posts, intelligently determine whether the monitored personnel are away from their posts according to the requirements of the work environment, and correctly alarm or remind is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides a method and system for detecting personnel leaving their posts, which can intelligently determine whether the monitored personnel are leaving their posts according to the working environment requirements, and can correctly alarm or remind. In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for detecting when a person leaves his post, comprising:

[0007] Collect video images, pre-process the video, and extract feature information from video frame images;

[0008] Use feature information to train the behavior recognition model and face recognition model respectively;

[0009] Obtaining behavior recognition information and face recognition information according to the behavior recognition model and the face recognition model;

[0010] Obtain recognition results based on the behavior recognition information and the face recognition information, refer to the recognition results of the adjacent frames, and correct the recognition results of the current frame;

[0011] Determine the personnel's off-duty status based on the identification results.

[0012] Optionally, the capturing of video images includes: installing a high-definition camera in a work scene of personnel at a construction site, factory or office area to capture real-time video images of the work area.

[0013] Optionally, the preprocessing of the video includes: performing median filtering on the collected video images to alleviate the impact of noise in the data; enhancing contrast to separate the foreground and background of the person and the background.

[0014] Optionally, the extracting of feature information from the video frame image includes: dividing the video image into multiple image frames, identifying feature points of the image frame based on the similarity between the image frame and reference feature points through image processing technology, and deriving the position of the feature point in the image frame, deriving first feature information based on the position of the feature point, and training the behavior recognition model through the first feature information.

[0015] Optionally, it also includes: distinguishing video image data, for color images, extracting each channel of the color image and the grayscale image as a data source; for grayscale images, using the grayscale image as a data source to obtain second feature information, and training a face recognition model through the second feature information.

[0016] Optionally, obtaining the recognition result based on the behavior recognition information and the face recognition information further includes:

[0017] Output multiple target positioning frames and their confidences at the output end of the face recognition model, and screen out targets based on the confidences;

[0018] A single frame image is divided into multiple target regions, each region is responsible for detecting the target within the region and outputting the target positioning frame and confidence, and setting a first threshold and a second threshold;

[0019] If the confidence of the target is higher than the specified first threshold, it is judged as true and the target is retained;

[0020] If the confidence of the target is lower than the specified second threshold, it is judged as false and the target is screened out;

[0021] If the confidence of the target is within the threshold range, the target is retained.

[0022] Optionally, the correcting the recognition result of the current frame by referring to the recognition result of the adjacent frame includes:

[0023] Select the previous and next frames of the current frame, compare the recognition results of the previous and next frames with the recognition result of the current frame, and correct the recognition result of the current frame if the deviation of the recognition result of the current frame is too large;

[0024] Select the first n target positioning frames as the recognition results according to the confidence level within the threshold interval determined by the current frame, filter out other target positioning frames of the frame, and compare the recognition results of the frame with those of the previous and next frames respectively;

[0025] If the Jaccard similarity between a recognition result of the current frame and the recognition results of the previous and next frames is less than a given threshold, then there is a deviation between the recognition result and the recognition results of the previous and next frames, and the recognition result is screened out in the current frame.

[0026] Optionally, judging the off-duty status of personnel according to the recognition result includes:

[0027] The behavior recognition model outputs multiple information, obtains a first recognition result according to the maximum correlation degree, and corrects the recognition result of the current frame by referring to the recognition results of the adjacent frames;

[0028] If the first recognition result is non-human, a warning message is issued;

[0029] If the first recognition result is a human body, the face recognition model is used to perform face recognition again to output a second recognition result, and the recognition result of the current frame is corrected by referring to the recognition results of the adjacent frames;

[0030] Compare the second recognition result with the preset personnel information and record the personnel's on-the-job status;

[0031] Compare the personnel's on-the-job status with the preset rules and output warning information.

[0032] Optionally, a personnel leaving the post detection system includes:

[0033] Acquisition module: used to collect video images,

[0034] Preprocessing module: used to preprocess the video and extract feature information from the video frame image;

[0035] Training module: used to train the behavior recognition model and face recognition model using feature information respectively;

[0036] Recognition module: used to obtain behavior recognition information and face recognition information based on the behavior recognition model and face recognition model;

[0037] Correction module: used to obtain recognition results based on behavior recognition information and face recognition information, refer to the recognition results of adjacent frames, and correct the recognition results of the current frame;

[0038] Judgment module: used to judge the off-duty status of personnel based on the recognition results.

[0039] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for detecting personnel leaving their posts, which has the following beneficial effects:

[0040] The present invention discloses a method for detecting personnel leaving their posts, comprising: collecting video images, preprocessing the videos, and extracting feature information from video frame images; respectively training a behavior recognition model and a face recognition model with the feature information; obtaining behavior recognition information and face recognition information according to the behavior recognition model and the face recognition model; obtaining recognition results according to the behavior recognition information and the face recognition information, and correcting the recognition results of the current frame with reference to the recognition results of the adjacent frames; and judging the personnel leaving their posts according to the recognition results. Outputting multiple information through the behavior recognition model, obtaining a first recognition result according to the maximum degree of association, and correcting the recognition results of the current frame with reference to the recognition results of the adjacent frames; if the first recognition result is a non-human body, a warning message is issued; if the first recognition result is a human body, performing face recognition again through the face recognition model to output the face recognition result, and correcting the recognition result of the current frame with reference to the recognition results of the adjacent frames; comparing the face recognition result with the preset personnel information, and recording the personnel on-the-job status; and comparing the personnel on-the-job status with the preset rules to output a warning message. It realizes intelligent judgment of whether the personnel in the monitoring are leaving their posts according to the requirements of the working environment, and can correctly alarm or remind. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0042] Figure 1 A schematic flow chart of a method for detecting personnel leaving their posts provided by the present invention.

[0043] Figure 2 This is a structural framework diagram of a personnel absence detection system provided by the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The embodiment of the present invention discloses a method for detecting personnel leaving their post, such as Figure 1 As shown, including:

[0046] Collect video images, pre-process the video, and extract feature information from video frame images;

[0047] Use feature information to train the behavior recognition model and face recognition model respectively;

[0048] Obtaining behavior recognition information and face recognition information according to the behavior recognition model and the face recognition model;

[0049] Obtain recognition results based on the behavior recognition information and the face recognition information, refer to the recognition results of the adjacent frames, and correct the recognition results of the current frame;

[0050] Determine the personnel's off-duty status based on the identification results.

[0051] Furthermore, the collecting of video images includes: installing a high-definition camera in a work scene of personnel at a construction site, factory or office area to collect real-time video images of the work area.

[0052] Furthermore, the video preprocessing includes: performing median filtering on the collected video images to alleviate the influence of noise in the data; enhancing contrast to separate the foreground and background of the person and the background.

[0053] In a specific embodiment, the preprocessing of the video further includes:

[0054] For each pixel in the video image, the degree of abnormality of each pixel is calculated according to the difference in grayscale values ​​between the pixel and adjacent pixels, multiple abnormal points are determined from all the pixels according to the probability that the pixel is an abnormal point, the possibility that each abnormal point belongs to a feature point is calculated according to the distribution of other abnormal points in the first distribution domain of the abnormal point, and multiple feature points are determined from all the abnormal points according to the possibility that the abnormal point belongs to a feature point;

[0055] Determine the possibility that the other feature points in the first distribution domain of each feature point are reference points according to the relative positions of other feature points in the first distribution domain of each feature point, and determine multiple reference points for each feature point from the other feature points in the first distribution domain of each feature point according to the possibility that the other feature points are reference points;

[0056] Calculate the similarity between each reference point of each feature point and each feature point according to the grayscale difference between the second distribution domain of each feature point and the second distribution domain of each reference point of each feature point, and the relative position of each feature point and each reference point of each feature point;

[0057] Calculate the contrast difference between each reference point of each feature point and each feature point according to the color value difference between the third distribution domain of each feature point and the pixel points in the third distribution domain of each reference point of each feature point, and the distance measurement between each feature point and each reference point of each feature point;

[0058] According to the similarity between each reference point of each feature point and each feature point, the possibility that each reference point of each feature point belongs to a feature point, the possibility that each reference point of each feature point is a reference point, and the contrast difference between each reference point of each feature point and each feature point, a weighted sum is performed on the grayscale value of each reference point of each feature point to determine the corrected grayscale value of each feature point;

[0059] According to the corrected grayscale values ​​of the feature points, a corrected video image is obtained.

[0060] Furthermore, the method of extracting feature information from video frame images includes: dividing the corrected video image into multiple image frames, identifying feature points of the image frame based on the similarity between the image frame and reference feature points through image processing technology, and obtaining the position of the feature point in the image frame, obtaining first feature information based on the position of the feature point, and training the behavior recognition model through the first feature information.

[0061] Further, the first feature information is obtained according to the position of the feature point, including:

[0062] The position information of all feature points corresponding to different image frames is used to form a feature point position set, and each feature point corresponds to a feature point position set; the feature point position set corresponding to each position information is selected, and the time continuity and time continuity duration reach the preset duration as the feature data set; the position information in the feature data set is substituted into the iterative model for debugging, and the weight parameters in the iterative model are adjusted; each feature point corresponds to a unique iterative model; the set of all iterative models is used to summarize the changes of the position information in all feature data sets over time to obtain the first feature information. Among them, the feature points are usually set in advance by designers or staff, and the reference feature points are usually selected from the joints. The movement changes of the joints are recorded to obtain the movement changes of the whole human body, and the iterative model can be a neural network model.

[0063] In a specific implementation, the behavior recognition model includes: obtaining video image information to be recognized; extracting feature information from the video image information; substituting the feature information into a preset behavior recognition model to obtain behavior information; wherein the behavior recognition model is obtained by training employee basic information, employee behavior information and a neural network model, and obtaining an employee basic database, the employee basic database including corresponding employee basic information and employee behavior information; training the neural network model based on the employee basic database, the employee basic information including name, age, employment status, occupation, medical history, and the character behavior information including feature information and habits; and obtaining recognition results based on the behavior information.

[0064] Specifically, the first feature information is substituted into the behavior recognition model to obtain multiple basic information of the characters; the degree of correlation between each basic information of the characters and the first feature information is calculated, and the basic information of the characters is sorted according to the degree of correlation; the basic information of the characters whose correlation degree reaches a preset threshold is screened out, and combined with the first feature information as the behavior information; the basic information of the characters with the largest correlation degree is selected as the recognition result.

[0065] The present invention extracts feature information based on the image information of the person to be identified, and the feature information represents the motion feature trajectory of the person to be identified; then the feature information is used as the key point and substituted into a behavior recognition model, and the feature information is used as part of the person's behavior information. According to the behavior recognition model, multiple corresponding basic information of the person is found, and then the basic information of the person is used as the behavior information, thereby achieving the purpose of human body recognition.

[0066] Furthermore, it also includes: distinguishing video image data, for color images, extracting each channel of the color image and the grayscale image as a data source; for grayscale images, using the grayscale image as a data source to obtain second feature information, and training a face recognition model through the second feature information.

[0067] In a specific implementation, the training face recognition model includes:

[0068] Build a face recognition model based on neural network:

[0069] The face recognition model based on neural network includes: an input layer, four convolutional layers, a maximum pooling layer, a fully connected layer and an output layer connected in sequence. The input layer has an input layer for each channel of each input image and outputs to the first convolutional layer; the first convolutional layer has a convolutional layer for each channel in the input layer, which extracts different features from the information of different channels, fuses the feature maps, and then outputs to the second convolutional layer; the second convolutional layer uses a multiple convolution method to extract targets of different scales to scales of different receptive fields, and then fuses the feature maps and performs batch normalization, and outputs to the third convolutional layer; the third convolutional layer uses a convolution weight reuse method to perform downsampling and feature extraction, and outputs to the fourth convolutional layer; the fourth convolutional layer connects the third convolutional layer and the input layer, retains the image feature information of the original image, and finally outputs to the maximum pooling layer;

[0070] Train a neural network-based face recognition model:

[0071] Fine-tune the learning rate; input the three RGB channels and the grayscale channel of the current frame image, process these four input channels separately and then fuse them; set different convolution kernel sizes for the input of different channels; during training, the output weight of the grayscale channel is higher than the output weight of the three RGB color image channels; set the layer that fuses the information of the shallow convolution layer and the deep convolution layer; use video frame extraction training.

[0072] Output of the neural network-based face recognition model: In the output layer, the output of the entire neural network model is divided into several small blocks, each of which is responsible for detecting multiple targets. Each target includes: relative position, positioning box size and confidence.

[0073] The present invention uses the RGB channels in the color image as input, and can obtain and utilize the information of each channel more specifically. Using the grayscale image as input and setting a larger loss weight can allow the model to learn more distinguishable color-independent features such as shape and size, reducing the dependence on the color features of the video image.

[0074] Furthermore, obtaining the recognition result based on the behavior recognition information and the face recognition information also includes:

[0075] Output multiple target positioning frames and their confidences at the output end of the face recognition model, and screen out targets based on the confidences;

[0076] A single frame image is divided into multiple target regions, each region is responsible for detecting the target within the region and outputting the target positioning frame and confidence, and setting a first threshold and a second threshold;

[0077] If the confidence of the target is higher than the specified first threshold, it is judged as true and the target is retained;

[0078] If the confidence of the target is lower than the specified second threshold, it is judged as false and the target is screened out;

[0079] If the confidence of the target is within the threshold range, the target is retained.

[0080] Furthermore, the reference to the recognition results of the adjacent frames to correct the recognition result of the current frame includes:

[0081] Select the previous and next frames of the current frame, compare the recognition results of the previous and next frames with the recognition result of the current frame, and correct the recognition result of the current frame if the deviation of the recognition result of the current frame is too large;

[0082] Select the first n target positioning frames as the recognition results according to the confidence level within the threshold interval determined by the current frame, filter out other target positioning frames of the frame, and compare the recognition results of the frame with those of the previous and next frames respectively;

[0083] If the Jaccard similarity between a recognition result of the current frame and the recognition results of the previous and next frames is less than a given threshold, then there is a deviation between the recognition result and the recognition results of the previous and next frames, and the recognition result is screened out in the current frame.

[0084] Using video frame extraction training can effectively reduce sample redundancy and improve sample diversity. The recognition accuracy of the current frame is improved by using the recognition results of the previous and next frames as a reference.

[0085] Furthermore, judging the off-duty status of personnel according to the recognition result includes:

[0086] The behavior recognition model outputs multiple information, obtains a first recognition result according to the maximum correlation degree, and corrects the recognition result of the current frame by referring to the recognition results of the adjacent frames;

[0087] If the first recognition result is non-human, a warning message is issued;

[0088] If the first recognition result is a human body, the face recognition model is used to perform face recognition again to output the face recognition result, and the recognition result of the current frame is corrected with reference to the recognition results of the adjacent frames;

[0089] Compare the face recognition results with the preset personnel information and record the personnel's on-the-job status;

[0090] Compare the personnel's on-the-job status with the preset rules and output warning information.

[0091] In a specific implementation, it also includes: setting corresponding employee duty standards according to the work area;

[0092] For work areas that require 24-hour duty, it is necessary to ensure that personnel on duty are detected at all times in the work area. The behavior recognition model is used to detect in real time whether there are personnel on duty in the work area, and a warning message is issued when the staff is off duty; when there are personnel on duty, the face recognition model is used to identify the personnel on duty and check the duty arrangement. The face recognition model is used to detect the current personnel on duty in real time and record them, and collect the personnel on duty status;

[0093] For work areas that do not require 24-hour duty, it is necessary to ensure that personnel on duty are detected in the work area within the specified time. The behavior recognition model is used to detect whether there are personnel on duty in the work area, and a warning message is issued when the staff leaves the post; when there are personnel on duty, the face recognition model is used to identify the on-duty personnel and check them with the duty arrangement. The face recognition model is used to detect the current on-duty personnel in real time and record them, and the on-duty status of the personnel is collected.

[0094] In a specific implementation, a scheduling module is also included, through which employees make temporary leave requests and leave request applications, and the system changes the duty arrangement in the work area according to the temporary leave requests and leave request applications. After the duty arrangement in the work area is updated, the updated duty arrangement in the work area is compared with the first recognition result and the second recognition result to see whether a warning message is output, and a record is made to collect the on-duty status of the personnel.

[0095] In a specific implementation, it also includes collecting the on-the-job status of personnel to form third characteristic information;

[0096] Extracting work area accident data according to the third characteristic information and constructing a work area accident knowledge graph;

[0097] Perform representation learning on entities in the work area accident knowledge graph and generate vector representations of entities;

[0098] Based on the fully connected deep neural network trained for multi-classification tasks, the corresponding work area accident labels are predicted using the vector representation of the work area accident.

[0099] The vector representation space is used to generate the main causes of the work area accident label accident, where the main causes of the accident are entities related to the work area accident label in the work area accident knowledge graph.

[0100] In a specific embodiment, the specific steps of constructing the work area accident knowledge graph are:

[0101] Extracting original multi-dimensional features of work area accidents, including at least work area accident environment features and work area accident event features;

[0102] Convert the original multi-dimensional features of work area accidents into numerical features;

[0103] Extract the accident labels of the working area according to the characteristics of the accident events in the working area;

[0104] The entities, relationships and concept entities of the work area accident characteristics are extracted to generate a work area accident knowledge graph, where an entity refers to a specific object or event corresponding to the work area accident characteristics; a concept entity refers to a specific value corresponding to the work area accident characteristics; and a relationship refers to the relationship between an instance entity and a concept entity, as well as the relationship between an instance entity and an instance entity.

[0105] In a specific implementation, the representation learning of entities in the work area accident knowledge graph to generate vector representations of entities includes: training a TransE model to represent entities in the work area accident knowledge graph to generate vector representations of entities; clustering the work area accident entities in the work area accident knowledge graph using a K-Means algorithm, and removing work area accident entities whose distance from the cluster center exceeds a distance threshold from the work area accident knowledge graph.

[0106] In a specific embodiment, the use of vector representation space to generate the main causes of work area accident labels includes: mapping each entity to Euclidean space to infer the possible main causes of work area accidents with accident labels, and searching for all work area accidents with the same label in the Euclidean space; for each entity in the Euclidean space, calculating the distance from the label to the entity; selecting the entity with the smallest distance, and considering the entity as the main cause of the accident corresponding to the label.

[0107] In a specific implementation, a personnel leaving the post detection system, such as Figure 2 As shown, including:

[0108] Acquisition module: used to collect video images,

[0109] Preprocessing module: used to preprocess the video and extract feature information from the video frame image;

[0110] Training module: used to train the behavior recognition model and face recognition model using feature information respectively;

[0111] Recognition module: used to obtain behavior recognition information and face recognition information based on the behavior recognition model and face recognition model;

[0112] Correction module: used to obtain recognition results based on behavior recognition information and face recognition information, refer to the recognition results of adjacent frames, and correct the recognition results of the current frame;

[0113] Judgment module: used to judge the off-duty status of personnel based on the recognition results.

[0114] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0115] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting when a person is away from his post, characterized in that: include: Collect video images, pre-process the video, and extract feature information from video frame images; Use feature information to train the behavior recognition model and face recognition model respectively; Obtaining behavior recognition information and face recognition information according to the behavior recognition model and the face recognition model; Obtain recognition results based on the behavior recognition information and the face recognition information, refer to the recognition results of the adjacent frames, and correct the recognition results of the current frame; Determine the personnel's off-duty status based on the identification results.

2. A method for detecting personnel leaving their post according to claim 1, characterized in that: The video image acquisition includes: installing a high-definition camera in a work scene of personnel at a construction site, factory or office area to acquire real-time video images of the work area.

3. A method for detecting personnel leaving their post according to claim 1, characterized in that: The video preprocessing includes: performing median filtering on the collected video images to alleviate the influence of noise in the data; enhancing the contrast and separating the foreground and background of the person and the background.

4. A method for detecting personnel leaving their post according to claim 1, characterized in that: The method of extracting feature information from a video frame image includes: dividing the video image into a plurality of image frames, identifying feature points of the image frame based on the similarity between the image frame and a reference feature point through image processing technology, and obtaining the position of the feature point in the image frame, obtaining first feature information based on the position of the feature point, and training a behavior recognition model through the first feature information.

5. A method for detecting personnel leaving their post according to claim 4, characterized in that: Also includes: Distinguish video image data, and for color images, extract each channel of the color image and the grayscale image as data sources; For grayscale images, the grayscale images are used as a data source to obtain second feature information, and the face recognition model is trained through the second feature information.

6. A method for detecting personnel leaving their post according to claim 1, characterized in that: The obtaining of the recognition result based on the behavior recognition information and the face recognition information also includes: Output multiple target positioning frames and their confidences at the output end of the face recognition model, and screen out targets based on the confidences; A single frame image is divided into multiple target regions, each region is responsible for detecting the target within the region and outputting the target positioning frame and confidence, and setting a first threshold and a second threshold; If the confidence of the target is higher than the specified first threshold, it is judged as true and the target is retained; If the confidence of the target is lower than the specified second threshold, it is judged as false and the target is screened out; If the confidence of the target is within the threshold range, the target is retained.

7. A method for detecting personnel leaving their post according to claim 6, characterized in that: The step of correcting the recognition result of the current frame by referring to the recognition result of the adjacent frame includes: Select the previous and next frames of the current frame, compare the recognition results of the previous and next frames with the recognition result of the current frame, and correct the recognition result of the current frame if the deviation of the recognition result of the current frame is too large; Select the first n target positioning frames as the recognition results according to the confidence level within the threshold interval determined by the current frame, filter out other target positioning frames of the frame, and compare the recognition results of the frame with those of the previous and next frames respectively; If the Jaccard similarity between a recognition result of the current frame and the recognition results of the previous and next frames is less than a given threshold, then there is a deviation between the recognition result and the recognition results of the previous and next frames, and the recognition result is screened out in the current frame.

8. A method for detecting personnel leaving their post according to claim 1, characterized in that: The judging of the personnel off-duty status according to the recognition result includes: The behavior recognition model outputs multiple information, obtains a first recognition result according to the maximum correlation degree, and corrects the recognition result of the current frame by referring to the recognition results of the adjacent frames; If the first recognition result is non-human, a warning message is issued; If the first recognition result is a human body, the face recognition model is used to perform face recognition again to output a second recognition result, and the recognition result of the current frame is corrected by referring to the recognition results of the adjacent frames; Compare the second recognition result with the preset personnel information and record the personnel's on-the-job status; Compare the personnel's on-the-job status with the preset rules and output warning information.

9. A personnel leaving post detection system, characterized in that: include: Acquisition module: used to collect video images, Preprocessing module: used to preprocess the video and extract feature information from the video frame image; Training module: used to train the behavior recognition model and face recognition model using feature information respectively; Recognition module: used to obtain behavior recognition information and face recognition information based on the behavior recognition model and face recognition model; Correction module: used to obtain recognition results based on behavior recognition information and face recognition information, refer to the recognition results of adjacent frames, and correct the recognition results of the current frame; Judgment module: used to judge the off-duty status of personnel based on the recognition results.

Images