Intelligent supervision system based on multi-network fusion, unmanned aerial vehicle inspection and AI large model

Abstract

This invention relates to the interdisciplinary fields of intelligent monitoring, artificial intelligence recognition, drone applications, and public safety management. Specifically, it relates to an intelligent supervision system based on multi-network integration, drone inspection, and AI large-scale models. The intelligent supervision system includes a multi-network integrated monitoring module, a large-scale model intelligent recognition module, an information database module, a drone disposal module, and a warning and penalty module. This invention establishes a multi-network integrated three-dimensional monitoring system, achieving deep correlation between pet dogs and their owners. The multi-network integrated monitoring system covers fixed areas and blind spots, forming a more comprehensive collaborative supervision system. This facilitates subsequent monitoring or penalties and can also help owners find lost pet dogs, improving the practicality of the supervision system. It breaks down data barriers between monitoring systems, government databases, and law enforcement systems, achieving a closed loop in the entire process of supervision, disposal, and judgment. Simultaneously, it significantly improves supervision efficiency and coverage.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of intelligent monitoring, artificial intelligence recognition, drone applications and public safety management, specifically to an intelligent supervision system based on multi-network integration, drone inspection and AI big data models. Background Technology

[0002] Urban dog ownership regulations are dog management guidelines formulated by various localities to regulate dog ownership behavior and ensure public safety and environmental hygiene. These regulations mainly include requirements for dog registration, rabies vaccination, one dog per household, and prohibitions on keeping aggressive and large dog breeds. When taking dogs out, they must be led by an adult, using a leash no longer than 1.5 meters, wearing a dog tag and muzzle, promptly cleaning up dog feces, and giving way to vulnerable groups such as the elderly, pregnant women, and children. The regulations also clearly define prohibited dog areas and prohibit behaviors such as barking disturbances, dog attacks, and abandonment or abuse of dogs. Violations will be subject to corresponding legal liabilities.

[0003] Due to limited regulatory capacity, the problem of unregulated dog ownership is prominent, with frequent violations such as walking dogs without leashes, single individuals leading multiple dogs, and large dogs not wearing muzzles in public. These violations easily lead to personal safety risks and public order disruptions. However, existing regulatory models have significant shortcomings: First, they rely on manual inspections, which are inefficient and have limited coverage, making it difficult to achieve 24 / 7, all-area supervision. Second, existing monitoring systems (such as the "Skynet Project") lack intelligent identification capabilities for dog ownership behavior and cannot automatically identify violations. Third, there is a time lag between the discovery and handling of violations, making it difficult to stop dangerous behaviors in a timely manner. Fourth, the connection between dogs and their owners is insufficient, making it difficult to quickly identify the responsible party after a violation and lacking reliable grounds for punishment. Fifth, the data from various monitoring systems are isolated, failing to form a collaborative regulatory network. Therefore, an intelligent regulatory system based on multi-network integration, drone inspections, and AI big data models is proposed. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides an intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an intelligent supervision system based on multi-network fusion, drone inspection, and AI large-scale model. The intelligent supervision system includes a multi-network fusion monitoring module, a large-scale model intelligent recognition module, an information database module, a drone disposal module, and a warning and penalty module. The multi-network fusion monitoring module is used to construct an inspection network; the large-scale model intelligent recognition module is used for accurate classification and recognition of dog-related behaviors; the information database module stores dog registration information and owner identity information; the drone disposal module is used to report illegal behaviors in real time and issue warnings; and the warning and penalty module is used to link with the law enforcement system to generate judgment criteria.

[0006] Preferably, the inspection network in the multi-network converged monitoring module includes a static inspection network and a dynamic inspection network;

[0007] The multi-network integrated monitoring module is based on the existing fixed monitoring points of the "Skynet Project" and "Skynet Project" with no less than 5 per square kilometer, to build a ground static monitoring network, cover high-frequency dog-raising scenarios, and build a static inspection network.

[0008] By connecting to the government's "One-Stop Government Service Platform" and combining it with a static inspection network, 10-20 drones equipped with 5G communication and AI edge computing capabilities are deployed to build a dynamic inspection network and fill the gaps in fixed monitoring.

[0009] Preferably, the large-scale intelligent recognition module includes a dog-raising behavior-specific recognition model, a dog feature extraction unit, and a violation behavior judgment unit, wherein the dog feature extraction unit extracts dog biological features based on the ViT module;

[0010] The dog-raising behavior identification model is based on the YOLO object detection algorithm and the Transformer large model. It is trained with a large number of dog-raising regulations or violations to achieve accurate classification and identification of dog-raising behaviors.

[0011] The dog feature extraction unit extracts biometric features of the identified dogs using the Transformer large model, including breed, coat color, body size, unique markings, ear shape, and tail shape biometric data, and generates a unique dog feature code for each dog.

[0012] The violation determination unit automatically determines whether a violation has occurred by comparing the identification results of the dog ownership behavior identification model with the preset dog ownership regulation threshold.

[0013] Preferably, the samples of the dog-raising behavior identification model include image samples and video samples; the number of image samples is no less than 500,000, of which the proportion of samples of violations is ≥60%; the sample collection time includes different lighting conditions such as daytime, nighttime, sunny days, rainy days, and snowy days;

[0014] After the image samples were collected, OpenCV was used to perform image denoising, brightness equalization and scale normalization preprocessing in sequence; for video samples, images were extracted frame by frame, one image was extracted every 5 frames, and invalid samples with blur or occlusion area ≥50% were removed; through data desensitization processing, the privacy information of pedestrian faces and license plates in the samples was blurred.

[0015] The samples are labeled using a "target + behavior + feature" method, and LabelImg and VGG Image Annotator are used for joint annotation. The annotation supports the linkage of bounding box and attribute annotation. With the help of two-person cross-annotation and random inspection and verification, the annotation consistency ≥0.9 is considered qualified. Among them, "target" is labeled with dog type and owner; "behavior" is labeled with dog-raising behavior type; "feature" is labeled with dog biological characteristics.

[0016] Preferably, the dog-raising behavior-specific recognition model is trained using the YOLOv9 target detection algorithm as the main framework, and integrates the Vision Transformer attention mechanism to improve the recognition accuracy of small targets and complex backgrounds. At the same time, it integrates a lightweight version of ChatGLM-4 as a semantic understanding module to complete the semantic judgment of the relationship between the dog and the owner, and output the recognition result, which includes the violation type, violation time, and recognition confidence.

[0017] Preferably, in the dog-raising behavior identification model, the preprocessed images are sequentially subjected to random flipping, random cropping, color jittering, and Gaussian noise addition; CIoU loss, Focal Loss, and cross-entropy loss are weighted and optimized in a 3:2:1 ratio; INT8 quantization, channel pruning, and TensorRT inference engine are used to compress and accelerate the dog-raising behavior identification model to meet the needs of drone edge deployment; a quarterly incremental training mechanism is established, with no less than 10,000 new violation samples added each quarter for continuous iterative optimization.

[0018] Preferably, the information database module is connected to the government dog registration system, and the stored content includes: dog registration information, dog identification code, owner identity information, dog license number, dog photo, immunization certificate, dog owner's property certificate or lease contract, violation history, and reward record;

[0019] The dog registration information includes breed, age, and vaccination status; the owner's identity information includes name, ID number, contact information, and address.

[0020] The information database module supports a fast feature code comparison function. After receiving the dog feature code sent by the large model intelligent recognition module, it completes the retrieval within 1 second and matches the corresponding dog and owner information.

[0021] Preferably, the drone handling module includes a violation response unit, an on-site handling unit, and an evidence collection and archiving unit; after receiving a violation behavior determined by the large model intelligent recognition module, the violation response unit automatically sends a dispatch instruction to the drone in the nearest area, the instruction including the violation location and violation type;

[0022] After the drone arrives at the scene, the on-site handling unit plays a warning message about the violation to the dog owner through the speaker on the drone, and at the same time activates a non-contact dispersal device to stop the dog's dangerous behavior.

[0023] The evidence collection and archiving unit drives the drone to capture and store videos and pictures of the violation scene in real time through a high-definition camera, forming evidence data that serves as a visual basis for subsequent judgment. The captured content includes information such as location, dog status, and owner behavior.

[0024] Preferably, the warning and penalty module includes a data integration unit, an enforcement linkage unit, and a record archiving unit; the data integration unit automatically integrates violation identification results, dog-owner matching information, and drone evidence data to form a complete chain of evidence of violation;

[0025] The aforementioned law enforcement linkage unit synchronizes the chain of evidence of violations to the government affairs system of the law enforcement department, allowing law enforcement personnel to view and review it online, serving as the legal basis for administrative penalties;

[0026] The record archiving unit archives the entire process of violations, handling, and penalties, forming a credit record for dog owners and providing data support for subsequent tiered supervision.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention establishes a multi-network integrated three-dimensional monitoring system to achieve a deep connection between pet dogs and their owners. The multi-network integrated monitoring system covers fixed areas and blind spots, forming a more complete collaborative supervision system. This facilitates subsequent monitoring or punishment, and can also help owners find lost pet dogs. It improves the practicality of the supervision system, breaks down data barriers between the monitoring system, government database, and law enforcement system, and realizes a closed loop for the entire process of supervision, handling, and punishment. At the same time, it greatly improves the efficiency and coverage of supervision.

[0029] 2. This invention establishes an AI large-scale model intelligent recognition module, which can accurately and deeply bind the owner during the supervision process, realize the automatic identification and accurate judgment of dog ownership violations, reduce the cost of manual intervention, and form a complete and reliable electronic evidence chain with the recording and archiving unit, supporting subsequent automated law enforcement. The entire process does not require manual intervention, which greatly improves the efficiency of supervision and effectively solves the problems of insufficient manpower, difficulty in obtaining evidence, and delayed handling in urban dog management.

[0030] 3. This invention utilizes drone patrols to achieve rapid response, on-site warnings, and dispersal, effectively shortening the time for handling violations and preventing safety accidents. By matching dog characteristics and collecting evidence throughout the process, a complete chain of evidence is formed, solving the problem of difficulty in identifying the responsible party. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the intelligent monitoring system architecture for dog ownership regulations of the present invention;

[0032] Figure 2 This is a schematic diagram of the workflow of the intelligent monitoring system for dog ownership regulations of the present invention;

[0033] Figure 3 This is a schematic diagram of the training process of the AI ​​model for recognizing dog-related violations of the present invention;

[0034] Figure 4 This is a schematic diagram illustrating the matching of violation determination with dog owners according to the present invention.

[0035] Figure 5 This is a schematic diagram illustrating the automatic scheduling and precise handling of drones according to the present invention.

[0036] Figure 6 This is a schematic diagram illustrating the evidence fixation and closed-loop penalty process of the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments, which illustrate the above and other technical features and advantages of the present invention. However, the following embodiments are merely preferred embodiments of the present invention and are not exhaustive.

[0038] Example 1:

[0039] like Figure 1-3 As shown, this invention provides an intelligent supervision system based on multi-network integration, drone inspection, and AI big data model. The intelligent supervision system includes a multi-network integrated monitoring module, a big data model intelligent recognition module, an information database module, a drone disposal module, and a warning and penalty module. The multi-network integrated monitoring module is used to construct the inspection network; the big data model intelligent recognition module is used for accurate classification and recognition of dog-owning behavior; the information database module stores dog registration information and owner identity information; the drone disposal module is used to report illegal behavior in real time and issue warnings; and the warning and penalty module is used to link with the law enforcement system to generate the basis for judgment.

[0040] In this embodiment, the inspection network in the multi-network converged monitoring module includes a static inspection network and a dynamic inspection network;

[0041] Construct a static inspection network with no fewer than 5 monitoring points per square kilometer, based on the existing "Skynet" and "Skynet" projects.

[0042] Deploy 10-20 drones equipped with 5G communication, edge computing, high-definition cameras, speakers, and ultrasonic deterrent capabilities, divide them into inspection grids, and configure 1-2 drones in each grid to build a dynamic inspection network.

[0043] In this embodiment, the large model intelligent recognition module includes a special recognition model for dog-raising behavior, a dog feature extraction unit, and a violation behavior determination unit. The dog feature extraction unit extracts the biological features of dogs based on the ViT module;

[0044] The special recognition model for dog-raising behavior collects no less than 500,000 dog-raising behavior samples. Among them, the proportion of violation behavior samples is ≥60%; the sample collection time includes different lighting conditions during day and night, sunny days, rainy days, and snowy days, and the violation types include leash / witout leash, muzzle / no muzzle, large dog / small dog, and the number of dogs led by a single person;

[0045] After the picture samples are collected, image denoising, brightness equalization, and scale normalization preprocessing are performed in sequence using OpenCV; for video samples, pictures are extracted frame by frame, 1 picture is extracted every 5 frames, and invalid samples with blur or occlusion area ≥50% are removed; through data desensitization processing, the privacy information of pedestrians' faces and license plates in the samples is blurred;

[0046] Based on the YOLOv8 algorithm and the ChatGLM large model for joint training, the feature extraction accuracy and the violation determination accuracy are optimized, and the trained model is deployed to the monitoring terminal and the drone edge computing module;

[0047] The samples are labeled in the way of "target + behavior + feature", and are jointly labeled using LabelImg and VGGImageAnnotator, supporting the linkage between bounding box and attribute annotation, cooperating with double-person cross annotation and sampling review, and the annotation consistency ≥0.9 is regarded as qualified, and the unqualified labeled samples are corrected twice; among them, "target" labels the dog type and the owner; "behavior" labels the dog-raising behavior type; "feature" labels the biological features of the dog;

[0048] The dog feature extraction unit, the biological features of the dog include: breed, coat color, body type, ear type (erect ear / folded ear), tail type, eye color, nose pattern (if there is a high-precision image).

[0049] In this embodiment, when training the special recognition model for dog-raising behavior, the YOLOv9 object detection algorithm is used as the main framework, and the VisionTransformer attention mechanism is integrated to improve the recognition accuracy of small targets such as puppies and in complex backgrounds. At the same time, the lightweight version of ChatGLM-4 is connected as a semantic understanding module to complete the semantic judgment of the association relationship between the dog and the owner, and the recognition result is output. The recognition result includes the violation type, violation time, and recognition confidence;

[0050] During training, the dataset was first divided into training, validation, and test sets in a 7:2:1 stratification. Then, the pre-trained YOLOv9 weights from COCO were loaded, and the first 10 layers of the backbone network were frozen and fine-tuned for 10 epochs (learning rate 1e-4). Subsequently, all layers were unfrozen, and a cosine annealing learning rate strategy (initial 5e-5, minimum 1e-6) was adopted, with batch size set to 32. The model was then trained for 80 epochs. The output of the trained YOLOv9 object detection model was then weighted and fused with the visual features extracted by VisionTransformer and the semantic features output by ChatGLM-4 according to a weight of 0.6:0.2:0.2. The fusion parameters were then optimized and trained for 20 epochs to finally obtain the dog-raising behavior-specific recognition model.

[0051] In this embodiment, to improve model robustness and meet the requirements of drone edge deployment, data augmentation is performed on the preprocessed images before training. In the dog-raising behavior identification model, the preprocessed images are subjected to random flipping (horizontal flip probability 0.5), random cropping (cropping ratio 0.7-1.0), color jittering, and Gaussian noise addition. CIoU loss, FocalLoss, and cross-entropy loss are weighted in a 3:2:1 ratio for optimized training. INT8 quantization, channel pruning, and the TensorRT inference engine are used to compress and accelerate the dog-raising behavior identification model to meet the requirements of drone edge deployment. Finally, a quarterly incremental training mechanism is established, with no less than 10,000 new violation samples added each quarter for continuous iterative optimization.

[0052] In this embodiment, the information database module is connected to the government dog registration system, and the stored content includes: dog registration information, dog identification code, owner identity information, dog license number, dog photos (multi-angle), immunization certificate (electronic version), dog owner's property certificate or lease contract (for verifying residence), violation history, and reward record (if there is no violation record).

[0053] The dog registration information includes breed, age, and vaccination status; the owner's identity information includes name, ID number, contact information, and address.

[0054] The information database module supports a fast feature code comparison function. After receiving the dog feature code sent by the large model intelligent recognition module, it completes the retrieval within 1 second and matches the corresponding dog and owner information.

[0055] In this embodiment, the drone handling module includes a violation response unit, an on-site handling unit, and an evidence collection and archiving unit;

[0056] After receiving a violation response unit from the large model intelligent recognition module, which determines the violation, the unit automatically sends a dispatch instruction to the nearest drone. The instruction includes the location of the violation and the type of violation (such as not leashed or large dogs without muzzles).

[0057] After a violation occurs, the system immediately locks the geographical coordinates of the fixed monitoring equipment or drone at the time the violation was first detected as the initial violation location. After receiving the dispatch instruction, the system continuously uses the monitoring equipment (or the drone that has arrived) to visually track the dog and its owner. It also uses multi-camera intersection positioning or mapping of drone real-time images with geographic information to calculate the latest latitude and longitude coordinates of the target in real time. After the drone takes off from the initial location, it receives the target's dynamic coordinates continuously updated by the cloud through the 5G network. The drone flight control system dynamically adjusts the flight path accordingly to ensure accurate arrival at the target location.

[0058] After the drone arrives at the scene, the on-site response unit plays a warning message about the violation to the dog owner through the drone's onboard speaker (such as "Your dog is not on a leash, which poses a safety risk. Please rectify the situation immediately"). At the same time, it activates a non-contact deterrent device (such as an ultrasonic deterrent device) to stop the dog's dangerous behavior. While triggering the drone's on-site response, it automatically sends an alarm and real-time video feed to the nearest law enforcement officer's terminal to request support.

[0059] The evidence collection and archiving unit drives a drone to capture and store videos and images of the violation scene in real time using a high-definition camera, forming evidence data that serves as a visual basis for subsequent judgments. The captured content includes location, dog status, owner behavior, duration of violation, on-site GPS coordinates, environmental landmark information, and interactive footage that clearly proves the relationship between the dog and its owner.

[0060] In this embodiment, the warning and penalty module includes a data integration unit, an enforcement linkage unit, and a record archiving unit; the data integration unit automatically integrates violation identification results, dog-owner matching information, and evidence data to form a complete chain of evidence of violation.

[0061] The enforcement linkage unit synchronizes the chain of evidence of violations to the government affairs system of the law enforcement department, allowing law enforcement personnel to view and review it online, which serves as the legal basis for administrative penalties;

[0062] The record archiving unit archives the entire process of violations, handling, and penalties, and generates tamper-proof electronic case files to form a dog owner's credit record, providing data support for subsequent tiered supervision.

[0063] In this embodiment, the intelligent monitoring system consists of a perception layer, a transmission layer, an intelligent processing layer, a data layer, and an application layer from top to bottom. Each layer achieves data interaction and functional linkage through standardized interfaces. The output ends of the perception layer and the application layer are connected to the input ends of the transmission layer. The transmission layer and the intelligent processing layer can transmit data bidirectionally, as can the intelligent processing layer and the data layer. The output end of the data layer is connected to the input end of the application layer.

[0064] The perception layer comprises a fixed monitoring sublayer and a mobile inspection sublayer. The fixed monitoring sublayer consists of surveillance cameras from the "Skynet Project" and "Skynet Project," covering fixed areas of the city. The mobile inspection sublayer consists of drones equipped with edge computing, high-definition video recording, loudspeaker, and deterrent functions, covering blind spots in the monitoring. Each module is labeled with a device icon and connected to the transmission layer by a solid line.

[0065] The transport layer consists of a core module comprising a 5G communication gateway, an edge computing node, and an MQTT / RESTful API protocol adapter. The 5G communication gateway is responsible for real-time data transmission, while the edge computing node is responsible for AI model inference on the drone side. The MQTT / RESTful API protocol adapter enables protocol conversion between different systems and connects the perception layer and the intelligent processing layer respectively via bidirectional arrows.

[0066] The intelligent processing layer consists of core modules including an AI large model recognition engine, a behavior determination unit, a dog feature extraction unit, and a drone scheduling unit. The AI ​​large model recognition engine includes a YOLOv9 target detection module, a ViT feature extraction module, and a ChatGLM-4 semantic understanding module. Each unit is connected through an internal bus, with the input end connected to the transmission layer and the output end connected to the data layer and the application layer.

[0067] The data layer includes a dog-owner association database, a violation evidence database, and a system log database. The dog-owner association database is connected to the government dog registration system to store dog information, owner information, and feature code data. The database module is marked with a data storage icon and connected to the intelligent processing layer and the application layer by a bidirectional arrow.

[0068] The application layer comprises a disposal terminal sublayer and an enforcement management sublayer. The disposal terminal sublayer consists of a drone loudspeaker module and an ultrasonic deterrent module, which receive dispatch instructions and execute on-site disposal. The enforcement management sublayer consists of an enforcement system terminal and a credit record management terminal, which are used for violation judgment and credit archiving. Credit archiving automatically synchronizes the administrative penalty results back to the intelligent supervision system of this invention and stores them in the dog owner's dog ownership credit record, providing data support for the subsequent implementation of hierarchical supervision by the system (such as increasing the frequency of regional inspections for dishonest dog owners), and realizing closed-loop management of dog ownership supervision.

[0069] Example 2:

[0070] like Figure 4-6 As shown in the example, the application of the system is carried out in a central park in a certain city. The park is a high-frequency activity area for dogs and has deployed the intelligent supervision system of this invention. The information database module of the system has completed seamless data docking with the municipal government dog registration system, realizing real-time retrieval and matching of dog and dog owner registration information.

[0071] In this embodiment, the static inspection network utilizes the seven existing fixed monitoring points of the "Skynet Project" within the park to complete the deployment and adaptation of a special identification model for dog-raising behavior, thereby enabling real-time capture and preliminary analysis of regional video streams.

[0072] Dynamic inspection network: Equipped with one intelligent drone, which has 5G communication, AI edge computing, high-definition camera, voice broadcasting and ultrasonic deterrence functions, it is responsible for dynamic inspection and precise on-site handling of the entire park area.

[0073] In this embodiment, taking the violation of a large dog not being leashed as an example, the full-process handling logic of the intelligent monitoring system of the present invention is explained in detail. The specific scenario is: At 8:00 AM on [Date], citizen Zhang** was walking his registered large golden retriever "DouDou" in a park without using a leash. The system triggered the automated handling process, and the specific steps are as follows:

[0074] S1, Multi-network integrated monitoring and AI intelligent recognition: Fixed surveillance cameras (static inspection network) on one side of the park's main road capture real-time video footage of dog owner Zhang** and his dog "DouDou". A dog-related behavior recognition model deployed at this fixed monitoring point / nearby edge computing node analyzes the video stream in real time and performs the following recognition operations in sequence:

[0075] Using the YOLOv9 target detection algorithm, target 1 (dog - DouDou) and target 2 (dog owner - Zhang**) are accurately selected in the video frame, achieving precise target positioning;

[0076] The VisionTransformer (ViT) feature extraction module was called to perform fine biometric extraction on the image of the dog "DouDou" and identify the core features of the dog as a Golden Retriever, with golden fur and a large body size.

[0077] The ChatGLM-4 semantic judgment module determined the following: Analyzing the spatial relationship and interactive behavior between target 1 and target 2, it determined that there is a "following" relationship between the two, and no key dog ​​ownership behavior of "leashing" was detected.

[0078] S2, Violation judgment and precise matching with dog owner: The violation judgment unit compares the recognition result with the preset dog ownership standard threshold of "large dogs must be leashed when going out". Since the dog "DouDou" is a large dog and was not leashed, the system automatically judges the behavior as a violation: large dog going out without a leash.

[0079] The dog feature extraction unit extracts the dog's biological characteristics and instantly generates a unique feature code for the dog "DouDou". The system sends this unique feature code to the information database module. The database completes the matching of the feature code with the data in the municipal government dog registration system within 0.5 seconds through an index retrieval mechanism, and successfully obtains the corresponding dog and dog owner information: dog owner Zhang**, contact number 138****0000, address XX Community XX.

[0080] S3, Automatic Drone Dispatch and Precise Handling: After receiving the violation judgment and dog owner matching information, the system violation response unit automatically sends a dispatch instruction to the intelligent drone that is patrolling the park. The instruction includes the precise GPS coordinates of the violation site obtained by fixed monitoring cross-positioning and the violation type as a large dog not on a leash.

[0081] After receiving the dispatch command, the drone initiated the handling process. The system continuously used multiple fixed surveillance cameras in the park to visually track the moving dog owner Zhang and his dog "DouDou". Through a multi-camera intersection positioning algorithm, the real-time location information of the target was updated every 0.5 seconds, and the location information was synchronized to the drone through the 5G network to realize the dynamic adjustment of the drone's flight path. The drone adjusted its flight path according to the dynamically updated coordinates, and after 30 seconds, it hovered precisely at a low altitude position about 5 meters in front of the dog owner Zhang. The high-definition camera on the drone transmitted the scene back to the system backend in real time.

[0082] The system uses equipment mounted on the drone to issue voice warnings on-site. If the dog exhibits aggressive behavior or the dog owner refuses to cooperate, an alternative non-contact deterrence plan is activated (remotely activating the directional ultrasonic deterrent device mounted on the drone to safely drive the dog away, preventing it from approaching nearby pedestrians and avoiding accidents).

[0083] The voice warning means that the drone plays a customized warning message through its onboard speaker, such as "Hello Mr. / Ms. Zhang, your large dog 'DouDou' is not on a leash as required, which violates the city's dog management regulations. Please put a leash on your dog immediately to eliminate the safety hazard."

[0084] S4, Evidence Collection and Closed-Loop Penalty: From the moment the system determines the violation, drones and fixed surveillance cameras around the park record the entire process. During the verbal intervention, the drones capture high-definition footage from multiple angles, collecting the following key evidence of the violation, and automatically anonymizing the dog owner's facial information:

[0085] The system data integration unit automatically integrates clear images of dog owner Zhang** and his dog "DouDou", close-up shots of "DouDou" not on a leash, on-site footage including park landmarks, recognition results, and dog-owner matching information to generate an unalterable electronic case file named "Violation Evidence Chain_202X0820_001".

[0086] The system's law enforcement linkage unit automatically synchronizes the aforementioned electronic case files to the municipal comprehensive administrative law enforcement bureau's government affairs system via a secure encrypted link. Law enforcement officers review the evidence chain of violations online, and after confirming that it is correct, issue an administrative penalty of "warning and imposing a fine of 200 yuan" on dog owner Zhang** in accordance with the relevant regulations on dog management in the city, and archive the credit record.

[0087] The above description is merely a preferred embodiment of the present invention and is illustrative rather than restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. An intelligent monitoring system based on multi-network integration, drone inspection, and AI big data models, characterized in that: The intelligent monitoring system includes a multi-network integrated monitoring module, a large-scale model intelligent recognition module, an information database module, a drone disposal module, and a warning and penalty module; the multi-network integrated monitoring module is used to construct an inspection network. The large-scale intelligent recognition module is used for accurate classification and recognition of dog-raising behaviors; the information database module stores dog registration information and owner identity information; the drone handling module is used to report illegal behaviors and issue warnings in real time. The warning and penalty module is used to link with the law enforcement system to generate the basis for judgment.

2. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model as described in claim 1, characterized in that, The inspection network in the multi-network integrated monitoring module includes a static inspection network and a dynamic inspection network; The multi-network integrated monitoring module is based on the existing fixed monitoring points of the "Skynet Project" and "Skynet Project" with no less than 5 per square kilometer, to build a ground static monitoring network, cover high-frequency dog-raising scenarios, and build a static inspection network. By connecting to the government's "One-Stop Government Service Platform" and combining it with a static inspection network, 10-20 drones equipped with 5G communication and AI edge computing capabilities are deployed to build a dynamic inspection network and fill the gaps in fixed monitoring.

3. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI large-scale model as described in claim 1, characterized in that, The large-scale intelligent recognition module includes a dog-raising behavior-specific recognition model, a dog feature extraction unit, and a violation behavior judgment unit. The dog feature extraction unit extracts the dog's biological features based on the ViT module. The dog-raising behavior identification model is based on the YOLO object detection algorithm and the Transformer large model. It is trained with a large number of dog-raising regulations or violations to achieve accurate classification and identification of dog-raising behaviors. The dog feature extraction unit extracts biometric features of the identified dogs using the Transformer large model, including breed, coat color, body size, unique markings, ear shape, and tail shape biometric data, and generates a unique dog feature code for each dog. The violation determination unit automatically determines whether a violation has occurred by comparing the identification results of the dog ownership behavior identification model with the preset dog ownership regulation threshold.

4. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model as described in claim 3, characterized in that, The samples for the dog-raising behavior identification model include image samples and video samples; the number of image samples is no less than 500,000, of which the proportion of samples of violations is ≥60%; the sample collection time includes different lighting conditions such as daytime, nighttime, sunny days, rainy days, and snowy days; After the image samples were collected, OpenCV was used to perform image denoising, brightness equalization and scale normalization preprocessing in sequence; for video samples, images were extracted frame by frame, one image was extracted every 5 frames, and invalid samples with blur or occlusion area ≥50% were removed; through data desensitization processing, the privacy information of pedestrian faces and license plates in the samples was blurred. The samples are labeled using a "target + behavior + feature" method, and LabelImg and VGG Image Annotator are used for joint annotation. The annotation supports the linkage of bounding box and attribute annotation. With the help of two-person cross-annotation and random inspection and verification, the annotation consistency ≥0.9 is considered qualified. Among them, "target" is labeled with dog type and owner; "behavior" is labeled with dog-raising behavior type; "feature" is labeled with dog biological characteristics.

5. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model as described in claim 1, characterized in that, The dog-raising behavior identification model is trained using the YOLOv9 target detection algorithm as the main framework, and integrates the Vision Transformer attention mechanism to improve the recognition accuracy of small targets and complex backgrounds. At the same time, it integrates a lightweight version of ChatGLM-4 as a semantic understanding module to complete the semantic judgment of the relationship between the dog and the owner, and output the recognition results, which include the violation type, violation time, and recognition confidence.

6. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model as described in claim 1, characterized in that, The preprocessed images in the dog-raising behavior identification model are sequentially subjected to random flipping, random cropping, color jittering, and Gaussian noise addition. Training is optimized using a weighted combination of CIoU loss, Focal Loss, and cross-entropy loss in a 3:2:1 ratio. The model is compressed and accelerated using INT8 quantization, channel pruning, and the TensorRT inference engine to meet the needs of edge deployment of drones. A quarterly incremental training mechanism is established, with at least 10,000 new violation samples added each quarter for continuous iterative optimization.

7. The intelligent supervision system based on multi-network fusion, drone inspection, and AI big data model as described in claim 1, characterized in that, The information database module is connected to the government dog registration system and stores the following information: dog registration information, dog identification code, owner identity information, dog license number, dog photo, immunization certificate, dog owner's property certificate or lease contract, violation history, and reward record. The dog registration information includes breed, age, and vaccination status; the owner's identity information includes name, ID number, contact information, and address. The information database module supports a fast feature code comparison function. After receiving the dog feature code sent by the large model intelligent recognition module, it completes the retrieval within 1 second and matches the corresponding dog and owner information.

8. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model as described in claim 1, characterized in that, The drone handling module includes a violation response unit, an on-site handling unit, and an evidence collection and archiving unit. After receiving a violation detected by the large model intelligent recognition module, the violation response unit automatically sends a dispatch instruction to the drone in the nearest area. The instruction includes the location and type of violation. After the drone arrives at the scene, the on-site handling unit plays a warning message about the violation to the dog owner through the speaker on the drone, and at the same time activates a non-contact dispersal device to stop the dog's dangerous behavior. The evidence collection and archiving unit drives the drone to capture and store videos and pictures of the violation scene in real time through a high-definition camera, forming evidence data that serves as a visual basis for subsequent judgment. The captured content includes information such as location, dog status, and owner behavior.

9. The intelligent monitoring system based on multi-network fusion, drone inspection, and AI big data model as described in claim 1, characterized in that, The warning and penalty module includes a data integration unit, an enforcement linkage unit, and a record archiving unit; the data integration unit automatically integrates violation identification results, dog-owner matching information, and drone evidence data to form a complete chain of evidence of violation. The aforementioned law enforcement linkage unit synchronizes the chain of evidence of violations to the government affairs system of the law enforcement department, allowing law enforcement personnel to view and review it online, serving as the legal basis for administrative penalties; The record archiving unit archives the entire process of violations, handling, and penalties, forming a credit record for dog owners and providing data support for subsequent tiered supervision.

Images