Abnormal behavior recognition method and behavior recognition device based on face features

Abstract

The invention belongs to the technical field of video monitoring analysis of petroleum working areas, and particularly relates to an abnormal behavior recognition method and behavior recognition device based on face features, electronic equipment and a computer readable storage medium. According to the invention, by using the face recognition and multi-target tracking recognition model, all workers in a working area can be automatically recognized and tracked, and manual real-time monitoring is not needed; by constructing the training database and utilizing the database to carry out model training, behaviors of workers can be identified more accurately, and identification errors caused by factors such as illumination and angles are reduced; the identity of each worker is accurately recognized by using a face recognition model, so that whether the worker has the corresponding action authority or not is judged; the action tracks of all the workers in the sub-regions are tracked through the multi-target tracking recognition model, and the cross-region workers are recognized through the face recognition model, so that the problem of monitoring cross-region behaviors is effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of video monitoring and analysis of oil working areas, and in particular relates to an abnormal behavior recognition method and a behavior recognition device based on facial features. Background Art

[0002] In the prior art, monitoring and managing the behavior of operators in the work area is an important task, especially in places involving the safety of important facilities and personnel, such as oil fields. In such an environment, it is necessary to ensure that operators act within the permitted work area and avoid potential accidents and risks by mistakenly entering unauthorized areas.

[0003] In this regard, the method usually adopted is to define the action area of ​​the operator and rely on the operator's self-consciousness to impose self-limitation. This method requires the operator's self-consciousness, which may lead to unconsciousness.

[0004] Therefore, at present, the behavior of operators is also monitored through video surveillance systems. Such systems can capture real-time video images in the work area and then monitor the behavior of operators by analyzing these video images.

[0005] However, this method has the following major problems: For large-scale work areas, a large number of monitoring personnel may be required to watch the video surveillance images in real time, which not only consumes human resources, but also may miss some abnormal behaviors due to the limited attention of people. At the same time, for large-scale work areas, it is necessary to monitor the actions of multiple targets at the same time, which may require a large amount of computing, resulting in high configuration requirements for monitoring equipment and possible response jams. In some work areas, different personnel may have different action permissions, and the same person may have action permissions for multiple work areas. For such cross-area behaviors, existing monitoring systems may find it difficult to accurately track and identify them.

[0006] Therefore, a more efficient and accurate technology is needed to solve the above problems in order to accurately monitor and manage the behavior of operators in the work area. Summary of the invention

[0007] In order to solve the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide an abnormal behavior recognition method and behavior recognition device based on facial features, so as to improve the safety in the working area, realize automatic monitoring and management, improve the accuracy of abnormal behavior recognition, and realize sustainable monitoring.

[0008] The technical solution adopted by the present invention is:

[0009] A method for identifying abnormal behavior based on facial features, comprising the following steps:

[0010] S1: Divide the working area into multiple sub-areas;

[0011] S2: Determine the action authority of the operator relative to each sub-area and obtain the facial features of each operator;

[0012] S3: Obtain historical video image data of the working area and build a training database based on the historical video image data;

[0013] S4: Build a multi-target tracking and recognition model and train it through the training database; build a face recognition model;

[0014] S5: Acquire real-time video image data of each sub-area and identify each operator through a face recognition model;

[0015] S6: Determine whether the operators in each sub-area have the corresponding action authority. If not, it is identified as abnormal; if so, it is identified as normal;

[0016] S7: Track the movement trajectories of all operators in the sub-area through the multi-target tracking and recognition model, and determine whether the operators have cross-area behavior; if not, identify it as normal; if yes, identify the cross-area operators through the face recognition model;

[0017] S8: Determine whether the cross-zone operator has the action authority in the corresponding area. If not, it is identified as abnormal; if yes, it is identified as normal;

[0018] S9: Continue steps S7 to S8 to complete the identification of abnormal behaviors in the working area.

[0019] As a preferred embodiment of the present invention, the method for constructing the training database includes:

[0020] Acquire historical video image data of the working area, and extract multiple video frames from the historical video image data;

[0021] Annotate the postures of multiple workers in the video frame to obtain the original picture; the posture annotation categories include: walking, standing, and various working postures corresponding to the work content;

[0022] Perform random rotation, flipping, random color channel conversion, random saturation transformation, and random chromaticity transformation on the original image to obtain a training image;

[0023] The training images are divided into a training set and a test set to obtain a training database.

[0024] As a preferred solution of the present invention, in step S7, time threshold determination and frequency threshold determination are also included, and the time threshold determination and frequency threshold determination include:

[0025] S71: Tracking the movement trajectories of all operators in the sub-area through a multi-target tracking and recognition model;

[0026] S72: Setting a time threshold T; setting a frequency threshold N, and setting the cross-zone frequency n=0;

[0027] S73: When it is detected that a certain operator moves from the atomic area to another sub-area, start timing and obtain the cross-area time t; set the operator as the first operator;

[0028] S74: Continue to track the first operator, and if the first operator does not move to the atomic area within T time, determine that the first operator has performed cross-area behavior;

[0029] If the first operator moves to the atomic region and t<T, it is determined that the first operator has not performed cross-region behavior, and n=n+1;

[0030] S75: Iteratively execute steps S73 to S74. If n<N, it is determined that the first operator has not performed cross-zone behavior; if n≥N, it is determined that the first operator has performed cross-zone behavior.

[0031] As a preferred solution of the present invention, the method for constructing a multi-target tracking and recognition model includes:

[0032] Build a general model of the YOLOv5s algorithm and replace the first CBS layer at the input of YOLOv5s with a 3×3 normal convolution kernel and a 3×3 hole convolution;

[0033] Prune the BN layer in the CBS layer of the backbone network;

[0034] Add regularization constraints to the loss function of the general model of the YOLOv5s algorithm, and obtain the loss function corresponding to the modified model of the YOLOv5s algorithm;

[0035] Build tracking algorithm model;

[0036] The output of the YOLOv5s algorithm correction model is used as the input of the tracking algorithm model, and the tracking algorithm model outputs the moving trajectory.

[0037] As a preferred embodiment of the present invention, the method for constructing the tracking algorithm model includes:

[0038] Obtain the output of the YOLOv5s algorithm correction model as input; perform cascade matching on the input to obtain the first output; perform IOU matching on the first output to obtain the second output;

[0039] Perform Hungarian algorithm matching on the second output to obtain the third output;

[0040] Determine whether the trajectory matching is successful; if unsuccessful, perform Kalman filter prediction on the third output; if successful, output the third output as the final result, and perform Kalman filter update and Kalman filter prediction on the third output in sequence for continuous tracking.

[0041] As a preferred embodiment of the present invention, a method for constructing a face recognition model includes:

[0042] Construct the convolution layer, j out =(j in *str)+f size ;

[0043] Construct a pyramid pooling layer,

[0044] Construct a softmax classifier, whose loss function is:

[0045] Among them, j out is the output feature, j in is the input feature, str is the step size of the previous layer, f size is the size of the convolution kernel, h in is the input quantity, h out is the output, k h is the height of the core, p h is the number of fills, n is the number of pools, e is a natural constant, fy i is the target feature distribution, f i is the target true distribution, i is the index of the sample, and j is the index of the category.

[0046] An abnormal behavior recognition device based on facial features:

[0047] It includes a working area division module: dividing the working area into multiple sub-areas;

[0048] Action authority determination and facial feature acquisition module: determines the action authority of operators relative to each sub-area and acquires the facial features of each operator;

[0049] Training database construction module: obtain historical video image data of the working area and construct a training database based on the historical video image data;

[0050] Model building and training module: build a multi-target tracking and recognition model and train it through the training database; build a face recognition model;

[0051] Real-time video image acquisition and face recognition module: acquires real-time video image data of each sub-area and identifies each operator through the face recognition model;

[0052] Action authority judgment module: judge whether the operators in each sub-area have the corresponding action authority. If not, it is identified as abnormal; if so, it is identified as normal;

[0053] Cross-zone behavior tracking and identification module: The multi-target tracking and identification model is used to track the movement trajectories of all operators in the sub-zone, and determine whether the operators have cross-zone behavior; if not, it is identified as normal; if so, the cross-zone operators are identified through the face recognition model;

[0054] Cross-zone action authority judgment module: judge whether the cross-zone operator has the action authority in the corresponding area. If not, it is identified as abnormal; if so, it is identified as normal.

[0055] An electronic device comprises a processor, a communication interface, a memory and a communication bus; wherein the processor, the communication interface and the memory communicate with each other via the communication bus; the memory is used to store computer programs; the processor is used to execute the programs stored in the memory to implement the above-mentioned abnormal behavior recognition method based on facial features.

[0056] A computer-readable storage medium is used to store computer-readable instructions, and when the instructions are executed, the above-mentioned abnormal behavior recognition method based on facial features is executed.

[0057] The beneficial effects of the present invention are:

[0058] The present invention can automatically identify and track all staff members in the work area by using face recognition and multi-target tracking recognition models, without the need for manual real-time monitoring, thereby greatly saving human resources; by building a training database and using the database for model training, the behavior of staff members can be more accurately identified, and recognition errors caused by factors such as lighting and angles can be reduced; by using the face recognition model, the method can accurately identify the identity of each staff member, thereby determining whether they have corresponding action authority; the action trajectories of all staff members in the sub-area are tracked by the multi-target tracking recognition model, and cross-area staff members are identified by the face recognition model, effectively solving the problem of monitoring cross-area behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the method of the present invention;

[0060] Figure 2 It is a flow chart of the method of time threshold judgment and frequency threshold judgment. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0062] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0063] like Figure 1 As shown, the abnormal behavior recognition method based on facial features of this embodiment includes the following steps:

[0064] S1: Divide the working area into multiple sub-areas;

[0065] S2: Determine the action authority of the operator relative to each sub-area and obtain the facial features of each operator;

[0066] S3: Obtain historical video image data of the working area and build a training database based on the historical video image data;

[0067] S4: Build a multi-target tracking and recognition model and train it through the training database; build a face recognition model;

[0068] S5: Acquire real-time video image data of each sub-area and identify each operator through a face recognition model;

[0069] S6: Determine whether the operators in each sub-area have the corresponding action authority. If not, it is identified as abnormal; if so, it is identified as normal;

[0070] S7: Track the movement trajectories of all operators in the sub-area through the multi-target tracking and recognition model, and determine whether the operators have cross-area behavior; if not, identify it as normal; if yes, identify the cross-area operators through the face recognition model;

[0071] S8: Determine whether the cross-zone operator has the action authority in the corresponding area. If not, it is identified as abnormal; if yes, it is identified as normal;

[0072] S9: Continue steps S7 to S8 to complete the identification of abnormal behaviors in the working area.

[0073] The method for constructing the training database includes:

[0074] Acquire historical video image data of the working area, and extract multiple video frames from the historical video image data;

[0075] Annotate the postures of multiple workers in the video frame to obtain the original picture; the posture annotation categories include: walking, standing, and various working postures corresponding to the work content;

[0076] Perform random rotation, flipping, random color channel conversion, random saturation transformation, and random chromaticity transformation on the original image to obtain a training image;

[0077] The training images are divided into a training set and a test set to obtain a training database.

[0078] like Figure 2 As shown, in step S7, time threshold determination and frequency threshold determination are also included, and the time threshold determination and frequency threshold determination include:

[0079] S71: Tracking the movement trajectories of all operators in the sub-area through a multi-target tracking and recognition model;

[0080] S72: Setting a time threshold T; setting a frequency threshold N, and setting the cross-zone frequency n=0;

[0081] S73: When it is detected that a certain operator moves from the atomic area to another sub-area, start timing and obtain the cross-area time t; set the operator as the first operator;

[0082] S74: Continue to track the first operator, and if the first operator does not move to the atomic area within T time, determine that the first operator has performed cross-area behavior;

[0083] If the first operator moves to the atomic region and t<T, it is determined that the first operator has not performed cross-region behavior, and n=n+1;

[0084] S75: Iteratively execute steps S73 to S74. If n<N, it is determined that the first operator has not performed cross-zone behavior; if n≥N, it is determined that the first operator has performed cross-zone behavior.

[0085] Among them, the method of building a multi-target tracking and recognition model includes:

[0086] Build a general model of the YOLOv5s algorithm and replace the first CBS layer at the input of YOLOv5s with a 3×3 normal convolution kernel and a 3×3 hole convolution;

[0087] Prune the BN layer in the CBS layer of the backbone network;

[0088] Add regularization constraints to the loss function of the general model of the YOLOv5s algorithm, and obtain the loss function corresponding to the modified model of the YOLOv5s algorithm;

[0089] Build tracking algorithm model;

[0090] The output of the YOLOv5s algorithm correction model is used as the input of the tracking algorithm model, and the tracking algorithm model outputs the moving trajectory.

[0091] Among them, the construction method of the tracking algorithm model includes:

[0092] Obtain the output of the YOLOv5s algorithm correction model as input; perform cascade matching on the input to obtain the first output; perform IOU matching on the first output to obtain the second output;

[0093] Perform Hungarian algorithm matching on the second output to obtain the third output;

[0094] Determine whether the trajectory matching is successful; if unsuccessful, perform Kalman filter prediction on the third output; if successful, output the third output as the final result, and perform Kalman filter update and Kalman filter prediction on the third output in sequence for continuous tracking.

[0095] Among them, the construction method of the face recognition model includes:

[0096] Construct the convolution layer, j out =(j in *str)+f size ;

[0097] Construct a pyramid pooling layer,

[0098] Construct a softmax classifier, whose loss function is:

[0099] Among them, j out is the output feature, j in is the input feature, str is the step size of the previous layer, f size is the size of the convolution kernel, h in is the input quantity, h out is the output, k h is the height of the core, p h is the number of fills, n is the number of pools, e is a natural constant, fy i is the target feature distribution, f i is the target true distribution, i is the index of the sample, and j is the index of the category.

[0100] The abnormal behavior recognition device based on facial features of this embodiment:

[0101] It includes a working area division module: dividing the working area into multiple sub-areas;

[0102] Action authority determination and facial feature acquisition module: determines the action authority of operators relative to each sub-area and acquires the facial features of each operator;

[0103] Training database construction module: obtain historical video image data of the work area, and construct a training database based on the historical video image data; the training database construction method includes: obtain historical video image data of the work area, and extract multiple video frames from the historical video image data; annotate the postures of multiple workers in the video frames to obtain original pictures; posture categories include: walking, standing, and various working postures corresponding to the work content; perform random rotation processing, flip processing, random color channel conversion, random saturation transformation, and random chromaticity transformation on the original pictures to obtain training pictures; divide the training pictures into training sets and test sets to obtain a training database;

[0104] Model building and training module: build a multi-target tracking and recognition model and train it through the training database; build a face recognition model;

[0105] Real-time video image acquisition and face recognition module: acquires real-time video image data of each sub-area and identifies each operator through the face recognition model;

[0106] Action authority judgment module: judge whether the operators in each sub-area have the corresponding action authority. If not, it is identified as abnormal; if so, it is identified as normal;

[0107] Cross-zone behavior tracking and identification module: The multi-target tracking and identification model is used to track the movement trajectories of all operators in the sub-zone, and determine whether the operators have cross-zone behavior; if not, it is identified as normal; if so, the cross-zone operators are identified through the face recognition model;

[0108] Cross-zone action authority judgment module: judge whether the cross-zone operator has the action authority in the corresponding area. If not, it is identified as abnormal; if so, it is identified as normal.

[0109] The electronic device of this embodiment includes a processor, a communication interface, a memory and a communication bus; wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; the processor is used to execute the programs stored in the memory to implement the above-mentioned abnormal behavior recognition method based on facial features.

[0110] The computer-readable storage medium of this embodiment is used to store computer-readable instructions, and when the instructions are executed, the above-mentioned abnormal behavior recognition method based on facial features is executed.

[0111] Example:

[0112] In a large oil extraction and production facility, we can use this technical solution to ensure the safety of employees while also ensuring the safe and efficient operation of the oil facility.

[0113] Division of work areas: We first divide the entire oil facility into multiple sub-areas, such as drilling area, storage area, processing facility, office area, etc.

[0114] Action authority determination: Then determine the action authority of each operator relative to each sub-area. For example, only drilling personnel can enter the drilling area, only warehouse managers can enter the storage area, and all employees can enter the office area.

[0115] Obtain facial features: Obtain facial feature information of all employees by installing facial recognition devices at the entrance of the facility.

[0116] Build a training database: Obtain historical video image data of the facility area and build a training database based on this data.

[0117] Build and train models: Build multi-target tracking and recognition models and face recognition models, and train them using a training database.

[0118] Real-time video image acquisition and face recognition: Install cameras in each sub-area to obtain real-time video image data of each sub-area, and identify each operator through the face recognition model.

[0119] Action authority judgment: Based on the face recognition results, it is judged whether the operating personnel in each sub-area have the corresponding action authority. If not, it is identified as abnormal; if so, it is identified as normal.

[0120] Cross-zone behavior tracking and identification: The movement trajectories of all operators in the sub-area are tracked through a multi-target tracking and identification model, and it is determined whether the operators have cross-zone behavior; if not, it is identified as normal; if so, the cross-zone operators are identified through a face recognition model.

[0121] Inter-zone action authority judgment: judge whether the inter-zone operator has the action authority in the corresponding area. If not, it is identified as abnormal; if so, it is identified as normal.

[0122] The system can help ensure that personnel only work in areas they are authorized to work in, thereby increasing the safety and efficiency of oil facilities. At the same time, the system can quickly detect and report to security personnel those who attempt to enter unauthorized areas.

[0123] The present invention has the following advantages:

[0124] 1. Improve the safety of the work area: By dividing the work area, determining the action rights of the operators relative to each sub-area, and using face recognition technology to determine the people entering each sub-area in real time, it is possible to effectively identify and prevent the entry and abnormal behavior of unauthorized personnel, thereby improving the safety of the work area.

[0125] 2. Automated monitoring and management: This method includes steps such as real-time video image acquisition, face recognition, authority judgment, and abnormal behavior identification, which greatly reduces the workload of manual monitoring and management and realizes automated monitoring and management of the work area.

[0126] 3. Improve the accuracy of abnormal behavior recognition: By building a multi-target tracking and recognition model and a face recognition model, and training with historical video image data, the accuracy of abnormal behavior recognition can be improved. At the same time, by setting time thresholds and frequency thresholds, it is possible to more accurately determine whether the operator has engaged in cross-zone behavior.

[0127] 4. Continuous monitoring: This method includes a continuous abnormal behavior identification step, which can continuously monitor the work area and identify and prevent abnormal behavior in real time.

[0128] In general, the main purpose of the present invention is to realize the automatic identification and management of abnormal behavior in the work area through the recognition of facial features and behavior analysis, thereby improving the safety of the work area, reducing the workload of manual management, and improving the accuracy of abnormal behavior identification.

[0129] The present invention is not limited to the above-mentioned optional implementation modes. Anyone can derive other various forms of products under the inspiration of the present invention. However, no matter what changes are made in the shape or structure, all technical solutions that fall within the scope defined by the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A method for identifying abnormal behavior based on facial features, characterized in that: The following steps are involved: S1: Divide the working area into multiple sub-areas; S2: Determine the action authority of the operator relative to each sub-area and obtain the facial features of each operator; S3: Obtain historical video image data of the working area and build a training database based on the historical video image data; S4: Build a multi-target tracking and recognition model and train it through the training database; build a face recognition model; S5: Acquire real-time video image data of each sub-area and identify each operator through a face recognition model; S6: Determine whether the operators in each sub-area have the corresponding action authority. If not, identify it as abnormal; If yes, it is identified as normal; S7: Track the movement trajectories of all operators in the sub-area through the multi-target tracking and recognition model, and determine whether the operators have cross-area behavior; if not, identify it as normal; If yes, the cross-zone operators are identified through the face recognition model; S8: Determine whether the cross-zone operator has the action authority in the corresponding area. If not, identify it as abnormal; If yes, it is identified as normal; S9: Continue steps S7 to S8 to complete the identification of abnormal behaviors in the working area.

2. The method for identifying abnormal behavior based on facial features according to claim 1, characterized in that: The method for constructing the training database includes: Acquire historical video image data of the working area, and extract multiple video frames from the historical video image data; Annotate the postures of multiple workers in the video frame to obtain the original picture; the posture annotation categories include: walking, standing, and various working postures corresponding to the work content; Perform random rotation, flipping, random color channel conversion, random saturation transformation, and random chromaticity transformation on the original image to obtain a training image; The training images are divided into a training set and a test set to obtain a training database.

3. The method for identifying abnormal behavior based on facial features according to claim 1, characterized in that: In step S7, time threshold determination and frequency threshold determination are also included, and the time threshold determination and frequency threshold determination include: S71: Tracking the movement trajectories of all operators in the sub-area through a multi-target tracking and recognition model; S72: Setting a time threshold T; setting a frequency threshold N, and setting the cross-zone frequency n=0; S73: When it is detected that a certain operator moves from the atomic area to another sub-area, start timing and obtain the cross-area time t; set the operator as the first operator; S74: Continue to track the first operator, and if the first operator does not move to the atomic area within T time, determine that the first operator has performed cross-area behavior; If the first operator moves to the atomic region and t<T, it is determined that the first operator has not performed cross-region behavior, and n=n+1; S75: Iteratively execute steps S73 to S74. If n<N, it is determined that the first operator has not performed cross-zone behavior; if n≥N, it is determined that the first operator has performed cross-zone behavior.

4. The method for identifying abnormal behavior based on facial features according to claim 1, characterized in that: Methods for building a multi-target tracking and recognition model include: Build a general model of the YOLOv5s algorithm and replace the first CBS layer at the input of YOLOv5s with a 3×3 normal convolution kernel and a 3×3 hole convolution; Prune the BN layer in the CBS layer of the backbone network; Add regularization constraints to the loss function of the general model of the YOLOv5s algorithm, and obtain the loss function corresponding to the modified model of the YOLOv5s algorithm; Build tracking algorithm model; The output of the YOLOv5s algorithm correction model is used as the input of the tracking algorithm model, and the tracking algorithm model outputs the moving trajectory.

5. The method for identifying abnormal behavior based on facial features according to claim 4, characterized in that: The construction method of the tracking algorithm model includes: Obtain the output of the YOLOv5s algorithm correction model as input; perform cascade matching on the input to obtain the first output; perform IOU matching on the first output to obtain the second output; Perform Hungarian algorithm matching on the second output to obtain the third output; Determine whether the trajectory matching is successful; if unsuccessful, perform Kalman filter prediction on the third output; if successful, output the third output as the final result, and perform Kalman filter update and Kalman filter prediction on the third output in sequence for continuous tracking.

6. The method for identifying abnormal behavior based on facial features according to claim 1, characterized in that: The construction method of the face recognition model includes: Construct the convolution layer, j out =(j in *str)+f size ; Construct a pyramid pooling layer, Construct a softmax classifier, whose loss function is: Among them, j out is the output feature, j in is the input feature, str is the step size of the previous layer, f size is the size of the convolution kernel, h in is the input quantity, h out is the output, k h is the height of the core, p h is the number of fills, n is the number of pools, e is a natural constant, fy i is the target feature distribution, f i is the target true distribution, i is the index of the sample, and j is the index of the category.

7. An abnormal behavior recognition device based on facial features, used in the abnormal behavior recognition method based on facial features according to any one of claims 1 to 6, characterized in that: It includes a working area division module: dividing the working area into multiple sub-areas; Action authority determination and facial feature acquisition module: determines the action authority of operators relative to each sub-area and acquires the facial features of each operator; Training database construction module: obtain historical video image data of the working area and construct a training database based on the historical video image data; Model building and training module: build a multi-target tracking and recognition model and train it through the training database; build a face recognition model; Real-time video image acquisition and face recognition module: acquires real-time video image data of each sub-area and identifies each operator through the face recognition model; Action authority judgment module: judges whether the operators in each sub-area have the corresponding action authority. If not, it is identified as abnormal; If yes, it is identified as normal; Cross-area behavior tracking and identification module: Track the movement trajectories of all operators in the sub-area through the multi-target tracking and identification model, and determine whether the operators have cross-area behavior; if not, it is identified as normal; If yes, the cross-zone operators are identified through the face recognition model; Cross-region action authority judgment module: judge whether the cross-region operator has the action authority of the corresponding area. If not, it is identified as abnormal; If yes, it is identified as normal.

8. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus; wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; the processor is used to execute the programs stored in the memory to implement the abnormal behavior recognition method based on facial features as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: Used to store computer-readable instructions, which, when executed, perform the abnormal behavior recognition method based on facial features as described in any one of claims 1 to 6.

Images