An abnormal intrusion monitoring method and related apparatus

Abstract

The application belongs to the field of data processing, and particularly relates to an abnormal intrusion monitoring method and related device, which comprises the following steps: in response to a warning instruction for a to-be-processed object, obtaining multi-dimensional intrusion monitoring information of the to-be-processed object; the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp; the intrusion detection unit is arranged at an electronic fence of the target camp; the intrusion detection unit comprises multiple information acquisition modules of different types; risk prediction of the multi-dimensional intrusion monitoring information is performed by using an intrusion risk prediction model to obtain risk prediction information of the to-be-processed object; dynamic intrusion disposal information is generated based on the risk prediction information; the dynamic intrusion disposal information at least comprises intrusion disposal scheduling information, an intrusion disposal mode and intrusion warning information. The method can perform risk prediction on multi-dimensional intrusion monitoring information, realize automatic intrusion monitoring of a to-be-identified object, improve abnormal intrusion monitoring efficiency and improve camp safety.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of data processing, and particularly relates to an abnormal intrusion monitoring method and related device. BACKGROUND

[0002] A camp is a base with multiple functions and multiple units. The camp usually includes various building facilities, such as a combat area, a weapon and equipment storage area, a dormitory, a mess hall, a medical facility, and the like, and is provided with measures such as warning, patrol, and entrance and exit inspection to ensure the safety of the camp.

[0003] In the related art, the safety warning of the camp is usually performed by manual patrol. However, the human senses, reaction speed, and ability are limited, and it is inevitable that there will be omissions or missed reports. Moreover, the camp buildings and facilities are relatively dispersed, and the coverage area of manual patrol is relatively limited, which cannot guarantee comprehensive and timely monitoring of all intrusion behaviors. In addition, although the defense facilities such as the fence and the sentry post enhance the safety of the camp, they cannot effectively prevent intrusion behaviors, and the defense facilities themselves are at risk of being broken or bypassed, making intrusion monitoring and security more difficult. It can be seen that the defense means of manual patrol and traditional structural defense facilities such as the fence and the sentry post in the related art are single, and cannot adapt to more and more complex intrusion means.

[0004] Therefore, in order to improve the safety of the camp, a new abnormal intrusion monitoring scheme is urgently needed. SUMMARY

[0005] The present application provides an abnormal intrusion monitoring method and related device to perform risk prediction on multi-dimensional intrusion monitoring information, realize automatic intrusion monitoring on the to-be-recognized object, improve the abnormal intrusion monitoring efficiency, and improve the safety of the camp.

[0006] In a first aspect, the present application provides an abnormal intrusion monitoring method, which comprises:

[0007] In response to a warning instruction for a to-be-processed object, multi-dimensional intrusion monitoring information of the to-be-processed object is acquired; the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp; the intrusion detection unit is deployed at an electronic fence of the target camp; the intrusion detection unit comprises multiple information collection modules of different types;

[0008] The risk prediction information of the to-be-processed object is obtained by performing risk prediction on the multi-dimensional intrusion monitoring information through an intrusion risk prediction model;

[0009] Dynamic intrusion disposal information is generated based on the risk prediction information; wherein the dynamic intrusion disposal information at least comprises: intrusion disposal scheduling information, an intrusion disposal mode, and intrusion warning information; the intrusion disposal scheduling information is used to indicate the disposal personnel and / or equipment required to be scheduled.

[0010] In a second aspect, the embodiments of the present application provide an abnormal intrusion monitoring device, which comprises:

[0011] An acquisition unit configured to acquire multi-dimensional intrusion monitoring information of the to-be-processed object in response to a pre-warning instruction for the to-be-processed object; the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp; the intrusion detection unit is deployed at an electronic fence of the target camp; the intrusion detection unit comprises a plurality of different types of information collection devices;

[0012] A risk prediction unit configured to perform risk prediction on the multi-dimensional intrusion monitoring information by using an intrusion risk prediction model to obtain risk prediction information of the to-be-processed object.

[0013] A generation unit configured to generate dynamic intrusion disposal information based on the risk prediction information; wherein the dynamic intrusion disposal information at least comprises intrusion disposal scheduling information, intrusion disposal mode, and intrusion warning information; the intrusion disposal scheduling information is used to indicate disposal personnel and / or equipment required to be dispatched.

[0014] In a third aspect, the embodiments of the present application provide a computing device, which comprises:

[0015] At least one processor, a memory, and an input-output unit;

[0016] The memory is configured to store a computer program, and the processor is configured to invoke the computer program stored in the memory to execute the abnormal intrusion monitoring method of the first aspect.

[0017] In a fourth aspect, a computer readable storage medium is provided, which comprises instructions, when the instructions are run on a computer, causing the computer to execute the abnormal intrusion monitoring method of the first aspect.

[0018] In the technical scheme provided by the embodiments of the present application, first, in response to a pre-warning instruction for a to-be-processed object, multi-dimensional intrusion monitoring information of the to-be-processed object is acquired. The multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp, the intrusion detection unit is deployed at an electronic fence of the target camp, and the intrusion detection unit includes multiple information acquisition modules of different types. In this way, by deploying the intrusion detection unit in the target camp and using information acquisition modules of different types, the multi-dimensional intrusion monitoring information can be more comprehensively acquired, and the perception ability of intrusion behavior can be improved. Further, the multi-dimensional intrusion monitoring information is predicted by an intrusion risk prediction model, and risk prediction information of the to-be-processed object is obtained. In this way, by predicting the multi-dimensional intrusion monitoring information by the intrusion risk prediction model, the intrusion risk level of the to-be-processed object can be quickly and accurately evaluated, more comprehensive risk prediction information can be provided, and the accuracy of intrusion risk monitoring can be improved. Finally, dynamic intrusion disposal information is generated based on the risk prediction information. The dynamic intrusion disposal information at least includes intrusion disposal scheduling information, an intrusion disposal method, and intrusion warning information. The intrusion disposal scheduling information is used to indicate the disposal personnel and / or equipment required to be dispatched. In this way, the response efficiency of risk prediction can be improved, and the security of the target camp can be improved. In particular, the delay caused by manual patrol can be avoided, and the processing efficiency of camp security problems can be improved.

[0019] In the technical scheme of the present application, the multi-dimensional intrusion monitoring information in the target camp is more comprehensively acquired, the perception ability of intrusion behavior is improved, the multi-dimensional intrusion monitoring information is processed by the intrusion risk prediction model, the detection and discrimination ability of intrusion behavior is improved, the response efficiency of intrusion disposal is effectively enhanced, the intrusion risk is reduced, and the security of the target camp is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0020] The accompanying drawings, which are included to provide a further understanding of the present application, form a part of the present application and illustrate the illustrative embodiments of the present application and the explanation of the present application, and do not constitute improper limitations on the present application. In the drawings:

[0021] Figure 1 is a flow diagram of an abnormal intrusion monitoring method according to an embodiment of the present application;

[0022] Figure 2 is a principle diagram of an intrusion risk prediction model according to an embodiment of the present application;

[0023] Figure 3 is a principle diagram of a multi-dimensional review network according to an embodiment of the present application;

[0024] Figure 4 is a structure diagram of an abnormal intrusion monitoring device according to an embodiment of the present application;

[0025] Figure 5 FIG. 1 is a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application.

[0028] A camp is a base for gathering multiple functions and units. The camp usually includes various buildings and facilities, such as dormitories, canteens, medical facilities, warehouses, etc., and is provided with measures such as guard, patrol, and entrance inspection to ensure the safety of the camp.

[0029] In the related art, the safety guard of the camp usually adopts manual patrol, but the human senses, reaction speed and ability are limited, and it is inevitable that there will be omissions or false reports. Moreover, the camp buildings and facilities are relatively dispersed, and the coverage area of manual patrol is relatively limited, which cannot guarantee comprehensive and timely monitoring of all intrusion behaviors. In addition, although the defense facilities such as fences and sentry posts enhance the safety of the camp, they cannot effectively prevent intrusion behaviors, and the defense facilities themselves are at risk of being broken or bypassed, making intrusion monitoring and security more difficult. It can be seen that the defense means of manual patrol and traditional structural defense facilities such as fences and sentry posts in the related art is single, and it is difficult to adapt to more and more complex intrusion means.

[0030] Therefore, in order to improve the safety of the camp, a new abnormal intrusion monitoring scheme is urgently needed.

[0031] To solve the above at least one technical problem, the embodiments of the present application provide an abnormal intrusion monitoring method and related device.

[0032] Specifically, in the abnormal intrusion monitoring scheme, first, in response to a warning instruction for a to-be-processed object, multi-dimensional intrusion monitoring information of the to-be-processed object is acquired. The multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp. The intrusion detection unit is deployed at an electronic fence of the target camp, and the intrusion detection unit includes multiple information acquisition modules of different types. In this way, by deploying the intrusion detection unit in the target camp and using information acquisition modules of different types, more comprehensive multi-dimensional intrusion monitoring information can be obtained, and the perception ability of intrusion behavior can be improved. Further, the multi-dimensional intrusion monitoring information is predicted by an intrusion risk prediction model to obtain risk prediction information of the to-be-processed object. In this way, the intrusion risk level of the to-be-processed object can be quickly and accurately evaluated by the intrusion risk prediction model to predict the risk of the multi-dimensional intrusion monitoring information, to provide more comprehensive risk prediction information, and to facilitate the improvement of the accuracy of intrusion risk monitoring. Finally, dynamic intrusion disposal information is generated based on the risk prediction information. The dynamic intrusion disposal information at least includes: intrusion disposal scheduling information, intrusion disposal mode, and intrusion warning information. The intrusion disposal scheduling information is used to indicate the disposal personnel and / or equipment required to be dispatched. In this way, the response efficiency of risk prediction can be improved, and the security of the target camp can be improved. In particular, it helps to avoid delays due to manual patrol, and improves the processing efficiency of camp security problems.

[0033] In the abnormal intrusion monitoring scheme, more comprehensive multi-dimensional intrusion monitoring information in the target camp is obtained, the perception ability of intrusion behavior is improved, and the multi-dimensional intrusion monitoring information is processed by the intrusion risk prediction model to improve the detection and discrimination ability of intrusion behavior, effectively enhance the response efficiency of intrusion disposal, reduce the intrusion risk, and ensure the security of the target camp.

[0034] The abnormal intrusion monitoring scheme provided by the embodiments of the present application can be executed by an electronic device, which can be a server, a server cluster, a cloud server. The electronic device can also be a terminal device such as a mobile phone, a computer, a tablet computer, a wearable device, or a special-purpose device (such as a special-purpose terminal device with an abnormal intrusion monitoring system, etc.). In an optional embodiment, a service program for executing the abnormal intrusion monitoring scheme can be installed on the electronic device.

[0035] Figure 1 A schematic diagram of an abnormal intrusion monitoring method provided by the embodiments of the present application is shown in FIG. 1, which includes the following steps: Figure 1

[0036] 101, in response to a warning instruction for a to-be-processed object, multi-dimensional intrusion monitoring information of the to-be-processed object is acquired.

[0037] ​In the embodiments of the present application, the to-be-processed object is a candidate intruding object monitored to enter the target camp area. Here, the candidate intruding object can be an individual object or a group object. It can be a person, specifically, the to-be-identified object can be a person or a group of people. It can also be a device, such as a drone, a vehicle, etc.

[0038] For example, if the target camp area is conducting intrusion monitoring, the warning instruction for the to-be-processed object can be a case where one or a group of suspicious personnel or illegal articles such as drones are found in a certain area in the target camp area. At this time, the to-be-processed object is a candidate intruding object monitored to enter the target camp area, which can be a single suspicious person or a drone, or a group of suspicious persons or drones. After monitoring a candidate intruding object, multi-dimensional intrusion monitoring information of the candidate intruding object can be obtained through the intrusion monitoring device, such as the position, speed, direction of travel, volume, weight, sound, image, etc. of the candidate intruding object, so as to further analyze and identify. Based on these multi-dimensional intrusion monitoring information, risk assessment and prediction can be performed on the candidate intruding object, and then dynamic intrusion disposal information is generated to assist security personnel to take corresponding intrusion disposal measures.

[0039] In the embodiments of the present application, the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in the target camp area. Specifically, the intrusion detection unit is deployed at the electronic fence of the target camp area. The intrusion detection unit includes a plurality of different types of information acquisition modules. Among them, the electronic fence of the target camp area is a safety facility for defining and protecting the boundary of the target area, which uses electronic technology for monitoring and control.

[0040] In one possible example, the electronic fence of the target camp can be composed of the following components: sonar sensors, photoelectric sensors, vibration sensors, video monitoring systems. Based on this, in 101, the electronic fence can be equipped with sonar sensors installed on the fence or around the ground. These sensors can monitor and identify the movement of objects or personnel through sound waves. When someone or something approaches or touches the fence, the sonar sensor will capture the relevant sound signal and trigger the alarm system. The electronic fence can also use photoelectric sensors installed on the fence in some or all areas. These sensors monitor activities around the fence by emitting an infrared light beam and detecting interruptions in the light beam. When someone or something passes through the infrared light beam, the photoelectric sensor detects the interruption of the light beam and sends an alarm. The electronic fence can also be equipped with vibration sensors installed on the fence structure. These sensors detect the touch or damage of external objects or personnel by sensing the vibration of the fence. When someone tries to climb over, hit or damage the fence, the vibration sensor will capture the vibration signal and trigger an alarm. The electronic fence can be combined with a video monitoring system, with cameras installed to cover key areas around the fence. These cameras can monitor the outside of the fence in real time and provide high-definition images or videos. Monitoring personnel can observe the video in real time through the monitoring center and take further action as needed. The above is an example of the electronic fence of the target camp. Through the integration of various sensors and monitoring technologies, the electronic fence can monitor and detect intrusion behavior around the fence in real time and provide multi-dimensional intrusion monitoring information to ensure the safety of the target area.

[0041] Based on the above example, the warning instruction for the object to be processed can be issued by the intrusion detection unit. The intrusion detection unit monitors the boundaries or specific areas of the target camp through different types of information collection modules. When suspicious intrusion behavior occurs, the intrusion detection unit can automatically trigger a warning instruction. These warning instructions can include but are not limited to the following cases:

[0042] Case 1: Trigger an alarm, the intrusion detection unit can trigger an alarm through sound, light flash, vibration, etc. to attract the attention of security personnel. This can alert the staff in the target camp in a timely manner and let them know that suspicious intrusion behavior is occurring.

[0043] Case 2: Send a notification, the intrusion detection unit can send a notification to designated security personnel through communication channels connected to the central management system. In this way, security personnel can quickly receive intrusion warning information and respond and handle accordingly.

[0044] Case 3: Trigger the monitoring video, the intrusion detection unit can be combined with the video monitoring system. When suspicious intrusion behavior is detected, the system can automatically obtain the relevant monitoring video from the camera and send it to the security personnel for real-time review and analysis.

[0045] Through these early warning instructions, the intrusion detection unit can timely alarm and notify relevant personnel, so as to quickly respond and cope, in order to protect the safety of the target camp.

[0046] In another example, the intrusion detection unit can be composed of several different types of information collection modules: high-definition cameras, infrared sensors, sound sensors, pressure sensors. Specifically, the intrusion detection unit can be equipped with high-definition cameras for real-time monitoring of the surrounding area of the target camp. These cameras can capture visual information of intrusion behavior, such as images of suspicious persons or vehicles, thereby providing visual monitoring information of the object to be processed. The intrusion detection unit can also contain infrared sensors for detecting infrared radiation of human bodies or other heat sources. When a person or animal enters the target area, the infrared sensor will respond, providing thermal signal monitoring information of the object to be processed. The intrusion detection unit can integrate sound sensors for detecting sound changes in the target area. For example, when someone knocks on the fence or makes an abnormal sound, the sound sensor can capture these signals and trigger an early warning, providing sound monitoring information of the object to be processed. The intrusion detection unit can also use pressure sensors, which are installed on the fence or the ground, for monitoring pressure changes in the target area. When a person or vehicle steps on or exerts pressure, the pressure sensor can detect these changes, providing pressure monitoring information of the object to be processed.

[0047] These information collection modules can be connected to the intrusion detection unit through wireless communication or wired connection, and real-time multi-dimensional intrusion monitoring information collected is sent to the central management system for processing and analysis, thereby providing comprehensive and multi-angle intrusion intelligence, helping to accurately judge the intrusion nature and risk level of the object to be processed.

[0048] In 102, the multi-dimensional intrusion monitoring information is risk predicted by the intrusion risk prediction model, and the risk prediction information of the object to be processed is obtained.

[0049] Specifically, the intrusion risk prediction model can be constructed based on historical data and machine learning algorithms. It will consider multiple factors such as the type of intrusion behavior, the time of intrusion, the behavior pattern of the intruder, and the related intrusion monitoring information. By analyzing and modeling these factors, the model can assess and predict the risk of the object to be processed. The prediction result of the intrusion risk prediction model can be a numerical value or a risk level, indicating the risk degree that the object to be processed may face. For example, it may predict three different risk levels: low, medium, and high, or the higher the score, the higher the risk. Through the intrusion risk prediction model, the purpose is to help the camp security personnel to take corresponding measures according to the risk prediction information of the object to be processed. For example, for high-risk objects to be processed, nearby security personnel or security equipment can be dispatched for response and emergency action, and for low-risk objects to be processed, timed patrol or enhanced monitoring can be performed.

[0050] Exemplarily, assuming that the intrusion risk prediction model of the target camp uses image recognition technology to identify whether a person has crossed the fence in a specific area. Then, the intrusion risk prediction model at least includes the following structures: feature extraction layer, hidden risk prediction layer, and prediction output layer. The hidden risk prediction layer includes multiple stacked connection layers, each of which processes the output of the previous layer to extract higher-level feature representations for subsequent risk prediction.

[0051] In this intrusion risk prediction model, the feature extraction layer can process the images captured by the surveillance camera to extract key features such as the outline of the pedestrian, the body posture, the gait, etc. In the next step, the first hidden risk prediction layer can learn some basic feature representations such as the presence of pedestrians, speed, direction, etc. from these features. Then in the subsequent hidden risk prediction layers, each connection layer can process the features from the previous layer to further extract higher-level features such as certain motion patterns, preferences, etc. of pedestrians in the visible area. In this way, the model can correspond the hidden risk prediction features of each layer to more abstract and more specific intrusion-related feature representations.

[0052] When the input data passes through all the hidden layers and finally reaches the output layer, the model can integrate the information of all the hidden layers to form a model output representing the risk degree of the object to be processed. This output can be a result of intrusion behavior classification or risk grading, such as unpredictable intrusion, low-risk, medium-risk, or high-risk intrusion.

[0053] As an optional embodiment, assuming that the intrusion risk prediction model at least includes the following structures: feature extraction layer, hidden risk prediction layer, and prediction output layer. Further, the hidden risk prediction layer includes multiple hidden risk prediction layers stacked in connection.

[0054] Based on the above structure, the multi-dimensional intrusion monitoring information is risk-predicted by the intrusion risk prediction model in 102 to obtain risk prediction information of the to-be-processed object, as shown in Figure 2 The method can be implemented as the following steps:

[0055] 201. Extracting, by a feature extraction layer of the intrusion risk prediction model, an intrusion behavior feature from the multi-dimensional intrusion monitoring information.

[0056] The intrusion behavior feature is extracted from the multi-dimensional intrusion monitoring information, such as the intrusion occurrence location and the intrusion time, which will be input to the feature extraction layer.

[0057] 202. Learning and extracting, by each hidden risk prediction layer of the intrusion risk prediction model, a first hidden risk feature from the intrusion behavior feature.

[0058] In the hidden risk prediction layer, the intrusion risk prediction model learns to extract risk features from the extracted features, and these hidden risk features may be related to the type of intrusion behavior, the behavior pattern of the intruder, etc.

[0059] 203. Performing, by each hidden risk prediction layer of the intrusion risk prediction model, feature fusion processing based on the second hidden risk feature output by the previous hidden risk prediction layer and the first hidden risk feature to obtain a third hidden risk feature.

[0060] In order to improve the prediction accuracy of the model, the hidden risk prediction layer can be connected with multiple hidden layers in a stacked manner, and each hidden layer will continue to learn useful feature representations from the previous layer. At each level of the hidden risk prediction layer, the model will perform feature fusion based on the features output by the previous layer and the input features of the current layer. For example, the output features of the previous hidden layer and the input features of the current layer can be spliced together as the input features of the current layer, or some other feature fusion method can be used.

[0061] 204. Performing, by a prediction output layer of the intrusion risk prediction model, risk prediction processing on the hidden risk features output by each of the multiple hidden risk prediction layers to obtain an intrusion risk prediction probability of the to-be-processed object.

[0062] After the feature extraction and feature fusion processing of the multiple hidden risk prediction layers, the model will input the features to the prediction output layer for classification or regression prediction, such as judging the intrusion risk level faced by the to-be-processed object.

[0063] In practical applications, as an optional embodiment, the process of obtaining the intrusion risk prediction probability is represented by the following expression:

[0064] y = σ(w(L+1) ·h (L) +b (L+1) )

[0065] wherein, y represents the intrusion risk prediction probability, σ(·) represents the expression of risk prediction processing, h (L) represents the hidden risk features output by the Lth hidden risk prediction layer, b (L+1) represents a bias scalar, w (L+1) represents the weight vector corresponding to the Lth hidden risk prediction layer.

[0066] As an optional embodiment, the training method of the intrusion risk prediction model can be implemented as follows: first, generate pseudo sample data matched with multi-dimensional intrusion monitoring information. These pseudo sample data can be generated by simulating various different types of intrusion behaviors and different scenes and environments to increase the diversity of the training samples of the model. Further, input the multi-dimensional intrusion monitoring information and the matched pseudo sample data into the intrusion risk prediction model, and train each hidden risk prediction layer in the intrusion risk prediction model from bottom to top. These data can include information of time, place, behavior pattern and other dimensions of intrusion. Then, obtain the contrast divergence difference between each multi-dimensional intrusion monitoring information and the matched pseudo sample data in each hidden risk prediction layer. The contrast divergence is a measure of the difference between the distributions of two data, which is used to measure the distance between the prediction result and the true result. Finally, adjust the model parameters in each hidden risk prediction layer based on the contrast divergence difference to complete the training of the intrusion risk prediction model. In this way, based on the contrast divergence difference, the model parameters in each hidden risk prediction layer are adjusted by optimization methods such as gradient descent. Through continuous iteration and adjustment, the model parameters are optimized, so that the model can more accurately predict the risk of the object to be processed.

[0067] Through the above training steps, the generated pseudo sample data can be used to train the intrusion risk prediction model, the model parameters are optimized, and the contrast divergence difference is used as a measurement index to gradually improve the prediction ability of the model, so as to improve the accuracy of the risk prediction of the object to be processed. This training method can increase the generalization ability of the model and improve the prediction performance of the model on unseen samples.

[0068] In 103, dynamic intrusion disposal information is generated based on risk prediction information.

[0069] In the embodiments of the present application, the dynamic intrusion disposal information at least includes: intrusion disposal scheduling information, intrusion disposal mode, intrusion warning information.

[0070] The intrusion disposal scheduling information is used to indicate the required disposal personnel and / or equipment. Specifically, the following intrusion disposal scheduling information can be generated. For example, disposal personnel scheduling information can be generated. Specifically, the most suitable security personnel is selected according to the risk prediction information to perform the disposal task. If it is predicted that there is a high-risk intrusion behavior in the area, the security personnel with disposal ability and close location can be indicated by the information to go to the area for disposal. For example, equipment scheduling information can be generated. Specifically, the related monitoring equipment can also be scheduled for disposal according to the risk prediction information. If it is predicted that there is a high-risk intrusion behavior in the area, the nearby camera can be scheduled for monitoring, and automatic processing operations such as video recording and image analysis can be performed. In this way, by generating the intrusion disposal scheduling information, the response time and effect of the intrusion disposal can be optimized. According to different intrusion risk prediction conditions, disposal resources can be reasonably allocated, and appropriate personnel and equipment can be scheduled for timely disposal to reduce the loss and threat caused by the intrusion.

[0071] In addition to the intrusion disposal scheduling information, the dynamic intrusion disposal information also includes intrusion disposal methods and intrusion warning information.

[0072] For the intrusion disposal method, it is assumed that the risk prediction model predicts that there is an intrusion risk in a certain area, and the corresponding intrusion disposal scheduling information has been generated. Next, an appropriate intrusion disposal method also needs to be selected according to the actual situation for processing. For example, the intrusion disposal method includes but is not limited to the following methods: enhancing physical security measures, such as strengthening the security access control of the perimeter of the area, increasing the security patrol density, etc.; strengthening technical security protection, such as enabling the intrusion detection system, strengthening the network firewall, updating the security patch, disabling the sensitive account, etc.; taking public security disposal measures, such as scheduling police forces, developing public security patrol plans, increasing patrol intensity, etc. According to the actual situation of the target camp, different intrusion disposal methods can be taken for processing, and corresponding disposal strategies can be established.

[0073] For the intrusion warning information, the intrusion warning information is mainly used to alert the user and provide related intrusion prediction and disposal suggestions. For example, the user can receive the intrusion warning information through a mobile phone, a computer, or other terminals, so as to timely understand the related situation and make correct disposal processing according to the warning suggestions. For example, the intrusion warning information includes but is not limited to: intrusion risk prediction, such as predicting the intrusion behavior that may occur in a certain area in the future; alarm information push, such as pushing the intrusion warning information and disposal suggestions to the key personnel in real time; related data information, such as providing intrusion history, similar security event cases, technical white paper, and other related data information, so that the user can better understand the intrusion situation and disposal method. Through the timely issuance and reasonable use of the intrusion warning information, the related security threat can be prevented in time, and the effectiveness of the intrusion disposal can be enhanced.

[0074] As an optional embodiment, before 103, a real-time positioning model can also be used to process the multi-dimensional intrusion monitoring information for trajectory positioning, to obtain the real-time action trajectory of the object to be processed.

[0075] Correspondingly, in 103, first, the real-time position of the object to be processed is determined according to the real-time action trajectory. Further, based on the intrusion behavior in the risk prediction information and the real-time position, the intrusion disposal scheduling route of the object to be processed is generated. Finally, the campsite position range covered by the intrusion disposal scheduling route in the target campsite is obtained, and the intrusion alarm information matched with the campsite position range is generated.

[0076] For example, assuming that in a certain target campsite, multiple intrusion monitoring devices such as cameras, sensors, etc. are deployed to monitor intrusion behavior. At the same time, the system also uses a real-time positioning model to process multi-dimensional intrusion monitoring information for trajectory positioning to obtain the real-time action trajectory of the object to be processed. Specifically, through the real-time positioning model, combined with the position data in the intrusion monitoring information, the real-time position of the object to be processed can be determined. This can be achieved through various positioning technologies such as GPS positioning, etc. Further, according to the intrusion behavior in the risk prediction information and the real-time position of the object to be processed, the intrusion disposal scheduling route can be generated. This includes indicating the disposal personnel and / or equipment that need to be dispatched, as well as the order and path of the processing action. Finally, by obtaining the campsite position range covered by the intrusion disposal scheduling route in the target campsite, the intrusion alarm information matched with the campsite position range can be generated. These alarm information can include the type of intrusion behavior, risk level, specific location, etc., and can be notified to relevant personnel through various ways such as mobile phone push, SMS notification, etc.

[0077] For example, assuming that the real-time positioning model monitors a potential intruder in a certain area of the target campsite, and according to the risk prediction information, judges that the intrusion behavior belongs to high risk. Then, an intrusion disposal scheduling route can be generated, indicating the nearest security personnel to go to the area for disposal, and ensuring to act according to the specified path. At the same time, intrusion alarm information is also generated, indicating the intrusion risk and specific location of the area, so that relevant personnel can take timely measures.

[0078] By using the real-time positioning model combined with the risk prediction information, the position and action trajectory of the object to be processed can be more accurately determined, and the corresponding intrusion disposal scheduling route and alarm information can be generated, to improve the efficiency and accuracy of intrusion disposal.

[0079] In the embodiments of the present application, by deploying intrusion detection units and using different types of information collection modules, multi-dimensional intrusion monitoring information in the target camp can be more comprehensively obtained to improve the perception ability of intrusion behavior. These intrusion detection units are deployed at the electronic fence of the target camp, which can timely detect intrusion behavior and collect various types of monitoring information through the information collection module. Next, by using an intrusion risk prediction model to predict the risk of multi-dimensional intrusion monitoring information, the intrusion risk level of the to-be-processed object can be quickly and accurately evaluated, and more comprehensive risk prediction information can be provided. Such risk prediction information can help decision makers better understand the intrusion risk and develop appropriate disposal strategies according to the risk level. Based on the risk prediction information, dynamic intrusion disposal information can be generated. Among them, the intrusion disposal scheduling information is used to indicate the disposal personnel and / or equipment that need to be dispatched in order to timely handle the intrusion behavior. The intrusion disposal mode will select appropriate security measures according to the specific situation, such as enhancing physical security and strengthening technical security protection. At the same time, the intrusion warning information will alert the relevant personnel and provide intrusion prediction, disposal suggestions and other information. By more comprehensively obtaining multi-dimensional intrusion monitoring information and combining the analysis of the intrusion risk prediction model, the detection and judgment ability of intrusion behavior can be improved, the response efficiency of intrusion disposal can be improved, and the intrusion risk can be reduced to ensure the safety of the target camp. In particular, for the delay problem existing in the traditional manual patrol mode, the technical means in the embodiments of the present application can avoid such limitations and improve the processing efficiency of security problems.

[0080] In the above or the following embodiments, the present application further provides an optimization correction method of risk prediction information.

[0081] Specifically, after 101, a multi-dimensional review network can be used to review the multi-dimensional intrusion monitoring information to obtain behavior risk review information of the to-be-processed object. The multi-dimensional review network includes a behavior review branch for processing different types of intrusion monitoring information. After 102, based on the behavior risk review information, a risk optimization model can be used to predict and correct the risk prediction information to obtain the optimized and corrected risk prediction information of the to-be-processed object.

[0082] Based on the above embodiments, the step of using a multi-dimensional review network to review different types of behavior data in the multi-dimensional intrusion monitoring information to obtain behavior risk review information of the to-be-processed object, as shown in Figure 3 , can be implemented as the following steps:

[0083] 301, pre-processing the multi-dimensional intrusion monitoring information to obtain intrusion monitoring information suitable for different behavior review branches;

[0084] 302, through the video review branch model in the multi-dimensional review network, the audio review branch model, the behavior trajectory branch review model, respectively identify the corresponding intrusion monitoring information, to obtain the intrusion behavior prediction information containing each branch review result.

[0085] Among them, the intrusion behavior prediction information is used to represent the probability prediction value of the to-be-processed object implementing various intrusion behaviors.

[0086] For example, assuming that a multi-dimensional monitoring device is deployed in a target camp, including video monitoring, audio monitoring, and behavior trajectory monitoring, etc. These monitoring devices will record and detect various intrusion behaviors, such as illegal intrusion of personnel, abnormal sound, etc. For the multi-dimensional review network, in the video review branch, the behavior of the to-be-processed object in the target camp can be analyzed by analyzing the video monitoring. For example, the video analysis algorithm can detect that the to-be-processed object wears camouflage clothes, carries suspicious articles, etc., and will give corresponding intrusion behavior prediction information, indicating the probability of the to-be-processed object involving illegal intrusion. In the audio review branch, the sound produced by the to-be-processed object can be analyzed by audio monitoring. For example, the sound analysis algorithm can detect abnormal noise, explosion sound, etc. made by the to-be-processed object, and will give corresponding intrusion behavior prediction information, indicating the probability of the to-be-processed object involving destructive behavior. In the behavior trajectory branch, the moving track and activity pattern of the to-be-processed object in the target camp can be analyzed by behavior trajectory monitoring. For example, the activity of the to-be-processed object in the non-working area can be identified, and corresponding intrusion behavior prediction information will be given, indicating the probability of the to-be-processed object involving border crossing behavior.

[0087] In this way, by reviewing different types of intrusion monitoring information through the multi-dimensional review network, comprehensive intrusion behavior prediction information can be obtained. These intrusion behavior prediction information represents the probability prediction value of the to-be-processed object implementing various intrusion behaviors, which helps to assess the risk and develop appropriate disposal measures. For example, if the intrusion behavior prediction information of the to-be-processed object shows a high probability, relevant alarms will be issued, and security personnel will be instructed to take appropriate action to handle the intrusion behavior.

[0088] By adopting the multi-dimensional review network and generating the intrusion behavior prediction information, the intrusion risk of the to-be-processed object can be more comprehensively and accurately assessed, and more accurate assessment basis can be provided to effectively deal with intrusion events and improve the security of the target camp.

[0089] As an optional embodiment, in 302, through the audio review branch model in the multi-dimensional review network, the corresponding intrusion monitoring information is identified to obtain the intrusion behavior prediction information containing each branch review result, which can be implemented as follows:

[0090] 3021, extracting a plurality of local acoustic features from the audio intrusion monitoring information by a multi-layer feature extraction layer of the audio recheck branch model.

[0091] For example, the original audio data (i.e., audio intrusion monitoring information) is subjected to Mel-Frequency Cepstral Coefficients (MFCC) feature extraction to obtain an audio feature sequence with different frequency and time characteristics. Mel-Frequency Cepstral Coefficients is a feature extraction method in the field of audio signal processing, which is used to convert frequency domain features into Mel-frequency scale cepstral coefficients. This conversion can better adapt to the characteristics of human hearing.

[0092] In the embodiments of the present application, shared convolution layers are used to set shared weights between the multi-layer feature extraction layers. In traditional network design, each feature extraction layer will have independent convolution kernel parameters, that is, the convolution kernels between different layers are independently trained. However, in the embodiments of the present application, shared convolution layers are selected to set shared weights. Specifically, in this example, shared weights can be set on the convolution layer of layer A, and these shared weights are used for convolution operations of layers B and C. This means that layers B and C use the same convolution kernel parameters as layer A for convolution operations, that is, they share the same convolution kernel. For example, assuming that a 3x3 convolution kernel parameter [w1, w2, w3; w4, w5, w6; w7, w8, w9] is used in the convolution layer of layer A, where wi represents the weight of the convolution kernel. Then, the same 3x3 convolution kernel parameter [w1, w2, w3; w4, w5, w6; w7, w8, w9] is also used in the convolution operations of layers B and C. In this way, the effect of sharing convolution layer weights can be achieved. Thus, the network parameter amount is reduced, the model efficiency is improved, and the features between different layers can be learned and represented together. By sharing weights, the feature information learned by all layers can be utilized, and the input changes can be better handled.

[0093] 3022, using capsule units to downsample the plurality of local acoustic features and re-group the plurality of local acoustic features to obtain a plurality of groups of local acoustic features by a Primary Capsule of the audio recheck branch model.

[0094] The Primary Capsule includes a plurality of capsule units, and each capsule unit can extract different local acoustic features and represent each feature as a correlation vector. The values on each vector represent the intensity and position information of the corresponding feature of each capsule unit.

[0095] For example, a group of local acoustic features is extracted by performing average pooling operations on the MFCC sequence to represent a certain sound pattern in the audio.

[0096] 3023, the digital capsule layer (Digit Capsule) of the audio review branch model transmits multiple groups of local acoustic features to the classification output layer through dynamic routing capsules to enhance the learning ability of local acoustic features.

[0097] Specifically, the capsule unit calculates the corresponding capsule output according to the input local acoustic features using a dynamic routing algorithm, while preserving spatial information and enhancing the learning ability of local acoustic features.

[0098] 3024, the classification output layer of the audio review branch model classifies the multiple groups of local acoustic features under each type of intrusion behavior after preliminary classification to obtain the intrusion behavior probability corresponding to the local acoustic features.

[0099] For example, the classification output layer can output probability prediction values of different intrusion behaviors (such as breaking door sound, breaking window sound, etc.), representing the possibility of each intrusion behavior occurring. Through these prediction probabilities, the intrusion behavior probability corresponding to the local acoustic features can be obtained.

[0100] For example, assume that the main capsule layer and the digital capsule layer are used for feature extraction and classification of audio signals. First, the MFCC method is used to extract features from the audio signals, resulting in a series of MFCC sequences. Each MFCC sequence is equivalent to an image and can be regarded as a small signal composed of multiple local acoustic features, each of which captures different acoustic properties. Then, each MFCC sequence is passed into the main capsule layer for processing. In the main capsule layer, each capsule unit receives input from multiple local acoustic features, learns their relationships using a dynamic routing mechanism, and generates a capsule output. By establishing connections between multiple capsule units, the capsule layer can aggregate local acoustic features to obtain more discriminative features. Specifically, the main capsule layer first inputs each MFCC sequence into a convolutional layer, which is used to capture different local acoustic features. Then, after pooling and nonlinear activation operations, each convolutional kernel produces a mapping of a small signal. For example, assume that three 3x3 convolutional kernels are used, each of which produces a 7x7 mapping. At this point, a feature tensor composed of three mappings is obtained. Next, the main capsule layer is used to process this feature tensor. In the main capsule layer, each capsule unit down-samples small signals with similar features and reorganizes them to generate a set of capsule outputs. Unlike traditional convolutional operations, capsule units not only learn the features of the input but also learn the relationships between them. The use of this dynamic routing mechanism helps improve the adaptability and generalization ability of the model. Finally, in the output of the main capsule layer, each capsule unit corresponds to an acoustic pattern, such as the sound of a vehicle entering, the sound of breaking a window, etc. These patterns have more abstract and robust feature representations that can better distinguish different acoustic events. In summary, the main capsule layer is an important level for down-sampling and encoding multiple local acoustic features, which helps to extract more discriminative and abstract features. Through the processing of the main capsule layer, acoustic events in the audio signal can be better understood, and more accurate classification and prediction can be performed.

[0101] In the above example, the main capsule layer can be regarded as an adaptive feature extractor. Through the down-sampling and reorganization operations of the capsule units, the main capsule layer can capture higher-dimensional acoustic features and convert them into different correlation vectors for further classification. Such a dynamic routing mechanism allows the model to adaptively extract local acoustic features rather than being limited to fixed pre-defined feature extraction.

[0102] Further, let's introduce the digital capsule layer again. For the digital capsule layer in the audio review branch model, in the above camp example, it can be applied to the representation learning and coding of sound patterns. Suppose there is an audio review branch model that reviews the speech signal to determine whether there is abnormal sound. First, the MFCC method can be used to extract features from the audio signal to obtain a series of MFCC sequences. Each MFCC sequence can be regarded as a small signal segment composed of multiple local acoustic features, and each feature reflects different properties of the sound. Then, these MFCC sequences are input into the digital capsule layer for feature learning and representation coding. In the digital capsule layer, each capsule unit can encode each MFCC sequence and learn the relationship between sequences to form the representation of sound patterns. The capsule unit here can be regarded as a low-level acoustic part extracted from the audio signal, and a set of weights is assigned to each part. Through the processing of the digital capsule layer, a set of capsule outputs representing the audio sound patterns can be obtained. Each capsule output represents a sound pattern, such as normal human voice, environmental noise, high-decibel sound, etc. These capsule outputs not only contain the feature information of the sound, but also contain the relationship between the sounds. Finally, these capsule outputs can be used for abnormal sound detection and classification. The capsule outputs can be connected to the next layer, such as a fully connected layer or a classifier, to determine whether there is abnormal sound in the audio signal. In this step, the results of the capsule output represent the confidence of different sound patterns, which can be used to determine whether there is abnormal sound. By adding the digital capsule layer, the audio review branch model can better understand the sound patterns in the audio signal and convert them into interpretable feature coding. Such a design can improve the performance and interpretability of the audio review model, providing more accurate and reliable results for sound anomaly detection and classification tasks.

[0103] Another optional embodiment, for the target camp's intrusion monitoring task, the video review branch model involved in 302 can be implemented as follows:

[0104] Input layer: The input of the video composite branch model is a video sequence data, such as a video stream captured by multiple surveillance cameras.

[0105] Convolutional layer: Use convolutional operation to extract features from each video frame, extract local spatial features and dynamic temporal features in the video. For example, since the surveillance camera usually captures video of the same area, a pre-trained convolutional kernel such as VGG16, Inception, etc. video composite branch model can be used for feature extraction.

[0106] Temporal convolution layer: To better learn the temporal features of video sequences, a temporal convolution layer can be added before each ResNet residual block. The size of the temporal convolution layer can be set to (1, 3), indicating that only convolution operations are performed in the time dimension, preserving the temporal information of the input data.

[0107] Residual block layer: Residual blocks are used to learn higher-order feature representations in depth. Each residual block consists of multiple convolution layers and batch normalization layers, which reduce the gradient vanishing problem and speed up the training process through skip connections. In each residual block, temporal convolution layers and convolution layers appear alternately to effectively learn temporal and spatial features.

[0108] Global average pooling layer: The feature data is globally averaged and pooled into a fixed-size feature vector. This reduces the number of parameters, avoids overfitting, and preserves important feature information.

[0109] Spatial convolution layer: Several spatial convolution layers are added at the end of the network to increase the receptive field and extract spatial features. These spatial convolution layers operate on the feature map, further improving the performance and accuracy of the video composite branch model.

[0110] Fully connected layer: Finally, the global average pooling layer and the output layer are connected to output classification probabilities. The fully connected layer can use the softmax function to map features to probabilities of normal or abnormal classes. If there are more subdivisions of intrusion behavior, the number of classes in the output layer can be increased.

[0111] The above structure can implement the video composite branch model, which can be used for intrusion monitoring tasks in the target camp by combining the background of the target camp. Through the video composite branch model, it can detect whether there is abnormal behavior in the video stream in real time, such as crawling, climbing, crossing boundaries, etc., and timely alarm relevant personnel. The video composite branch model can play an important role in ensuring the safety of the target camp.

[0112] In another optional embodiment, the behavior trajectory branch review model involved in 302 for the intrusion monitoring task of the target camp can be implemented as follows:

[0113] Input layer: The input of the behavior trajectory branch review model is a behavior trajectory sequence data, such as the behavior trajectory recorded by the sensors (such as video monitoring, infrared sensors, etc.) in the target camp.

[0114] Position encoding layer: To add position information to sequence data, a position encoding layer can be used. The position encoding layer adds position information to each time step of the behavior trajectory in the sequence, so that the behavior trajectory branch review model can learn the context relationship between behavior trajectories.

[0115] Stacked Transformer encoding layer: The encoding part of the Transformer behavior trajectory branch review model is used to learn the global information of the input sequence and generate the representation of the corresponding behavior trajectory. The Transformer behavior trajectory branch review model can learn the global relationship in the sequence through the self-attention mechanism (Self-Attention), thereby effectively capturing the temporal dependency and contextual relationship between behavior trajectories.

[0116] Classification output layer: A classification output layer is added at the end of the behavior trajectory branch review model, which uses a Softmax layer to map the representation of the behavior trajectory to the probability of different categories. Specifically, the output layer can be set to multiple categories (such as normal behavior, abnormal behavior, etc.) according to task requirements, for classifying behavior trajectories.

[0117] Through the above structure, a behavior trajectory branch review model can be realized, which analyzes and classifies the recorded behavior trajectories in combination with the background of the target camp. For example, in the intrusion monitoring task of the target camp, the behavior trajectory branch review model can learn the features of normal behavior trajectories and identify different types of abnormal behavior trajectories, such as running, sneaking, and interacting. The behavior trajectory branch review model can output the class probability of each behavior trajectory, thereby providing accurate intrusion behavior prediction information. Through the behavior trajectory branch review model, real-time monitoring and analysis of behavior in the target camp can be achieved, improving the accuracy and efficiency of intrusion detection and further ensuring the safety of the target camp.

[0118] In another embodiment of the present application, an abnormal intrusion monitoring device is also provided, as shown in Figure 4 The device includes the following units:

[0119] The acquisition unit is configured to acquire multi-dimensional intrusion monitoring information of the to-be-processed object in response to a warning instruction for the to-be-processed object; the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in the target camp; the intrusion detection unit is deployed at an electronic fence of the target camp; the intrusion detection unit includes multiple different types of information collection devices;

[0120] The risk prediction unit is configured to perform risk prediction on the multi-dimensional intrusion monitoring information through an intrusion risk prediction model to obtain risk prediction information of the to-be-processed object;

[0121] The generation unit is configured to generate dynamic intrusion disposal information based on the risk prediction information; wherein the dynamic intrusion disposal information at least includes: intrusion disposal scheduling information, intrusion disposal mode, and intrusion warning information; the intrusion disposal scheduling information is used to indicate the required disposal personnel and / or equipment.

[0122] Further optionally, the device further comprises a review unit configured to, after the acquisition unit acquires the multi-dimensional intrusion monitoring information of the to-be-processed object, review the multi-dimensional intrusion monitoring information by using a multi-dimensional review network to obtain behavior risk review information of the to-be-processed object; wherein the multi-dimensional review network comprises a behavior review branch for processing different types of intrusion monitoring information.

[0123] The review unit is further configured to, after the risk prediction unit predicts the risk of the multi-dimensional intrusion monitoring information by using an intrusion risk prediction model to obtain risk prediction information of the to-be-processed object, correct the risk prediction information by using a risk optimization model based on the behavior risk review information to obtain the risk prediction information of the to-be-processed object after optimization correction.

[0124] Further optionally, the review unit reviews different types of behavior data in the multi-dimensional intrusion monitoring information by using a multi-dimensional review network to obtain behavior risk review information of the to-be-processed object, and is specifically configured to:

[0125] preprocess the multi-dimensional intrusion monitoring information to obtain intrusion monitoring information suitable for different behavior review branches;

[0126] identify corresponding intrusion monitoring information by using a video review branch model, an audio review branch model, and a behavior trajectory branch review model in the multi-dimensional review network to obtain intrusion behavior prediction information containing review results of each branch;

[0127] The intrusion behavior prediction information is used to represent a probability prediction value of the to-be-processed object implementing various intrusion behaviors.

[0128] Further optionally, the review unit identifies corresponding intrusion monitoring information by using a video review branch model, an audio review branch model, and a behavior trajectory branch review model in the multi-dimensional review network to obtain intrusion behavior prediction information containing review results of each branch, and is specifically configured to:

[0129] extract a plurality of local acoustic features from the audio intrusion monitoring information by using a multi-layer feature extraction layer of the audio review branch model; and share weights are set between the multi-layer feature extraction layers by using a shared convolution layer;

[0130] use a main capsule layer of the audio review branch model to down-sample the plurality of local acoustic features by using a capsule unit, and re-group the plurality of local acoustic features to obtain a plurality of groups of local acoustic features;

[0131] The digital capsule layer of the audio review branch model is used to transmit multiple groups of local acoustic features to a classification output layer through dynamic routing capsules, so as to enhance the learning ability of the local acoustic features;

[0132] The classification output layer of the audio review branch model is used to perform classification probability prediction on the multiple groups of local acoustic features under each type of intrusion behavior after preliminary classification, so as to obtain the intrusion behavior probability corresponding to the local acoustic features.

[0133] Further optionally, the device further comprises a real-time positioning unit configured to:

[0134] Before the generation unit generates the dynamic intrusion disposal information based on the risk prediction information, a real-time positioning model is used to perform trajectory positioning processing on the multi-dimensional intrusion monitoring information, so as to obtain a real-time action trajectory of the to-be-processed object;

[0135] The generation unit is specifically configured to: determine a real-time position of the to-be-processed object according to the real-time action trajectory; generate an intrusion disposal scheduling route of the to-be-processed object based on the intrusion behavior in the risk prediction information and the real-time position; obtain a camp position range covered by the intrusion disposal scheduling route in a target camp, and generate intrusion alarm information matched with the camp position range.

[0136] Further optionally, the intrusion risk prediction model at least comprises the following structures: a feature extraction layer, a hidden risk prediction layer, and a prediction output layer; the hidden risk prediction layer comprises a plurality of stacked hidden risk prediction layers. The risk prediction unit is specifically configured to:

[0137] The feature extraction layer of the intrusion risk prediction model is used to extract an intrusion behavior feature from the multi-dimensional intrusion monitoring information;

[0138] Each hidden risk prediction layer of the intrusion risk prediction model is used to learn and extract a first hidden risk feature from the intrusion behavior feature; and a feature fusion processing is performed based on a second hidden risk feature output by a previous hidden risk prediction layer and the first hidden risk feature, so as to obtain a third hidden risk feature;

[0139] The prediction output layer of the intrusion risk prediction model is used to perform risk prediction processing on the hidden risk features output by each of the plurality of hidden risk prediction layers, so as to obtain an intrusion risk prediction probability of the to-be-processed object.

[0140] Further optionally, the processing process of obtaining the intrusion risk prediction probability is represented by the following expression:

[0141] y = σ (w (L+1) ·h (L) +b (L+1) )

[0142] wherein, y represents the intrusion risk prediction probability, σ(·) represents the expression of risk prediction processing, h (L) represents the hidden risk feature output by the Lth hidden risk prediction layer, b (L+1) represents the bias scalar, w (L+1) represents the weight vector corresponding to the Lth hidden risk prediction layer.

[0143] Further optionally, the device further comprises a training unit configured to:

[0144] generate pseudo sample data matched with the multi-dimensional intrusion monitoring information;

[0145] input the multi-dimensional intrusion monitoring information and the matched pseudo sample data into the intrusion risk prediction model, and train each hidden risk prediction layer in the intrusion risk prediction model from bottom to top;

[0146] obtain the contrast divergence difference between each multi-dimensional intrusion monitoring information and the matched pseudo sample data in each hidden risk prediction layer;

[0147] adjust the model parameters in each hidden risk prediction layer based on the contrast divergence difference, to complete the training of the intrusion risk prediction model.

[0148] In the embodiments of the present application, the multi-dimensional intrusion monitoring information in the target camp is more comprehensively obtained through the abnormal intrusion monitoring device, the perception ability of the intrusion behavior is improved, and the detection and discrimination ability of the intrusion behavior is improved through the intrusion risk prediction model processing the multi-dimensional intrusion monitoring information, the response efficiency of the intrusion disposal is effectively enhanced, the intrusion risk is reduced, and the safety of the target camp is ensured.

[0149] In another embodiment of the present application, an electronic device is also provided, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete communication with each other through the communication bus;

[0150] The memory is used to store a computer program.

[0151] The processor is used to execute the program stored on the memory, and realize the abnormal intrusion monitoring method provided in the method embodiments.

[0152] The communication bus 1140 mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc.

[0153] For ease of representation, Figure 5 Only one thick line is used in the middle, but it does not mean that there is only one bus or one type of bus.

[0154] The communication interface 1120 is used for communication between the above electronic device and other devices.

[0155] The memory 1130 can include a Random Access Memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0156] The processor 1110 mentioned above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0157] Correspondingly, the embodiment of the present application also provides a computer readable storage medium storing a computer program, and the computer program is executed to implement each step that can be executed by the electronic device in the above method embodiment.

Claims

1. An abnormal intrusion monitoring method characterized by comprising: The method comprises the steps of: in response to a pre-warning instruction for a to-be-processed object, acquiring multi-dimensional intrusion monitoring information of the to-be-processed object; the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp; the intrusion detection unit is deployed at an electronic fence of the target camp; the intrusion detection unit comprises a plurality of different types of information acquisition modules; using a multi-dimensional review network, the multi-dimensional intrusion monitoring information is reviewed and processed to obtain behavior risk review information of the to-be-processed object, wherein the multi-dimensional review network comprises a behavior review branch for processing different types of intrusion monitoring information; including: pre-processing the multi-dimensional intrusion monitoring information to obtain intrusion monitoring information suitable for different behavior review branches; through the video review branch model, the audio review branch model, and the behavior trajectory branch review model in the multi-dimensional review network, the corresponding intrusion monitoring information is identified respectively to obtain intrusion behavior prediction information containing the review results of each branch; wherein the intrusion behavior prediction information is used to represent the probability prediction value of the to-be-processed object implementing various intrusion behaviors; by using an intrusion risk prediction model, the multi-dimensional intrusion monitoring information is predicted to obtain risk prediction information of the to-be-processed object, including: based on the behavior risk review information, using a risk optimization model to predict and correct the risk prediction information to obtain the risk prediction information of the to-be-processed object after optimization and correction; based on the risk prediction information, dynamic intrusion disposal information is generated; wherein the dynamic intrusion disposal information at least includes: intrusion disposal scheduling information, intrusion disposal mode, and intrusion warning information; the intrusion disposal scheduling information is used to indicate the disposal personnel and / or equipment required to be dispatched.

2. The abnormal intrusion monitoring method of claim 1, wherein, The method comprises the steps of: through the multi-layer feature extraction layer of the audio review branch model, a plurality of local acoustic features are extracted from the audio intrusion monitoring information; the shared convolutional layer is used to set shared weights between the multi-layer feature extraction layers; through the main capsule layer of the audio review branch model, the plurality of local acoustic features are down-sampled and re-grouped using capsule units to obtain a plurality of groups of local acoustic features; through the digital capsule layer of the audio review branch model, the plurality of groups of local acoustic features are transmitted to the classification output layer through dynamic routing capsules to enhance the learning ability of the local acoustic features; through the classification output layer of the audio review branch model, the plurality of groups of local acoustic features after preliminary classification are classified under each type of intrusion behavior to obtain the intrusion behavior probability corresponding to the local acoustic features.

3. The abnormal intrusion monitoring method of claim 1, wherein, Before the dynamic intrusion disposal information is generated based on the risk prediction information, the method further comprises the steps of: using a real-time positioning model, the multi-dimensional intrusion monitoring information is processed to obtain the real-time action trajectory of the to-be-processed object; the dynamic intrusion disposal information is generated based on the risk prediction information, comprising: determining a real-time position of the object to be processed according to the real-time action track; generating an intrusion disposal dispatch route of the object to be processed based on the intrusion behavior in the risk prediction information and the real-time position; obtaining a campsite position range covered by the intrusion disposal dispatch route in a target campsite, and generating intrusion alarm information matched with the campsite position range.

4. The abnormal intrusion monitoring method of claim 1, wherein, The intrusion risk prediction model at least includes the following structures: a feature extraction layer, a hidden risk prediction layer, and a prediction output layer; the hidden risk prediction layer includes a plurality of stacked hidden risk prediction layers; The risk prediction of the multi-dimensional intrusion monitoring information by the intrusion risk prediction model includes: extracting intrusion behavior features from the multi-dimensional intrusion monitoring information by the feature extraction layer of the intrusion risk prediction model; learning and extracting first hidden risk features from the intrusion behavior features by each hidden risk prediction layer of the intrusion risk prediction model; performing feature fusion processing based on second hidden risk features output by a previous hidden risk prediction layer and the first hidden risk features to obtain third hidden risk features; performing risk prediction processing on the hidden risk features output by each of the plurality of hidden risk prediction layers by the prediction output layer of the intrusion risk prediction model to obtain an intrusion risk prediction probability of the object to be processed.

5. The abnormal intrusion monitoring method according to claim 4, characterized by, The processing process of the intrusion risk prediction probability is expressed as the following expression: wherein, denotes the intrusion risk prediction probability, denotes an expression of the risk prediction process, denotes a hidden risk feature output by the Lth hidden risk prediction layer, denotes a bias scalar, denotes a weight vector corresponding to the Lth hidden risk prediction layer.

6. The abnormal intrusion monitoring method of claim 1, wherein, The training method of the intrusion risk prediction model further includes: generating pseudo sample data matched with the multi-dimensional intrusion monitoring information; inputting the multi-dimensional intrusion monitoring information and the matched pseudo sample data into the intrusion risk prediction model, and training each hidden risk prediction layer in the intrusion risk prediction model from bottom to top; obtaining a contrast divergence difference between each multi-dimensional intrusion monitoring information and the matched pseudo sample data in each hidden risk prediction layer; adjusting model parameters in each hidden risk prediction layer based on the contrast divergence difference to complete the training of the intrusion risk prediction model.

7. An abnormal intrusion monitoring apparatus characterized by comprising: The device includes: An acquisition unit is configured to acquire multi-dimensional intrusion monitoring information of a to-be-processed object in response to a pre-warning instruction for the to-be-processed object; the multi-dimensional intrusion monitoring information is monitored by an intrusion detection unit in a target camp; the intrusion detection unit is deployed at an electronic fence of the target camp; the intrusion detection unit includes multiple different types of information collection devices; a multi-dimensional review network is used to review and process the multi-dimensional intrusion monitoring information to obtain behavior risk review information of the to-be-processed object, including: pre-processing the multi-dimensional intrusion monitoring information to obtain intrusion monitoring information adapted to different behavior review branches; through a video review branch model, an audio review branch model, and a behavior trajectory branch review model in the multi-dimensional review network, corresponding intrusion monitoring information is identified respectively to obtain intrusion behavior prediction information containing review results of each branch; wherein the intrusion behavior prediction information is used to represent probability prediction values of various intrusion behaviors implemented by the to-be-processed object; wherein the multi-dimensional review network includes behavior review branches for processing different types of intrusion monitoring information. A risk prediction unit is configured to perform risk prediction on the multi-dimensional intrusion monitoring information through an intrusion risk prediction model to obtain risk prediction information of the to-be-processed object, including: based on the behavior risk review information, a risk optimization model is used to perform prediction correction processing on the risk prediction information to obtain optimized and corrected risk prediction information of the to-be-processed object. A generation unit is configured to generate dynamic intrusion disposal information based on the risk prediction information; wherein the dynamic intrusion disposal information at least includes: intrusion disposal scheduling information, intrusion disposal mode, and intrusion pre-warning information; the intrusion disposal scheduling information is used to indicate disposal personnel and / or equipment required to be scheduled.

8. A computing device, comprising: The computing device includes: at least one processor, a memory, and an input-output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the abnormal intrusion monitoring method according to any one of claims 1 to 6.

Images