Scene-based monitoring model customization method, equipment and medium

Abstract

The invention provides a scene-based monitoring model customization method and device and a medium, and belongs to the technical field of security monitoring, and the scene-based monitoring model customization method comprises the steps: obtaining the safety demands and data features of different monitoring places, and according to the safety demands and data features of the monitoring places, determining the security demands and data features of the monitoring places; designing monitoring models of different monitoring places according to a customization strategy; according to the attention mechanism, respectively training corresponding monitoring models for different monitoring places; monitoring effect data of the monitoring model are collected in real time, and monitoring parameters of the monitoring model are adjusted in real time according to the monitoring effect data and safety requirements and data features of the monitoring place. According to the technical scheme, the problems that in the prior art, when a unified model is adopted for monitoring, targeted event monitoring cannot be carried out according to different monitoring scenes, and a generalization model cannot improve the monitoring accuracy of a specific scene can be solved.

Description

Technical Field

[0001] This application belongs to the field of security monitoring technology, specifically involving a method, device, and medium for customizing a scene-based monitoring model. Background Technology

[0002] A security monitoring system transmits video signals within a closed loop and constitutes an independent and complete system from camera to image display and recording. It can reflect the monitored object in real time, vividly, and realistically, replacing long-term manual surveillance and recording the footage via video recorders.

[0003] Typically, a security monitoring system consists of multiple cameras installed in different monitoring locations. Each camera extracts video data from its respective location and then uploads it to a unified model for processing. This allows for the analysis of monitored objects and events at different locations.

[0004] However, when existing technologies use a unified model for monitoring, they cannot perform targeted event monitoring based on different monitoring locations, and generalized models cannot improve the monitoring accuracy of specific scenarios. Summary of the Invention

[0005] The technical solution described in this application aims to provide a scenario-based monitoring model customization method, device, and medium, which can solve the problems of existing technologies that, when using a unified model for monitoring, cannot perform targeted event monitoring according to different monitoring scenarios, and that generalized models cannot improve the monitoring accuracy of specific scenarios.

[0006] According to a first aspect of this application, embodiments of this application provide a method for customizing a scenario-based monitoring model, comprising: Obtain the security requirements and data characteristics of different monitoring locations, and design monitoring models for different monitoring locations according to customized strategies based on the security requirements and data characteristics of the monitoring locations; Based on the attention mechanism, corresponding monitoring models are trained for different monitoring locations. The system collects monitoring performance data from the monitoring model in real time and adjusts the monitoring parameters of the model based on the monitoring performance data, the security requirements of the monitoring site, and the data characteristics.

[0007] Preferably, in the above-mentioned monitoring model customization method, the security requirements and data characteristics of different monitoring locations are obtained, and monitoring models for different monitoring locations are designed according to the security requirements and data characteristics of the monitoring locations and a customization strategy, including: Through scene understanding algorithms, the types and behaviors of objects in the monitored locations can be identified from the surveillance videos. Based on object type and object behavior, the location type of the monitored location is determined through spatiotemporal context. Based on the security characteristics of the venue type and combined with the spatiotemporal context, the security requirements of the monitored venue are mapped, and the data characteristics are summarized. The system acquires real-time surveillance videos of monitored locations, learns the contextual relationships, and updates security requirements and data characteristics based on these contextual relationships and the security characteristics of the location type.

[0008] Preferably, in the above-mentioned monitoring model customization method, the security requirements and data characteristics of different monitoring locations are obtained, and monitoring models for different monitoring locations are designed according to the security requirements and data characteristics of the monitoring locations and a customization strategy, including: Based on the security requirements of the monitored location, select a matching monitoring model from the model library; Based on the data characteristics of the monitored locations, set the model parameters and loss function of the monitoring model; According to the monitoring standards of the monitoring model, the monitoring video acquired by the monitoring equipment is preprocessed to obtain image frame data suitable for the monitoring model; Design a feedback closed-loop model for the monitoring model. The feedback closed-loop model includes alarm threshold setting, human-computer interaction design, and online learning mechanism. Use the feedback closed-loop model to perform closed-loop evaluation and feedback on the monitoring effect data of the monitoring model, so as to adjust the model parameters and loss function of the monitoring model.

[0009] Preferably, in the above-mentioned monitoring model customization method, based on the attention mechanism, corresponding monitoring models are trained for different monitoring locations, including: Based on the security needs and data characteristics of different monitoring locations, determine the corresponding monitoring areas and monitoring objects; And according to security requirements and data characteristics, match the corresponding attention weights for the monitoring area and the monitoring object; Using attention weights, calculate the attention values ​​that the monitoring model matches for the monitored area and the monitored object, respectively; Based on the attention value, the collected surveillance video is used to train the surveillance model and obtain surveillance effect data.

[0010] Preferably, in the above-mentioned monitoring model customization method, the monitoring effect data of the monitoring model is collected in real time, and the monitoring parameters of the monitoring model are adjusted in real time according to the monitoring effect data, the security requirements of the monitoring site, and the data characteristics, including: The monitoring effect data of the monitoring model is compared with the security requirements of the monitoring site to obtain the monitoring effect error. By combining data characteristics and monitoring parameters, root cause analysis is performed on the monitoring effect error to obtain the root cause analysis results, and corresponding negative feedback adjustment strategies are matched to the root cause analysis results. The monitoring parameters of the monitoring model are adjusted using a negative feedback adjustment strategy.

[0011] Preferably, in the above-mentioned monitoring model customization method, the root cause analysis of the monitoring effect error is performed by combining data characteristics and monitoring parameters to obtain the root cause analysis results, and the corresponding negative feedback adjustment strategy is matched for the root cause analysis results, including: By calling the model knowledge graph and combining it with data features, a multi-dimensional vector corresponding to the monitoring effect error is constructed. Multi-dimensional vector matching is performed between multi-dimensional feature vectors and similar root cause cases in the model knowledge graph to extract similar root cause analysis cases from the multi-dimensional vector matching. An expert experience decision tree is constructed based on fuzzy logic. The negative feedback adjustment strategy corresponding to similar root cause analysis cases is handled by the fuzzy membership function through the expert experience decision tree.

[0012] Preferably, the above-mentioned monitoring model customization method, based on the attention mechanism, trains corresponding monitoring models for different monitoring locations, including: The surveillance video is input into the surveillance model for learning; In the monitoring model, two subnets, left and right sequence contexts, of a bidirectional LSTM network are used respectively. Forward and backward propagation methods are used to capture the forward and backward temporal features in the monitoring video, respectively. By fusing forward and reverse temporal features according to temporal relationships, the fused dependency temporal information is obtained. By using attention mechanisms and combining them with time-series information, we can predict the classification labels and behavioral trends of monitored objects. By matching classification labels and behavioral trends with corresponding features in surveillance videos, monitoring effectiveness data can be obtained.

[0013] Preferably, the above-mentioned monitoring model customization method adjusts the monitoring parameters of the monitoring model in real time based on the monitoring effect data and the security requirements and data characteristics of the monitoring site, including: Based on the real-time monitoring footage from the monitoring equipment, determine whether the corresponding monitoring scene in the monitored area has changed; When a change in the monitoring scenario is determined, the model parameters corresponding to the monitoring scenario are matched according to the pre-defined model parameter library to adapt to the change in the monitoring scenario.

[0014] According to a second aspect of this application, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the scenario-based monitoring model customization method provided by any of the above technical solutions.

[0015] According to a third aspect of this application, this application also provides a computer storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the scenario-based monitoring model customization method provided by any of the above technical solutions.

[0016] The technical solution of this application has at least the following technical effects: The scenario-based monitoring model customization solution provided in this application acquires the security requirements and data characteristics of different monitoring locations. Based on these requirements and characteristics, a customized monitoring model can be designed for each location. Furthermore, an attention mechanism is used to train the monitoring model suitable for that location. After the monitoring model analyzes and processes the monitoring video, its monitoring parameters are adjusted using negative feedback based on the model's performance data and the location's security requirements and data characteristics. This further matches the monitoring model with the location, enabling targeted event monitoring in different scenarios and addressing the problem that existing generalized models cannot improve the accuracy of monitoring in specific scenarios. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a scenario-based monitoring model customization method provided in this application embodiment; Figure 2 for Figure 1 A flowchart illustrating the design method of the first monitoring model provided in the illustrated embodiment; Figure 3 for Figure 1 A flowchart illustrating the design method of the second monitoring model provided in the illustrated embodiment; Figure 4 for Figure 1 A flowchart illustrating the training method of the first monitoring model provided in the illustrated embodiment; Figure 5 for Figure 1 A flowchart illustrating the training method for the second monitoring model provided in the illustrated embodiment; Figure 6 for Figure 1 The illustrated embodiment provides a flowchart of a method for real-time adjustment of monitoring parameters; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0019] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. For those skilled in the art, various modifications and variations can be made to this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0020] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0021] In this application, unless otherwise expressly specified and limited, the terms "above" and "below" the second feature can refer to direct contact between the first and second features, or indirect contact between the first and second features through an intermediate medium. In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0022] The existing technology has the following drawbacks: Current security monitoring systems using a unified model cannot perform targeted event monitoring based on different monitoring scenarios, and generalized models cannot improve the monitoring accuracy of specific scenarios.

[0023] To address the aforementioned issues, the following embodiments of this application provide a scenario-based monitoring model customization solution. This solution employs customized strategies to design monitoring models tailored to the specific security needs and data characteristics of different locations. Specifically, 1) Based on location input information and historical monitoring information, scene characteristics are determined, and statistical analysis of different monitoring information is performed. Customized strategies are adopted to design models according to the specific security needs and data characteristics of different locations (such as shopping malls, banks, schools, etc.). Classification is then performed. 2) Train corresponding models for different monitoring scenarios, with each model focusing on different aspects (in shopping malls, focus on people flow and behavior analysis; in banks, focus on ensuring the security of funds and information; in schools, focus on student activities and campus order, etc.). 3) Assign corresponding monitoring scenarios to the monitoring equipment and determine whether the monitoring scenarios match, and make timely adjustments.

[0024] To achieve the above objectives, see [link to relevant documentation]. Figure 1 , Figure 1 A flowchart illustrating a scenario-based monitoring model customization method provided in this application embodiment is shown below. Figure 1 As shown, this scenario-based monitoring model customization method includes: S110: Obtain the security requirements and data characteristics of different monitoring locations, and design monitoring models for different monitoring locations according to customized strategies based on the security requirements and data characteristics of the monitoring locations.

[0025] The monitoring locations in this application include banks, shopping malls, schools, and traffic intersections. Corresponding security requirements include intrusion detection, traffic flow analysis, customer behavior analysis, and crowd density; data characteristics include resolution, frame rate, and target detection and tracking. By customizing strategies based on security requirements and data characteristics, monitoring models for different locations can be designed to adapt to the needs of various monitoring sites and improve the monitoring accuracy of specific scenarios.

[0026] Specifically, as a preferred embodiment, such as Figure 2 As shown, step S110 involves: obtaining the security requirements and data characteristics of different monitoring locations; and designing monitoring models for different monitoring locations according to a customized strategy based on these requirements and data characteristics, including: S111: Using scene understanding algorithms, identify the object types and object behaviors in the monitored locations from the surveillance videos.

[0027] For surveillance videos, this application embodiment needs to detect the meaning of each element in the monitored area to identify object types and object behaviors. Specifically, this includes: object type identification and detection, including identifying people, vehicles, objects, etc. The type of object is the primary clue for inferring the nature of the scene. In addition, semantic segmentation includes classifying each pixel in the surveillance image to divide it into objects such as roads, sidewalks, sky, buildings, and vegetation, thereby constructing a "map" of the monitored scene. In addition, for object behaviors, it is necessary to specifically analyze the behavior of people or objects, such as walking, running, loitering, gathering, falling down, and vehicles driving illegally.

[0028] S112: Determine the location type of the monitored location based on the object type and object behavior, through the spatiotemporal context.

[0029] Based on semantic understanding, the scene understanding algorithm provided in this application can categorize the entire monitoring scene into a high-level category and establish a context model. For example, it can determine whether the monitored location is an "intersection," "bank lobby," "airport security checkpoint," "retail store," "park," "school playground," or "construction site." The spatiotemporal context relationships established by the context model include temporal context, spatial context, and social context. Spatial context is used to understand the positional relationships between objects. For example, in a bank scene, a person lingering at the vault entrance for an extended period is more noteworthy than lingering in the rest area. Temporal context always understands the temporal rationality of behavior. For example, movement of people in an office area outside of working hours is abnormal. Social context: understands social norms. For example, suddenly running and screaming in a crowd is abnormal, but normal on a playground.

[0030] S113: Based on the security characteristics of the location type and combined with the spatiotemporal context, the security requirements of the monitored location are mapped, and the data characteristics are summarized.

[0031] Security requirements essentially necessitate defining anomalous situations in the monitored area. This involves identifying "abnormal situations" through the aforementioned security characteristics and mapping them to security requirements based on the spatiotemporal context. Specific methods include intrusion detection, abnormal behavior detection, object abandonment / loss detection, and traffic statistics and congestion analysis. Intrusion detection involves triggering an alarm when an object's behavior occurs within a designated restricted area of ​​the monitored site; in this case, the security requirement corresponds to area control. Abnormal behavior detection includes monitoring abnormal behavior and responses within the monitored site; in this case, the security requirement corresponds to public safety and timely response. For the detection of lost or unattended items, the security requirement corresponds to counter-terrorism and explosive ordnance prevention. For traffic statistics and congestion analysis, the security requirement corresponds to crowd management and crowd control.

[0032] S114: Real-time acquisition of surveillance video from monitored locations, learning of contextual relationships, and updating of security requirements and data characteristics based on contextual relationships and the security characteristics of the location type.

[0033] Different monitoring locations have different security requirements and data characteristics, as shown in Table 1: Table 1 – Correspondence between Monitoring Locations, Security Needs, and Data Characteristics

[0034] Finally, the pre-trained model provides an initial understanding and classification of the monitored locations, and through online learning and feedback optimization, updates to security requirements and data characteristics are achieved. Specifically, online learning includes continuously learning new "normal" patterns in the scenario during operation, and self-updating the baseline model (e.g., learning new patterns of commuter traffic). Feedback optimization includes optimizing the anomaly detection threshold and model parameters for the monitored scenario through feedback signals.

[0035] In addition, as a preferred embodiment, such as Figure 3 As shown, step S110 above involves: obtaining the security requirements and data characteristics of different monitoring locations; and designing monitoring models for different monitoring locations according to a customized strategy based on the security requirements and data characteristics of the monitoring locations, including: S115: Select a matching monitoring model from the model library based on the security requirements of the monitored location.

[0036] Different security requirements necessitate different monitoring models. This application pre-defines a model library, which includes, but is not limited to: moving target detection models, temporal behavior recognition models, high-precision recognition models, density regression networks, crowd counting models, and scene segmentation models. Different monitoring models are required for different security needs, such as intrusion detection, abnormal behavior recognition, face / license plate recognition, crowd density estimation, and item abandonment / loss. Details are shown in Table 2. Table 2 – Correspondence between security requirements and monitoring models

[0037] S116: Set the model parameters and loss function of the monitoring model according to the data characteristics of the monitored location.

[0038] S117: Preprocess the monitoring video acquired by the monitoring equipment according to the monitoring standards of the monitoring model to obtain image frame data suitable for the monitoring model.

[0039] Based on the data characteristics shown in Table 1, the data characteristics first need to be preprocessed. This preprocessing includes setting the model parameters and loss function for the monitoring model. For example: For insufficient / variable lighting: apply image enhancement algorithms (such as CLAHE), low-light image restoration models, or train on infrared / thermal imaging data; for targets with large scale variations: use multi-scale feature pyramid networks or detection models with adaptive RoI pooling; for limited computing resources: adopt lightweight models after model pruning, quantization, and knowledge distillation (such as YOLO-Lite, MobileNet backbone); for the need to process video streams: design keyframe extraction strategies instead of processing every frame to save computing power.

[0040] S118: Design a feedback closed-loop model for the monitoring model. The feedback closed-loop model includes alarm threshold settings, human-computer interaction design, and online learning mechanism. Use the feedback closed-loop model to perform closed-loop evaluation and feedback on the monitoring effect data of the monitoring model in order to adjust the model parameters and loss function of the monitoring model.

[0041] The feedback closed-loop model provided in this application includes alarm threshold settings, human-computer interaction design, and an online learning mechanism. Specifically, the alarm threshold customization is crucial: the criteria for "abnormality" differ for each scenario. Alarm thresholds need to be calibrated in a real-world environment to avoid false alarms (e.g., misinterpreting a passing plastic bag as an intrusion) and missed alarms. The human-computer interaction design ensures the system provides security personnel with actionable alarm information, such as "Intrusion detected in area A, screenshot below," rather than complex probability values. The online learning mechanism involves a feedback loop where the system records each alarm confirmation or rejection by security personnel to fine-tune the model, making it increasingly adaptable to the specific environment.

[0042] Specifically, taking security monitoring of bank vaults / perimeter surveillance as an example: its security requirements include: extremely high security, zero-tolerance intrusion, real-time alarms, and accurate post-incident evidence collection. Data characteristics include: controllable environment, potentially insufficient lighting, simple background, and sparse but critical targets. The required customized model design includes: the core model uses a high-precision target detection model (such as YOLOv8) for initial screening of "people / vehicles." It combines Siamese networks or Re-ID technology for continuous tracking of intrusion targets. Preprocessing of video data and monitoring models includes: integrating infrared video streams as the primary data source at night. Focusing ROI on key areas such as vault doors and safes to improve analysis efficiency. Subsequent closed-loop feedback and integration includes: setting extremely low alarm thresholds, with any unauthorized entry immediately triggering the highest level alarm. Upon alarm, the PTZ camera automatically zooms in on the incident area and records high-definition video for evidence collection.

[0043] Figure 1 The technical solution provided in the illustrated embodiment, after step S110: obtaining the security requirements and data characteristics of different monitoring locations, and designing monitoring models for different monitoring locations according to a customized strategy based on the security requirements and data characteristics of the monitoring locations, further includes: S120: Based on the attention mechanism, train corresponding monitoring models for different monitoring locations.

[0044] The technical solution provided in this application essentially uses an attention mechanism, which is a type of attention level or attention weight. For different monitoring environments, this application, by introducing an attention mechanism and combining it with a multilayer perceptron (MLP), can enhance the monitoring model's ability to capture and process key information.

[0045] Specifically, as a preferred embodiment, such as Figure 4 As shown, in the above-mentioned monitoring model customization method, step S120: based on the attention mechanism, train corresponding monitoring models for different monitoring locations, including: S121: Determine the corresponding monitoring area and monitoring object based on the security requirements and data characteristics of different monitoring locations.

[0046] S122: Match corresponding attention weights to the monitored area and monitored object according to security requirements and data characteristics.

[0047] S123: Using attention weights, calculate the attention values ​​that the monitoring model matches for the monitoring area and the monitoring object, respectively.

[0048] S124: Based on the attention value, train the monitoring model using the collected monitoring video and obtain monitoring effect data.

[0049] The technical solution provided in this application, based on the aforementioned security requirements and data characteristics, determines the monitoring area and monitoring object. This enables precise monitoring of target objects in the corresponding monitoring locations. For example, each monitoring model focuses on different aspects: in a shopping mall, the focus is on pedestrian flow and behavior analysis; in a bank, the focus is on ensuring the security of funds and information; and in a school, the focus is on student activities and campus order. Furthermore, according to the aforementioned security requirements and data characteristics, corresponding attention weights, or attention levels, can be matched to specific monitoring objects and monitoring areas, thus accurately identifying and determining the objects requiring attention. Then, the attention weights are used to calculate the attention values ​​matched by the monitoring model for the monitoring area and monitoring object respectively. Using these attention values, the monitoring model is trained with the monitoring video, thereby accurately obtaining monitoring effect data.

[0050] In addition, as a preferred embodiment, such as Figure 5 As shown, the above-mentioned monitoring model customization method, S120: Based on the attention mechanism, train corresponding monitoring models for different monitoring locations, including: S125: Input the surveillance video into the surveillance model for learning.

[0051] S126: In the monitoring model, two subnets, left and right sequence contexts, of a bidirectional LSTM network are used respectively. Forward and backward propagation methods are used to capture the forward and backward temporal features in the monitoring video respectively.

[0052] S127: Fuse forward and reverse time series features according to the time series relationship to obtain the fused dependency time series information.

[0053] S128: Using an attention mechanism and combining it with time-series information, predict the classification labels and behavioral trends of the monitored objects.

[0054] S129: Match the classification labels and behavioral trends with the corresponding features in the surveillance video to obtain the monitoring effect data.

[0055] The technical solution provided in this application integrates a bidirectional LSTM network with an attention mechanism to enable causal relationship queries between monitored objects and monitored areas in surveillance videos. Specifically, two subnets of the bidirectional LSTM network are used to propagate forward and backward, respectively, to obtain forward and backward temporal features. This allows the determination of dependent temporal information in the surveillance video, i.e., the temporal dependencies between monitored objects. Combining the aforementioned attention mechanism and dependent temporal information, the type and behavioral trends of monitored objects can be predicted. Finally, matching the aforementioned classification labels and behavioral trends with the corresponding features in the surveillance video determines the specific monitoring effect of the video.

[0056] in addition, Figure 1 The technical solution provided in the illustrated embodiment, after step S120: training corresponding monitoring models for different monitoring locations according to the attention mechanism, further includes: S130: Real-time acquisition of monitoring effect data from the monitoring model, and real-time adjustment of the monitoring parameters of the monitoring model based on the monitoring effect data, the security requirements of the monitoring site, and the data characteristics.

[0057] This application's embodiments enable model customization to meet specific security monitoring needs: Based on the specific security requirements and data characteristics of different locations (such as shopping malls, banks, and schools), we employ customized strategies to design models. For example, in shopping malls, the focus is on pedestrian flow and behavior analysis; in banks, the emphasis is on ensuring the security of funds and information; and in schools, the focus is on student activities and campus order. Addressing these diverse needs, this application's embodiments achieve precise customized design by adjusting model hyperparameters, selecting appropriate loss functions, and optimizing the network structure.

[0058] Specifically, as a preferred embodiment, such as Figure 6 As shown, step S130 above involves: real-time acquisition of monitoring effect data from the monitoring model; and real-time adjustment of the monitoring parameters of the monitoring model based on the monitoring effect data, the security requirements of the monitoring location, and the data characteristics, including: S131: Compare the monitoring effect data of the monitoring model with the security requirements of the monitoring site to obtain the monitoring effect error.

[0059] Comparing monitoring performance data with security requirements first requires identifying key performance indicators and feedback channels. These performance indicators include the following: False alarm rate: The frequency of incorrect system alarms. This is the most common tuning driver.

[0060] False negative rate: The frequency at which the system fails to identify real events.

[0061] Detection rate / recall rate: The proportion of successfully identified events out of all real events.

[0062] Accuracy / Precision: The proportion of actual events among all alarms.

[0063] Average response time: The delay from the occurrence of an event to the system alarm.

[0064] The feedback channel includes the following: Alarm confirmed: Security personnel confirm that the alarm is genuine.

[0065] Mark a false alarm: Security personnel reject or mark an alarm as a false alarm.

[0066] Missed report: Security personnel later discovered an incident in the video footage that the system did not trigger an alarm.

[0067] Annotated data: Re-annotate specific video segments to provide richer supervision signals.

[0068] By combining the above performance indicators and feedback channels with the performance requirements, the root cause of the problem can be analyzed.

[0069] S132: Combine data characteristics and monitoring parameters to perform root cause analysis on monitoring effect errors, obtain root cause analysis results, and match corresponding negative feedback adjustment strategies for the root cause analysis results.

[0070] Monitoring performance errors include excessively high false positive rates and excessively high false negative rates. Excessively high false positive rates are mainly caused by factors such as excessively low detection thresholds, weak model generalization ability, or learning incorrect features. Specifically, Examples of scenarios where the detection threshold is too low include: swaying leaves, changes in light and shadow, and small animals running by, all of which are identified as "intrusions." This indicates that the sensitivity parameter for motion detection or the confidence threshold for target detection is set too low.

[0071] The model's generalization ability is too strong, or it has learned incorrect features. An example scenario is that it misreports any red object as a "fire." This indicates that the model's learning of the features of "flame" is not accurate enough.

[0072] In addition, a high false negative rate may be due to factors such as an excessively high detection threshold, the model's inability to handle new scenarios, and scenario "drift." Specific scenario examples are as follows: Examples of scenarios where the detection threshold is too high include: only people very close to the wall trigger the alarm, while people at a distance are ignored. This indicates that the confidence threshold or target size filtering parameter is set too high. Monitoring models may fail to handle new scenarios, as seen in examples such as: pedestrian detection performance drops sharply at night or in rainy or snowy weather. This suggests that the model's training data lacks diversity for these scenarios. Furthermore, regarding scene "drift," examples include: the model is effective initially at construction sites, but as new buildings are constructed, the background permanently changes, causing the background subtraction algorithm to fail.

[0073] In one preferred embodiment, in the above-mentioned monitoring model customization method, step S132: combining data features and monitoring parameters to perform root cause analysis on the monitoring effect error, obtaining the root cause analysis results, and matching the root cause analysis results with corresponding negative feedback adjustment strategies, including: By calling the model knowledge graph and combining it with data features, a multi-dimensional vector corresponding to the monitoring effect error is constructed. Multi-dimensional vector matching is performed between multi-dimensional feature vectors and similar root cause cases in the model knowledge graph to extract similar root cause analysis cases from the multi-dimensional vector matching. An expert experience decision tree is constructed based on fuzzy logic. The negative feedback adjustment strategy corresponding to similar root cause analysis cases is handled by the fuzzy membership function through the expert experience decision tree.

[0074] S133: Adjust the monitoring parameters of the monitoring model through a negative feedback adjustment strategy.

[0075] Adjustments to monitoring parameters include confidence thresholds, sensitivity parameters, ROI, and target size range. Specifically, the confidence threshold controls the monitoring model's "confidence" in predicting results. A high false positive rate increases the threshold (e.g., from 0.5 to 0.7); a high false negative rate decreases the threshold (e.g., from 0.7 to 0.4). Regarding sensitivity parameters, for traditional motion detection algorithms, the sensitivity is directly adjusted; a high false positive rate decreases the sensitivity. For ROI, the monitoring area is dynamically adjusted. For example, in a road construction area, the construction zone is temporarily excluded from intrusion detection ROI. Additionally, regarding target size range, excessively large or small detection boxes are filtered out to adapt to specific scenarios.

[0076] In addition, as a preferred embodiment, in the above-mentioned monitoring model customization method, step S130: adjusting the monitoring parameters of the monitoring model in real time according to the monitoring effect data and the security requirements and data characteristics of the monitoring site, including: Based on the real-time monitoring footage from the monitoring equipment, determine whether the corresponding monitoring scene in the monitored area has changed; When a change in the monitoring scenario is determined, the model parameters corresponding to the monitoring scenario are matched according to the pre-defined model parameter library to adapt to the change in the monitoring scenario.

[0077] The technical solution provided in this application adapts to the monitoring needs of mobile scenarios by determining whether the monitoring scene in the monitoring location has changed. Specifically, it ensures that the model can be easily extended to new modal data, such as fusing infrared images and thermal imaging data, or flexibly adapt to different scenarios, such as community security monitoring or campus security monitoring. For example, when expanding from relatively simple community security monitoring to campus security monitoring with more complex personnel structures and more diverse environments, the model can quickly adjust its parameters to adapt to the new scenario requirements and play an effective monitoring role.

[0078] The scenario-based monitoring model customization method provided in this application obtains the security requirements and data characteristics of different monitoring locations. Based on these requirements and characteristics, a customized monitoring model can be designed for each location. Furthermore, an attention mechanism is used to train the monitoring model suitable for that location. After the monitoring model parses and processes the monitoring video, its monitoring parameters are adjusted using negative feedback based on the model's monitoring performance data and the location's security requirements and data characteristics. This further matches the monitoring model with the location, enabling targeted event monitoring in different scenarios using a location-specific monitoring model. This addresses the problem that existing generalized models cannot improve the accuracy of monitoring in specific scenarios.

[0079] In addition, the following embodiments of this application provide product embodiments, the beneficial effects of which are the same as those of the scenario-based monitoring model customization method provided in the above embodiments, and other technical features in the product embodiments are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0080] See Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device includes: The memory, the processor, and the computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the scenario-based monitoring model customization method provided by any of the above technical solutions.

[0081] like Figure 7As shown, the electronic device can include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory ROM 1002 or a program loaded from a storage device 1003 into a random access memory RAM 1004. The RAM 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. The communication device 1009 can operate the electronic device to exchange data with other devices wirelessly or via wired communication. Although the diagram shows a model building device with various systems, it should be understood that it is not required to implement or have all of the systems shown. It is possible to implement or have more or fewer systems alternatively.

Claims

1. A method for customizing a scenario-based monitoring model, characterized in that, include: Obtain the security requirements and data characteristics of different monitoring locations, and design monitoring models for different monitoring locations according to the security requirements and data characteristics of the monitoring locations and in accordance with customized strategies; Based on the attention mechanism, corresponding monitoring models are trained for different monitoring locations. The monitoring model's monitoring performance data is collected in real time, and the monitoring parameters of the monitoring model are adjusted in real time based on the monitoring performance data, the security requirements of the monitoring location, and the data characteristics.

2. The method as described in claim 1, characterized in that, The process of acquiring the security requirements and data characteristics of different monitoring locations, and designing monitoring models for different monitoring locations according to a customized strategy based on these requirements and data characteristics, includes: Using scene understanding algorithms, the types and behaviors of objects in the monitored locations are identified from the surveillance videos of the monitored locations. Based on the object type and object behavior, the location type of the monitored location is determined through spatiotemporal context. Based on the security characteristics corresponding to the location type and combined with the spatiotemporal context, the security requirements of the monitored location are mapped, and the data characteristics corresponding to the security requirements are summarized. The system acquires real-time surveillance videos of the monitored locations, learns the spatiotemporal context relationships, and updates the security requirements and data characteristics based on the spatiotemporal context relationships and the security features.

3. The method as described in claim 1, characterized in that, The process of acquiring the security requirements and data characteristics of different monitoring locations, and designing monitoring models for different monitoring locations according to a customized strategy based on these requirements and data characteristics, includes: Based on the security requirements of the monitored locations, a matching monitoring model is selected from the model library; Based on the data characteristics of the monitored locations, set the model parameters and loss function of the monitoring model; According to the monitoring standards of the monitoring model, the monitoring video acquired by the monitoring equipment is preprocessed to obtain image frame data suitable for the monitoring model; Design a feedback closed-loop model corresponding to the monitoring model. The feedback closed-loop model includes alarm threshold setting, human-computer interaction design, and online learning mechanism. Use the feedback closed-loop model to perform closed-loop evaluation feedback on the monitoring effect data of the monitoring model, so as to adjust the model parameters and loss function of the monitoring model.

4. The method as described in claim 1, characterized in that, The step of training corresponding monitoring models for different monitoring locations based on the attention mechanism includes: Based on the security needs and data characteristics of different monitoring locations, determine the corresponding monitoring areas and monitoring objects; Based on the security requirements and data characteristics, match corresponding attention weights for the monitoring area and the monitoring object; Using the attention weights, calculate the attention values ​​that the monitoring model matches for the monitoring area and the monitoring object, respectively; Based on the attention value, the monitoring model is trained using the collected monitoring videos, and monitoring effect data is obtained.

5. The method as described in claim 1, characterized in that, The process involves real-time acquisition of monitoring performance data from the monitoring model, and real-time adjustment of the monitoring parameters of the monitoring model based on the monitoring performance data, the security requirements of the monitored location, and data characteristics. This includes: The monitoring effect data of the monitoring model is compared with the security requirements of the monitoring site to obtain the monitoring effect error. By combining the data characteristics and the monitoring parameters, a root cause analysis is performed on the monitoring effect error to obtain the root cause analysis results, and a corresponding negative feedback adjustment strategy is matched for the root cause analysis results. The monitoring parameters of the monitoring model are adjusted using the negative feedback adjustment strategy.

6. The method as described in claim 5, characterized in that, The step of performing root cause analysis on the monitoring effect error by combining the data features and the monitoring parameters, obtaining the root cause analysis results, and matching the root cause analysis results with corresponding negative feedback adjustment strategies includes: The model knowledge graph is invoked, and a multi-dimensional vector corresponding to the monitoring effect error is constructed by combining the data features. Multi-dimensional vector matching is performed between the multi-dimensional feature vectors and similar root cause cases in the model knowledge graph to extract similar root cause analysis cases from the multi-dimensional vector matching. An expert experience decision tree is constructed based on fuzzy logic. The negative feedback adjustment strategy corresponding to the similar root cause analysis cases is then processed using the fuzzy membership function through the expert experience decision tree.

7. The method as described in claim 1, characterized in that, The step of training corresponding monitoring models for different monitoring locations based on the attention mechanism includes: The surveillance video is input into the surveillance model for learning; In the monitoring model, two subnets, left and right sequence contexts, are used in a bidirectional LSTM network. The forward and backward propagation methods are used to capture the forward and backward temporal features in the monitoring video, respectively. The forward and reverse temporal features are fused according to the temporal relationship to obtain the fused dependency temporal information; Using the attention mechanism and the dependent time-series information, the classification labels and behavioral trends of the monitored objects are predicted; The classification labels and behavioral trends are matched with corresponding features in the surveillance video to obtain the surveillance effect data.

8. The method as described in claim 1, characterized in that, The step of adjusting the monitoring parameters of the monitoring model in real time based on the monitoring effect data and the security requirements and data characteristics of the monitoring location includes: Based on the real-time monitoring footage from the monitoring equipment, determine whether the corresponding monitoring scene in the monitored location has changed; When a change in the monitoring scenario is determined, the model parameters corresponding to the monitoring scenario are matched according to a predetermined model parameter library to adapt to the change in the monitoring scenario.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the scenario-based monitoring model customization method as described in any one of claims 1 to 8.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the scenario-based monitoring model customization method as described in any one of claims 1 to 8.

Images