Visual video event detection system and method

Abstract

The invention relates to a visual video event detection system and method, and belongs to the technical field of event detection. The system comprises a user interaction module, a large language model engine module, an atomic energy warehouse module, a lightweight atomic energy engine module, a visual rule editor module, a rule compiler module, a resource dynamic scheduling module, a cross-device cooperation module, a feedback closed loop module and an incremental learning module. According to the system, a closed-loop system from semantic analysis to visual debugging to distributed execution to incremental optimization is constructed through the system composed of the modules, the technical threshold is lowered, the rule generation and debugging efficiency is improved, the adaptability, robustness and resource utilization efficiency of the system are enhanced, autonomous evolution of the zero-code video event detection system is achieved, and the system is suitable for popularization and application. And non-technical users can easily define complex visual detection rules.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of event detection, and particularly relates to a visual video event detection system and method. BACKGROUND

[0002] With the wide application of video monitoring technology, the demand for accurate detection of various events in videos is increasing. Traditional video event detection systems often rely on professional technicians to manually code to achieve detection rules, which has high technical threshold, low rule generation efficiency, and non-technical personnel cannot participate.

[0003] At the same time, the existing system has deficiencies in rule debugging, cross-device deployment, real-time performance, and adaptability, such as long time-consuming for rule optimization, low accuracy for cross-device collaboration, slow adaptation to new scenarios, high false positive rate, and difficulty in root cause analysis, resulting in high system maintenance cost and difficulty in meeting diversified actual application needs.

[0004] In addition, the traditional scheme has large resource consumption, low generation efficiency, and insufficient coverage of long-tail scenarios, which seriously restricts the popularization and application of video event detection technology. Therefore, there is an urgent need for a video event detection system and method that can reduce the technical threshold, improve the rule generation and debugging efficiency, and enhance the adaptability and robustness of the system. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the application is to provide a visual video event detection system and method, which builds a closed-loop system and method from semantic analysis-visual debugging-distributed execution-incremental optimization, reduces the technical threshold, improves the rule generation and debugging efficiency, enhances the adaptability, robustness and resource utilization efficiency of the system, realizes the autonomous evolution of the zero-code video event detection system, and enables non-technical users to easily define complex visual detection rules.

[0006] In a first aspect of the application, a visual video event detection system is provided, comprising: a user interaction module, a large language model engine module, an atomic capability warehouse module, a lightweight atomic capability engine module, a visual rule editor module, a rule compiler module, a resource dynamic scheduling module, a cross-device collaboration module, a feedback loop module, and an incremental learning module;

[0007] The user interaction module is configured to receive natural language descriptions input by a user through a Web / APP interface and receive operations of the user for rule debugging and configuration through a visual interface.

[0008] The large language model engine module is configured to disassemble sentence components of the natural language descriptions input by the user, deeply analyze the user's intention to obtain a semantic analysis result, and generate a debuggable detection rule chain according to the semantic analysis result.

[0009] The atomic capability warehouse module is used for scanning and mapping when generating the debuggable detection rule chain, and converting semantic elements into executable modules;

[0010] The lightweight atomic capability engine module is used for executing the executable modules and completing video event detection in combination with visual detection technology;

[0011] The visual rule editor module is used for receiving a test video uploaded by a user or a synthesized video of an adversarial sample generated by using the large language model engine module, and updating a video event detection result in real time when receiving rule debugging and configuration operations of the user through a visual interactive interface of the user interaction module;

[0012] The rule compiler module is used for splitting the debugged rule chain into an edge executable subgraph and a cloud subgraph after the user publishes the debugged rule chain;

[0013] The resource dynamic scheduling module is used for dynamically scheduling resources when deploying the debugged rule chain;

[0014] The cross-device collaboration module is used for realizing distributed logical verification of the debugged rule chain on the edge and the cloud through cross-camera Re-ID for feature extraction and real-time matching when cross-camera event detection is involved;

[0015] The feedback closed loop module is used for inputting original alarm video segments and environmental sensor data into a data processing engine to generate an adversarial sample classification archive when a feedback closed loop triggering condition is received;

[0016] The incremental learning module is used for dynamically selecting an update path according to a sample size and a false alarm rate threshold, and performing hierarchical deployment and closed loop verification after model updating.

[0017] The second aspect of the application also provides a visual video event detection method, applied to the visual video event detection system, and comprising the following steps:

[0018] The user interaction module receives a natural language description input by the user through a Web / APP interface;

[0019] The large language model engine module disassembles sentence components of the natural language description input by the user, deeply analyzes a user intention to obtain a semantic analysis result, and generates a debuggable detection rule chain according to the semantic analysis result;

[0020] The atomic capability warehouse module is used for scanning and mapping when generating the debuggable detection rule chain, and converting semantic elements into executable modules;

[0021] The lightweight atomic capability engine module is used for executing the executable modules and completing video event detection in combination with visual detection technology;

[0022] The visual rule editor module receives a test video uploaded by a user or a synthesized video of an adversarial sample generated by using the large language model engine module, and updates the video event detection result in real time when receiving a rule debugging and configuration operation performed by a user through a visual interaction interface of the user interaction module;

[0023] The rule compiler module splits the debugged rule chain into an edge executable subgraph and a cloud subgraph after the user publishes the debugged rule chain;

[0024] The resource dynamic scheduling module dynamically schedules resources when deploying the debugged rule chain;

[0025] The execution phase establishes a bidirectional control flow and performs intelligent alarm filtering;

[0026] The cross-device collaboration module performs feature extraction and real-time matching through cross-camera Re-ID when cross-camera event detection is involved, so as to realize distributed logic verification of the debugged rule chain on the edge and the cloud;

[0027] The feedback loop module inputs the original alarm video segment and environmental sensor data into the data processing engine to generate an adversarial sample classification archive when receiving a feedback loop trigger condition;

[0028] The incremental learning module dynamically selects an update path according to a sample size and a false alarm rate threshold, and performs hierarchical deployment and closed-loop verification after model updating.

[0029] The third aspect of the present application also proposes an electronic device, comprising a processor and a memory.

[0030] The processor is configured to execute any one of the visual video event detection methods described above by invoking programs or instructions stored in the memory.

[0031] The fourth aspect of the present application also proposes a computer readable storage medium, which stores programs or instructions, and the programs or instructions enable a computer to execute any one of the visual video event detection methods described above.

[0032] The present application has the following advantages:

[0033] 1) Reduce technical threshold: through natural language understanding and visual interaction, zero coding operation of the user is realized, non-technical personnel can create professional rules in a short time, and the demand description pass rate is greatly improved, solving the problem that non-technical personnel cannot participate in the traditional scheme due to high technical threshold.

[0034] 2) Improve generation and debugging efficiency: By using the LLM engine to generate rule chains, the time spent on rule generation is significantly reduced. At the same time, the visual rule debugging function enables rule optimization to be completed in a short time, which saves a lot of time and costs compared with the traditional manual coding method.

[0035] 3) Enhance system robustness and adaptability: Through the feedback loop of adversarial testing and incremental learning, accelerate the adaptation speed to new scenarios and improve the detection accuracy and stability of the system in different scenarios.

[0036] 4) Optimize resource consumption: The lightweight atomic capability engine and dynamic resource scheduling mechanism reduce the memory usage and feature transmission volume of edge devices, enabling the system to operate in a lightweight and efficient manner.

[0037] 5) Improve cross-device collaboration accuracy: Through cross-device collaboration modules and distributed logic verification, ensure a low false matching rate in multi-camera cascading scenarios and improve the accuracy of complex event detection.

[0038] 6) Achieve accurate detection and continuous optimization: The false alarm accuracy positioning function improves the root cause analysis accuracy and reduces blind parameter tuning. At the same time, the adaptive learning mechanism makes the system false alarm rate continue to decrease and continuously optimizes the detection performance. Attached Figure Description

[0039] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. It is obvious that the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings.

[0040] Figure 1 A diagram of a visual video event detection system provided in an embodiment of the present invention;

[0041] Figure 2 A diagram illustrating a visual video event detection method provided in an embodiment of the present invention;

[0042] Figure 3 A diagram illustrating a method for real-time updating video event detection results provided in an embodiment of the present invention;

[0043] Figure 4 This is a diagram illustrating a method for establishing a bidirectional control flow and performing intelligent alarm filtering, provided by an embodiment of the present invention.

[0044] Figure 5 A diagram illustrating a distributed logic verification method provided in an embodiment of the present invention;

[0045] Figure 6 A diagram illustrating a method for generating an adversarial example classification archive provided in an embodiment of the present invention;

[0046] Figure 7 This is a schematic block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0047] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0048] Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts disclosed in this invention.

[0049] In the description of this invention, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The terms "installed," "connected," and "linked" should be interpreted broadly; for example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of methods and systems consistent with some aspects of the invention as detailed in the appended claims.

[0051] This invention proposes a visual video event detection system, method, electronic device, and storage medium. It constructs a closed-loop system from semantic parsing to visual debugging, distributed execution, and incremental optimization, which lowers the technical threshold, improves the efficiency of rule generation and debugging, enhances the system's adaptability, robustness, and resource utilization efficiency, and enables the autonomous evolution of a zero-code video event detection system, allowing non-technical users to easily define complex visual detection rules.

[0052] Before introducing the embodiments of the present invention, the technical terms involved in the present invention will be introduced first.

[0053] Atomic Capability Repository: A database that stores metadata for various basic executable functional modules, including parameter definitions and calling rules for atomic-level capabilities such as visual detection, feature extraction, and logical judgment, providing reusable basic components for rule chain generation.

[0054] Structured rule chain: A structured execution sequence formed by connecting multiple atomic capability modules in a logical relationship. Through logical associations such as condition judgment and timing control between nodes, it transforms the user's natural language intent into a specific detection process that can be executed by the system.

[0055] Visualized node diagram: A schematic diagram that graphically displays the logical structure of the rule chain. Each node corresponds to an atomic capability module, and the lines between nodes represent logical relationships, allowing users to intuitively view and adjust the rule logic.

[0056] Adversarial example synthesis videos: Test videos containing special interference scenarios or edge cases generated by large language models are used to verify the detection stability of the rule chain in complex and extreme scenarios and improve the comprehensiveness of rule debugging.

[0057] Edge-executable subgraph and cloud-based subgraph: These are two execution units formed after the rule chain is split by the compiler. The edge-executable subgraph is deployed on the front-end edge device and is responsible for local real-time data processing. The cloud-based subgraph is deployed on the cloud server and is responsible for global coordination and complex logic calculations, achieving distributed and efficient execution.

[0058] Cross-camera Re-ID: Cross-camera pedestrian re-identification technology extracts target features captured by different cameras and performs accurate matching to achieve trajectory association of the same target across multiple cameras, providing a foundation for cross-device complex event detection.

[0059] Incremental learning module: This module updates and optimizes only some parameters of the model based on newly added false positive / false negative sample data without retraining the entire model. It can efficiently adapt to new scenario data and reduce the consumption of computing resources.

[0060] Layered deployment: Based on device performance and functional requirements, the model and rule chain are deployed in layers across different hardware environments such as edge terminals, edge servers, and the cloud, achieving reasonable resource allocation and efficient execution.

[0061] System Implementation Examples

[0062] Figure 1 This is a diagram of a visual video event detection system provided in an embodiment of the present invention.

[0063] In a first aspect, the present invention proposes a visual video event detection system, combining... Figure 1It includes: user interaction module 101, large language model engine module 102, atomic capability repository module 103, lightweight atomic capability engine module 104, visual rule editor module 105, rule compiler module 106, resource dynamic scheduling module 107, cross-device collaboration module 108, feedback closed loop module 109, and incremental learning module 110.

[0064] The user interaction module 101 is used to receive natural language descriptions input by users through the Web / APP interface and to receive operations performed by users through the visual interface for rule debugging and configuration.

[0065] Specifically, in this embodiment of the invention, the user interaction module 101 receives a natural language description input by the user through the Web / APP interface, such as: "Detected goods piled up at the warehouse entrance for more than 1 hour".

[0066] The large language model engine module 102 is used to decompose the sentence components of the natural language description input by the user, deeply analyze the user's intent to obtain semantic parsing results, and generate an debuggable detection rule chain based on the semantic parsing results.

[0067] The atomic capability repository module 103 is used to scan and map semantic elements into executable modules when generating debuggable detection rule chains.

[0068] Specifically, in this embodiment of the invention, the large language model engine module 102 initiates multi-level semantic parsing: first, it decomposes the core elements through dependency parsing, such as: detection object: warehouse entrance, goods; trigger action: stacking; spatiotemporal constraint: exceeding 1 hour, automatically completing implicit conditions such as goods must be stationary to be considered stacked; the atomic capability warehouse module 103 simultaneously scans the metadata table of the atomic capability warehouse, maps semantic elements to executable modules, and dynamically generates a structured rule chain based on the mapping results, such as: [area detection (warehouse entrance)] → [target tracking (goods)] → [stationary judgment] → [timer > 3600s] → alarm, while extracting implicit logic, and the rule editor renders it in real time as a visual draggable node graph: the nodes are arranged in the execution order, such as: detection → filtering → timer → alarm).

[0069] The lightweight atomic capability engine module 104 is used to execute executable modules and combine visual detection technology to complete video event detection.

[0070] The visual rule editor module 105 is used to receive test videos uploaded by users or adversarial sample synthesis videos generated by the large language model engine module. When users perform rule debugging and configuration operations through the visual interactive interface of the user interaction module, the video event detection results are updated in real time.

[0071] Specifically, in this embodiment of the invention, the visual rule editor module 105 is used to receive test videos uploaded by users or adversarial sample synthetic videos generated by the large language model engine module. When the user performs rule debugging and configuration operations through the visual interactive interface of the user interaction module, the method for updating the video event detection results in real time is described in the method class embodiment.

[0072] The rule compiler module 106 is used to split the debugged rule chain into an edge executable subgraph and a cloud subgraph after the user publishes the debugged rule chain.

[0073] Specifically, in this embodiment of the invention, the method by which the rule compiler module 106 splits the debugged rule chain into an edge executable subgraph and a cloud subgraph after the user publishes the debugged rule chain is described in the method class embodiment.

[0074] The resource dynamic scheduling module 107 is used to perform dynamic resource scheduling when deploying and debugging the rule chain.

[0075] Specifically, in this embodiment of the invention, the method for dynamic resource scheduling by the resource dynamic scheduling module 107 when deploying and debugging a rule chain is described in the method class embodiment.

[0076] The cross-device collaboration module 108 is used to perform feature extraction and real-time matching through cross-camera Re-ID when cross-camera event detection is involved, so as to realize the distributed logical verification of the debugged rule chain at the edge and the cloud.

[0077] Specifically, in this embodiment of the invention, the cross-device collaboration module 108 is used to perform feature extraction and real-time matching through cross-camera Re-ID when cross-camera event detection is involved, and the method for realizing distributed logical verification of the debugged rule chain at the edge and in the cloud is described in the method class embodiment.

[0078] The feedback closed-loop module 109 is used to input the original alarm video clip and environmental sensor data into the data processing engine to generate an adversarial example classification archive when it receives the feedback closed-loop trigger condition.

[0079] Specifically, in this embodiment of the invention, the method for the feedback closed-loop module 109 to input the original alarm video clip and environmental sensor data into the data processing engine to generate an adversarial example classification archive when it receives the feedback closed-loop trigger condition is described in the method class embodiment.

[0080] The incremental learning module 110 is used to dynamically select the update path based on the sample size and false alarm rate threshold, and to perform hierarchical deployment and closed-loop verification after the model update is completed.

[0081] Specifically, in this embodiment of the invention, the model is an optimizable machine learning model upon which each atomic capability in the system depends, including but not limited to the following two types: visual analysis models in atomic capabilities and lightweight classification or regression models in logical judgments; the method of selecting an update path and performing hierarchical deployment and closed-loop verification after completing the model update is described in detail in the following method embodiment.

[0082] Method Implementation Examples

[0083] Figure 2 This diagram illustrates a visual video event detection method provided in an embodiment of the present invention.

[0084] In a second aspect, the present invention also proposes a visual video event detection method, applied to the aforementioned visual video event detection system, combined with... Figure 2 It includes eleven steps from S21 to S211:

[0085] S21: The user interaction module receives a natural language description input by the user through the Web / APP interface.

[0086] Specifically, in this embodiment of the invention, the user interaction module receives a natural language description input by the user through the Web / APP interface, such as: "Detected goods piled up at the warehouse entrance for more than 1 hour".

[0087] S22: The large language model engine module decomposes the sentence components of the natural language description input by the user, deeply analyzes the user's intent to obtain semantic parsing results, and generates an debuggable detection rule chain based on the semantic parsing results.

[0088] Specifically, in this embodiment of the invention, the method of the large language model engine module decomposes the sentence components of the natural language description input by the user, deeply analyzes the user's intent to obtain semantic parsing results, and generates an debuggable detection rule chain based on the semantic parsing results is described in detail below.

[0089] S23: The atomic capability repository module scans and maps semantic elements into executable modules when generating debuggable detection rule chains.

[0090] Specifically, in this embodiment of the invention, the atomic capability repository module performs scanning and mapping when generating debuggable detection rule chains, and the method for converting semantic elements into executable modules is described in detail below.

[0091] S24: The lightweight atomic capability engine module executes executable modules and combines visual detection technology to complete video event detection.

[0092] S25: The visual rule editor module receives test videos uploaded by users or adversarial sample synthesis videos generated by the large language model engine module. When users perform rule debugging and configuration operations through the visual interactive interface of the user interaction module, the module updates the video event detection results in real time.

[0093] Specifically, in this embodiment of the invention, the method for the visual rule editor module to receive test videos uploaded by users or adversarial sample synthetic videos generated by the large language model engine module, and to update the video event detection results in real time when users perform rule debugging and configuration operations through the visual interactive interface of the user interaction module, is described in detail below.

[0094] S26: After a user publishes a debugged rule chain, the rule compiler module splits the debugged rule chain into an edge executable subgraph and a cloud subgraph.

[0095] Specifically, in this embodiment of the invention, the method by which the rule compiler module splits the debugged rule chain into an edge executable subgraph and a cloud subgraph after the user publishes the debugged rule chain is described in detail below.

[0096] S27: The resource dynamic scheduling module performs dynamic resource scheduling when the rule chain is deployed and debugged.

[0097] Specifically, in this embodiment of the invention, the method for dynamic resource scheduling by the resource dynamic scheduling module when deploying and debugging the rule chain is described in detail below.

[0098] S28: During the execution phase, a two-way control flow is established and intelligent alarm filtering is performed.

[0099] Specifically, in this embodiment of the invention, the method for establishing a bidirectional control flow and performing intelligent alarm filtering during the execution phase is described in detail below.

[0100] S29: When cross-camera event detection is involved, the cross-device collaboration module performs feature extraction and real-time matching through cross-camera Re-ID, realizing distributed logical verification of the debugged rule chain at the edge and in the cloud.

[0101] Specifically, in this embodiment of the invention, when cross-camera event detection is involved, the cross-device collaboration module performs feature extraction and real-time matching through cross-camera Re-ID, and the method for distributed logical verification of the debugged rule chain at the edge and in the cloud is described in detail below.

[0102] S210: When the feedback closed-loop module receives the feedback closed-loop trigger condition, it inputs the original alarm video clip and environmental sensor data into the data processing engine to generate an adversarial example classification archive.

[0103] Specifically, in this embodiment of the invention, the method for generating adversarial example classification archives is described in detail below. The original alarm video clips here include both simulated alarm clips generated from test videos during the debugging phase and real alarm clips that are marked by the user or automatically captured by the system after the system actually triggers an alarm during the operation phase.

[0104] S211: The incremental learning module dynamically selects the update path based on the sample size and false alarm rate threshold, and performs hierarchical deployment and closed-loop verification after the model update is completed.

[0105] Specifically, in this embodiment of the invention, the layered deployment includes: edge devices adopting a differential hot update mechanism; cloud services achieving seamless switching through container rolling upgrades, based on feedback sample feature analysis, such as: 80% of false alarms originate from small-sized targets, automatically triggering association rule optimization, here using a Bayesian optimization algorithm to search for the best parameter combination, for example: adjusting the minimum size of cargo detection from [50,50] pixels to [120,120], and storing the parameter changes in a structured manner in the rule database, finally, the large language model engine module actively captures the semantics of parameter changes, such as "min_size=120x120" mapping to "exclude small objects", synchronously updating the entity constraints in the event ontology library to ensure the logical consistency between natural language parsing and rule execution, the inference results of each frame of the edge device are compared with the cloud benchmark model, and if the error continuously exceeds the threshold, it is automatically rolled back; the cloud conducts A / B testing through traffic mirroring, and if the false alarm rate of the new model increases by ≥5%, it immediately triggers version reversal, forming a closed-loop guarantee.

[0106] Here, closed-loop verification automatically initiates a triple verification pipeline after the deployment of a new model / rule: historical sample backtesting, adversarial stress testing, and shadow mode A / B testing. Verification results trigger intelligent decisions: if successful (e.g., when the false positive rate of historical backtesting decreases by >40% and the accuracy of A / B testing improves by >5%), a new version is pushed out to all users; if a warning is issued (e.g., if resource consumption increases by >10% (e.g., edge memory usage changes from 85MB to 102MB), model distillation and compression are automatically initiated); if unsuccessful (e.g., when the recall rate of adversarial testing is <70%), the version is rolled back, and the defect type is marked. Finally, a three-dimensional dashboard is generated: accuracy dimension: F1 score / recall / false positive rate trend curves; resource dimension: edge CPU / memory / bandwidth consumption heatmap; business dimension: alarm accuracy / decrease in manual review volume.

[0107] Historical sample backtesting, such as: extracting 200 sets of historical samples from the test library, including false positive / false negative scenarios marked by users, and generating a comparison report through parallel inference of the old and new models; adversarial stress testing, such as: the adversarial engine dynamically generates extreme perturbation samples, such as videos of falling in 90% dense fog, to quantify the improvement of robustness; shadow mode A / B testing, such as: switching 5% of real-time traffic in the cloud to the new rule chain, and running the old and new models simultaneously on edge devices to calculate the difference in metrics in real time.

[0108] Furthermore, in the aforementioned visual video event detection method, the large language model engine module decomposes the sentence components of the natural language description input by the user, deeply analyzes the user's intent to obtain semantic parsing results, and generates an debuggable detection rule chain based on the semantic parsing results, including:

[0109] The large language model engine module decomposes core elements through dependency parsing, automatically completes implicit conditions, queries the event ontology library to match predefined entity relationships to obtain semantic parsing results, and generates debuggable detection rule chains based on the semantic parsing results.

[0110] The atomic capability repository module scans and maps semantic elements into executable modules when generating debuggable detection rule chains, including:

[0111] Based on the semantic parsing results, the metadata table of the atomic capability repository is scanned synchronously, semantic elements are mapped to executable modules, and a structured rule chain is dynamically generated based on the mapping results. The rule editor renders the structured rule chain as a visual, draggable node graph.

[0112] Specifically, in this embodiment of the invention, the large language model engine module initiates multi-level semantic parsing:

[0113] First, the core elements are broken down through dependency parsing, such as: detection object: warehouse entrance, goods; trigger action: stacking; spatiotemporal constraint: exceeding 1 hour, automatically completing implicit conditions such as goods must be stationary to be considered stacked; the atomic capability warehouse module synchronously scans the metadata table of the atomic capability warehouse, maps semantic elements to executable modules, and dynamically generates structured rule chains based on the mapping results, such as: [area detection (warehouse entrance)] → [target tracking (goods)] → [stationary judgment] → [timer > 3600s] → alarm. At the same time, implicit logic is extracted, and the rule editor renders it in real time as a visual draggable node graph: nodes are arranged in execution order, such as: detection → filtering → timer → alarm.

[0114] Figure 3 A diagram illustrating a method for real-time updating video event detection results provided in an embodiment of the present invention.

[0115] Furthermore, in the aforementioned visual video event detection method, the visual rule editor module receives test videos uploaded by users or adversarial sample synthetic videos generated using the large language model engine module. When the user performs rule debugging and configuration operations through the visual interactive interface of the user interaction module, the module updates the video event detection results in real time, combining... Figure 3 It includes four steps, S31 to S34:

[0116] S31: The visual rule editor module receives test videos uploaded by users or adversarial sample synthesis videos generated by the large language model engine module, initiates asynchronous detection requests to the atomic capability repository, and decomposes the debuggable detection rule chain into parallel execution task flows.

[0117] Specifically, in this embodiment of the invention, for example, the parallel execution task flow can be a region detection module: outputting coordinate heatmaps frame by frame, such as: identifying doorway coordinates and binding cargo IDs for target tracking; a behavior analysis module: generating time-series animations of target trajectories, such as: determining when the displacement is less than 5 pixels and starting a timer when the target is stationary; and a logic node: calculating state transition probabilities in real time, such as the confidence curve for "falling down and not getting up".

[0118] S32: The visual interactive interface synchronously renders a 3D visual dashboard, displaying alarm frames and trigger logic branches.

[0119] Specifically, in this embodiment of the invention, for example, the left screen: the original video stream, detection box, and logic trigger hot zone are superimposed in layers, and the confidence intensity is represented by red and blue spectra; the right screen: the rule chain status is dynamically displayed, and the nodes flash red and green signals according to the real-time results.

[0120] S33: When a user selects a false alarm frame, the false alarm frame context is automatically frozen and root cause tracing is initiated.

[0121] Specifically, in this embodiment of the invention, a false alarm frame is such as: a tree branch shaking being mistakenly identified as an intrusion.

[0122] S34: When the user drags the parameter slider, it triggers local re-inference, and the false alarm box changes color.

[0123] Specifically, in this embodiment of the invention, adjusting the parameter slider, for example, by adjusting the minimum target size from 50x50 to 120x120, local re-inference only takes 300ms, and the false alarm box changes from red to semi-transparent gray.

[0124] The root cause tracing includes: extracting optical flow features and intermediate data of the rule chain within a preset time period before the false alarm frame, and locating the failure node through a decision tree.

[0125] Specifically, in this embodiment of the invention, intermediate data, such as the region detection module outputting a confidence score of 0.92 in this frame, and failure nodes, such as the "moving object size filtering" threshold being too low.

[0126] Furthermore, in the aforementioned visual video event detection method, after the user publishes the debugged rule chain, the rule compiler module splits the debugged rule chain into an edge executable subgraph and a cloud subgraph, including:

[0127] The rule compiler initiates topology decomposition, which, based on the atomic capability computing power requirement library and the real-time device status table, splits the debugged rule chain into an edge executable subgraph and a cloud subgraph.

[0128] The debugged rule chain is split into an edge executable subgraph and a cloud subgraph, including: the first task is assigned to the camera's built-in processor, and the second task is encapsulated as a containerized microservice and scheduled to a cloud GPU node.

[0129] Specifically, in this embodiment of the invention, after the user publishes the debugged rule chain, the rule compiler immediately starts topology decomposition. Based on the atomic capability computing power requirement library and the real-time device status table, the rule chain is split into an edge executable subgraph and a cloud subgraph: the first task is a low-latency basic task, such as: regional intrusion detection is assigned to the camera's built-in processor; the second task is a high-computing-power task, such as: cross-target behavior analysis is encapsulated as a containerized microservice and scheduled to the cloud GPU node. Here, monitoring points can be dynamically inserted: heartbeat data is sent to the cloud every 10 minutes.

[0130] The resource dynamic scheduling module performs dynamic resource scheduling when the rule chain has been deployed and debugged, including:

[0131] When deploying and debugging the rule chain, dynamic injection technology is used. Edge devices obtain a lightweight inference engine and encrypted weight package through a secure channel. The cloud receives the Docker image and starts the stream processing pipeline. The scheduling engine evaluates the status of the edge devices and the cloud.

[0132] Figure 4 This diagram illustrates a method for establishing a bidirectional control flow and performing intelligent alarm filtering, as provided in an embodiment of the present invention.

[0133] Furthermore, in the aforementioned visual video event detection method, the execution phase establishes a bidirectional control flow and performs intelligent alarm filtering, including four steps from S41 to S44:

[0134] S41: For each preset number of frames processed, the edge node extracts metadata, compresses it into a preset size heartbeat packet, and uploads it to the cloud.

[0135] S42: Rule weight differential updates are sent from the cloud at preset time intervals;

[0136] S43: When an alarm is triggered, the edge device first performs a local decision and simultaneously uploads an event snapshot and intermediate features to the cloud for verification.

[0137] S44: The cloud-based logic engine filters based on the environmental context. If a valid event is confirmed, a tiered alert is pushed.

[0138] Specifically, in this embodiment of the invention, for example: every 5 frames processed by the edge node, metadata is extracted, compressed into a 1.2KB heartbeat packet, and uploaded to the cloud. The metadata is the target number / confidence level. The cloud sends out rule weight differential updates every 30 seconds. The edge device first performs local decision-making, such as: real-time audible and visual alarm for regional intrusion, and simultaneously uploads the event snapshot and intermediate features to the cloud for verification. The cloud logic engine combines the environmental context, such as: verifying rainstorm weather through the meteorological API for secondary filtering. If it is confirmed as a valid event, it pushes alarms in a tiered manner, such as: APP pop-up / SMS / work order system.

[0139] In some embodiments, when the network is interrupted, the edge device automatically switches to degraded mode, executes local decisions according to the last valid rule and caches 200 event queues, and incrementally synchronizes to the cloud to complete the audit chain after the network is restored.

[0140] Here, the preset number of frames, preset size, and preset time period can be flexibly set according to the actual situation.

[0141] Figure 5 This is a diagram illustrating a distributed logic verification method provided in an embodiment of the present invention.

[0142] Furthermore, in the aforementioned visual video event detection method, the cross-device collaboration module is used for cross-camera event detection. It performs feature extraction and real-time matching via cross-camera Re-ID, enabling distributed logical verification of the debugged rule chain at the edge and in the cloud. Figure 5 It includes six steps, S51 to S56:

[0143] S51: The user interaction module receives cross-space event rules defined by the user through natural language, and the large language model engine module decomposes them into a three-stage atomic task chain;

[0144] S52: The rule compiler initiates distributed scheduling, and the first spatial edge device extracts the feature vector of the target object in real time;

[0145] S53: The Re-ID module searches for targets with a similarity greater than a preset percentage in the feature library of the second space;

[0146] S54: Based on spatiotemporal constraint parameters, it is packaged into feature capsules and pushed to the Re-ID routing center via a low-latency message queue.

[0147] Specifically, in this embodiment of the invention, when a user defines cross-spatial event rules through natural language, such as: goods transported from warehouse A to warehouse B are not registered within 30 minutes, the large language model engine module parses the spatiotemporal logical relationship and decomposes it into a three-stage atomic task chain: camera A captures the departure event, cross-device Re-ID matching, and camera B identifies the missing registration behavior; the rule compiler then starts distributed scheduling, and the edge device of warehouse A extracts the target goods feature vector in real time. Here, the target goods feature vector is fused with 256-dimensional encoding of appearance / motion / texture; the Re-ID module searches for targets with similarity >95% in the feature library of warehouse B, and packages them into feature capsules with spatiotemporal constraint parameters such as: the target appearance time must be between T1+5min and T1+30min, and pushes them to the Re-ID routing center through a low-latency message queue.

[0148] It should be understood that here, the first space corresponds to the first camera, the second space corresponds to the second camera, warehouse A corresponds to camera A, and warehouse B corresponds to camera B. The preset percentage can be flexibly set according to the actual situation.

[0149] S55: The debugged rule chain is executed at the edge and in the cloud;

[0150] S56: Cloud-based logic validator integrates results from the edge and the cloud;

[0151] The three-stage atomic task chain includes: the first camera capturing the departure event, cross-device Re-ID matching, and the second camera recognizing the missing registration behavior.

[0152] Specifically, in this embodiment of the invention, after the edge device completes the detection of goods leaving / appearing, the routing center broadcasts the feature capsule to all associated devices within a 500m radius of warehouse B based on the target location prediction. Each device performs cosine similarity matching in its local feature library, such as with a threshold > 0.95, and sends back the spatiotemporal coordinates of the candidate target. The cloud performs Re-ID matching and timeout determination, and the cloud aggregator verifies spatiotemporal continuity, such as path conflict detection: the target cannot move 5 kilometers in 3 minutes. Finally, the integrity of the behavior is verified by the logical atomic capability, and the absence of hand registration is detected by the video action recognition model. The cloud logical verifier integrates the results from both ends, and an alarm is triggered when all three stages of verification are passed.

[0153] Figure 6 This diagram illustrates a method for generating an adversarial sample classification archive, as provided in an embodiment of the present invention.

[0154] Furthermore, in the aforementioned visual video event detection method, when the feedback loop module receives the feedback loop trigger condition, it inputs the original alarm video clip and environmental sensor data into the data processing engine to generate an adversarial example classification archive, and combines it with... Figure 6It includes five steps, S61 to S65:

[0155] S61: When the feedback loop module receives a user-marked false alarm / missed alarm or a visual video event, the system triggers feedback and closes the feedback loop.

[0156] S62: The visual video event system performs spatiotemporal positioning, automatically captures key frame sequences within a preset time period before and after the alarm is triggered, and marks the main target that triggered the alarm in continuous frames through a target tracking algorithm;

[0157] S63: Combines light intensity and weather data recorded by environmental sensors to dynamically overlay simulated rain, fog, and reflections;

[0158] S64: Use adversarial generative techniques to create perturbation samples;

[0159] S65: Processed keyframes and augmented data are stored according to positive / negative samples.

[0160] Specifically, in this embodiment of the invention, for example, when a user marks a false alarm / missed alarm or the visual video event system triggers automatic feedback, the original alarm video segment here contains a 30-second frame sequence before and after the alarm. The visual video event first performs spatiotemporal positioning, automatically extracts a 5-second key frame sequence before and after the alarm trigger, and uses a target tracking algorithm to mark the main target that triggered the alarm, such as goods, vehicles, or personnel, in the continuous frames. Then, it enters the multimodal enhancement stage, where it combines the light intensity and weather data recorded by environmental sensors to dynamically superimpose simulated rain, fog, reflection, and other noise. At the same time, it uses adversarial generation technology to create perturbation samples. For example, it replaces the forklift texture in the false alarm with a cargo pattern to enhance the robustness of the model. The processed key frames and enhancement data are stored according to positive / negative sample classification. Missed alarm scenarios are classified into positive sample libraries, false alarm scenarios are stored in negative sample libraries, and adversarial samples are archived separately.

[0161] The incremental learning module dynamically selects the update path based on the sample size and false positive rate threshold, including:

[0162] If the number of samples is less than the first preset number of samples or there is a need for local parameter optimization, lightweight online learning at the edge is triggered.

[0163] If the number of samples exceeds the second preset number or the false positive rate exceeds the preset percentage, then cloud-based deep retraining will be initiated.

[0164] For example: for small-scale samples, such as less than 100 samples or local parameter optimization needs, lightweight online learning at the edge is triggered first; when the cumulative number of samples exceeds 1,000 or the false positive rate exceeds 15%, deep retraining in the cloud is started.

[0165] It should be understood that the first and second preset number of entries can be flexibly set according to the actual situation.

[0166] A third aspect of the present invention also provides an electronic device comprising: a processor and a memory;

[0167] The processor executes any of the above-mentioned visual video event detection methods by calling programs or instructions stored in memory.

[0168] In a fourth aspect, the present invention also provides a computer-readable storage medium storing a program or instructions that cause a computer to execute a visual video event detection method as described in any of the preceding claims.

[0169] Figure 7 This is a schematic block diagram of an electronic device provided in an embodiment of the present invention.

[0170] like Figure 7 As shown, the electronic device includes at least one processor 701, at least one memory 702, and at least one communication interface 703. The various components of the electronic device are coupled together via a bus system 704. The communication interface 703 is used for information transmission with external devices. It is understood that the bus system 704 is used to implement communication between these components. In addition to a data bus, the bus system 704 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 7 The general designated all buses as Bus System 704.

[0171] It is understood that the memory 702 in this embodiment can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0172] In some implementations, memory 702 stores elements such as executable units or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0173] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. The program implementing any method in the visual video event detection method provided in this embodiment of the invention can be included in the application programs.

[0174] In this embodiment of the invention, the processor 701 executes the steps of various embodiments of the visual video event detection method provided by the present invention by calling the program or instructions stored in the memory 702, specifically, the program or instructions stored in the application program.

[0175] The user interaction module receives natural language descriptions input by the user through the Web / APP interface;

[0176] The large language model engine module decomposes the sentence components of the natural language description input by the user, deeply analyzes the user's intent to obtain semantic parsing results, and generates an debuggable detection rule chain based on the semantic parsing results;

[0177] The atomic capability repository module scans and maps semantic elements into executable modules when generating debuggable detection rule chains;

[0178] The lightweight atomic capability engine module executes executable modules and combines visual detection technology to complete video event detection.

[0179] The visual rule editor module receives test videos uploaded by users or adversarial sample synthesis videos generated by the large language model engine module. When users perform rule debugging and configuration operations through the visual interactive interface of the user interaction module, the video event detection results are updated in real time.

[0180] After a user publishes a debugged rule chain, the rule compiler module splits the debugged rule chain into an edge executable subgraph and a cloud subgraph.

[0181] The resource dynamic scheduling module performs dynamic resource scheduling when the rule chain has been deployed and debugged.

[0182] During the execution phase, a two-way control flow is established and intelligent alarm filtering is implemented.

[0183] When cross-camera event detection is involved, the cross-device collaboration module performs feature extraction and real-time matching through cross-camera Re-ID, enabling distributed logical verification of the debugged rule chain at the edge and in the cloud.

[0184] When the feedback closed-loop module receives the feedback closed-loop trigger condition, it inputs the original alarm video clip and environmental sensor data into the data processing engine to generate adversarial example classification archive;

[0185] The incremental learning module dynamically selects the update path based on the sample size and false alarm rate threshold, and performs hierarchical deployment and closed-loop verification after the model update is completed.

[0186] Any method in the visual video event detection method provided in this embodiment of the invention can be applied to, or implemented by, the processor 701. The processor 701 can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by the integrated logic circuits in the hardware of the processor 701 or by instructions in software form. The processor 701 can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor.

[0187] The steps of any method in the visual video event detection method provided in this embodiment of the invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software units in the decoding processor. The software units can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory 702, and processor 701 reads the information in memory 702 and combines it with hardware to complete the steps of the method.

[0188] Those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of the invention and form different embodiments.

[0189] Those skilled in the art will understand that the descriptions of the various embodiments have different focuses, and for parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0190] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention. All such modifications and variations fall within the scope defined by the appended claims. The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0191] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A visualized video event detection system, characterized in that, Comprise: User interaction module, large language model engine module, atomic ability warehouse module, lightweight atomic ability engine module, visual rule editor module, rule compiler module, resource dynamic scheduling module, cross-device collaboration module, feedback loop module and incremental learning module; Among them, the user interaction module is used to receive the natural language description input by the user through the Web / APP interface and receive the operation of the user through the visual interface for rule debugging and configuration; The large language model engine module is used to disassemble the sentence components of the natural language description input by the user, deeply analyze the user's intention to obtain semantic analysis results, and generate a debuggable detection rule chain according to the semantic analysis results; The atomic ability warehouse module is used to scan and map when generating a debuggable detection rule chain, and convert semantic elements into executable modules; The lightweight atomic ability engine module is used to execute executable modules and complete video event detection in combination with visual detection technology; The visual rule editor module is used to receive the test video uploaded by the user or the adversarial sample synthetic video generated using the large language model engine module, and update the video event detection results in real time when receiving the rule debugging and configuration operation of the user through the visual interaction interface of the user interaction module; The rule compiler module is used to split the debugged rule chain into edge executable subgraph and cloud subgraph after the user publishes the debugged rule chain; The resource dynamic scheduling module is used for dynamic scheduling of resources when deploying the debugged rule chain; The cross-device collaboration module is used to extract and match features in real time through cross-camera Re-ID when cross-camera event detection is involved, so as to realize distributed logical verification of the debugged rule chain on the edge and the cloud; The feedback loop module is used to input the original alarm video segment and environmental sensor data into the data processing engine to generate an adversarial sample classification archive when receiving a feedback loop trigger condition; The incremental learning module is used to dynamically select an update path according to the sample size and false alarm rate threshold, and perform hierarchical deployment and closed-loop verification after model updating.

2. The method of visualizing video event detection is applied to the system of visualizing video event detection of claim 1, characterized in that, Comprise: The user interaction module receives the natural language description input by the user through the Web / APP interface; The large language model engine module disassembles the sentence components of the natural language description input by the user, deeply analyzes the user's intention to obtain semantic analysis results, and generates a debuggable detection rule chain according to the semantic analysis results; The atomic ability warehouse module scans and maps when generating a debuggable detection rule chain, and converts semantic elements into executable modules; The lightweight atomic ability engine module executes executable modules and completes video event detection in combination with visual detection technology; The visual rule editor module receives the test video uploaded by the user or the adversarial sample synthetic video generated using the large language model engine module, and updates the video event detection results in real time when receiving the rule debugging and configuration operation of the user through the visual interaction interface of the user interaction module; The rule compiler module splits the debugged rule chain into edge executable subgraph and cloud subgraph after the user publishes the debugged rule chain; The resource dynamic scheduling module performs dynamic scheduling of resources when deploying and debugging the rule chain; The execution phase establishes a bidirectional control flow and performs intelligent alarm filtering; The cross-device collaboration module extracts and matches features in real time through cross-camera Re-ID when cross-camera event detection is involved, enabling distributed logic verification of the debugged rule chain on the edge and in the cloud; The feedback loop module receives feedback loop trigger conditions, inputs the original alarm video segment and environmental sensor data into the data processing engine to generate adversarial sample classifications for archiving; The incremental learning module dynamically selects an update path based on sample size and false positive rate thresholds, performs hierarchical deployment and closed-loop verification after model updating.

3. The method of claim 2, wherein, The large language model engine module disassembles the sentence components of the user's input natural language description, deeply analyzes the user's intent to obtain semantic analysis results, and generates a debuggable detection rule chain based on the semantic analysis results, including: The large language model engine module disassembles core elements through dependency syntax analysis, automatically completes implicit conditions, queries the event ontology library to match predefined entity relationships to obtain semantic analysis results, and generates a debuggable detection rule chain based on the semantic analysis results; The atomic capability warehouse module scans and maps the debuggable detection rule chain when generating it, converting semantic elements into executable modules, including: Based on the semantic analysis results, the metadata table of the atomic capability warehouse is scanned synchronously, and the semantic elements are mapped to executable modules. Based on the mapping results, a structured rule chain is dynamically generated, and the rule editor renders the structured rule chain into a visual, draggable node graph.

4. The method of claim 2, wherein, The visual rule editor module receives user-uploaded test videos or adversarial sample synthetic videos generated using the large language model engine module. When the user performs rule debugging and configuration operations through the visual interactive interface of the user interaction module, the video event detection results are updated in real time, including: The visual rule editor module receives user-uploaded test videos or adversarial sample synthetic videos generated using the large language model engine module, initiates an asynchronous detection request to the atomic capability warehouse, and disassembles the debuggable detection rule chain into parallel execution task streams; The visual interactive interface synchronously renders a three-dimensional visual dashboard, displaying alarm frames and trigger logic branches; When receiving the user's frame selection operation, the context of the false positive frame is automatically frozen, and root cause tracing is initiated; The user adjusts the parameter slider to trigger local re-inference, and the false positive frame changes color. The root cause tracing includes extracting the optical flow features within a preset time period before the false positive frame and the intermediate data in the rule chain, and locating the failed node through a decision tree.

5. The method of claim 2, wherein, The rule compiler module disassembles the debugged rule chain into edge executable subgraphs and cloud subgraphs after the user publishes the debugged rule chain, including: The rule compiler initiates topology disassembly, disassembles the debugged rule chain into edge executable subgraphs and cloud subgraphs based on the atomic capability computing power demand library and real-time device state table; Disassembling the debugged rule chain into edge executable subgraphs and cloud subgraphs includes assigning the first task to the camera's built-in processor and packaging the second task as a containerized microservice to schedule to a cloud GPU node. The resource dynamic scheduling module performs dynamic scheduling of resources when deploying and debugging the rule chain, including: The dynamic injection technology is adopted when deploying and debugging the rule chain, the edge device acquires the lightweight inference engine and encrypted weight package through a secure channel, the cloud receives the Docker image and starts the stream processing pipeline, and the scheduling engine evaluates the edge device and cloud state.

6. The method of claim 2, wherein, The execution stage establishes a bidirectional control flow and performs intelligent alarm filtering, including: The edge node extracts metadata and compresses it into a preset size heartbeat packet every time it processes a preset number of frames and uploads it to the cloud; The cloud issues rule weight differential updates every preset time period; When an alarm is triggered, the edge device first executes local decision-making, and simultaneously uploads event snapshots and intermediate features to the cloud for review; The cloud logic engine filters in combination with environmental context, and if it confirms that it is a valid event, it will push the alarm in stages.

7. The method of claim 2, wherein, The cross-device collaboration module is used when cross-camera event detection is involved, and features are extracted and matched in real time through cross-camera Re-ID to achieve distributed logic verification of the rule chain on the edge and the cloud, including: The user interaction module receives a cross-space event rule defined by the user in natural language, and the large language model engine module is decomposed into a three-segment atomic task chain; The rule compiler starts distributed scheduling, and the first space edge device extracts the target feature vector in real time; The Re-ID module searches for a target with a similarity greater than a preset percentage in the feature library of the second space; With the spatiotemporal constraint parameters, the feature capsules are packaged and pushed to the Re-ID routing center through a low-latency message queue. The rule chain is executed on the edge and the cloud; The cloud logic verifier synthesizes the edge and cloud results. The three-segment atomic task chain includes: the first camera captures a leaving event, cross-device Re-ID matching, and the second camera recognizes a missing registration behavior.

8. The method of claim 2, wherein, The feedback loop module receives a feedback loop trigger condition and inputs the original alarm video segment and environmental sensor data into the data processing engine to generate an adversarial sample classification archive, including: The feedback loop module receives a user-labeled false positive / false negative or visual video event system-triggered feedback to trigger the feedback loop. The visual video event system performs spatiotemporal positioning, automatically intercepts key frame sequences within a preset time period before and after the alarm trigger, and labels the subject causing the alarm in consecutive frames through a target tracking algorithm. In combination with the light intensity and weather data recorded by the environmental sensor, simulated rain and fog and reflections are dynamically superimposed. Adversarial generation techniques are used to create perturbation samples. The processed key frames and enhanced data are stored as positive / negative samples. The incremental learning module dynamically selects an update path based on the sample size and false positive rate threshold, including: When the sample size is less than a first preset number or local parameter optimization is required, the edge lightweight online learning is triggered. When the sample size is greater than a second preset number or the false positive rate is greater than a preset percentage, the cloud deep retraining is started.

9. An electronic device, comprising: including: a processor and a memory; The processor calls the program or instructions stored in the memory to execute the visual video event detection method according to any one of claims 2-8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores programs or instructions, which make the computer execute the visual video event detection method according to any one of claims 2 to 8.

Images