Wild animal AI monitoring and early warning system and method thereof

Abstract

The invention relates to the technical field of artificial intelligence and wild animal protection, in particular to a wild animal AI monitoring and early warning system and method, comprising a data acquisition and edge detection module, a cloud recognition and desensitization module, a manual verification and data management module, a model training module and a model deployment and early warning module which are connected in sequence, the system preliminarily screens target data through a first AI model on an edge side, uploads the target data to a cloud end, submits low-confidence data to manual verification and incorporates the low-confidence data into a central training data set after fine-grained identification and privacy desensitization are carried out through a second AI model, continuously optimizes model performance through incremental training, and improves the accuracy of data processing. And finally, deploying the updated model to an edge side and realizing intelligent graded early warning to form a self-evolutionary monitoring closed loop.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and wildlife conservation technology, and in particular to a wildlife AI monitoring and early warning system and method. Background Technology

[0002] With increasing awareness of ecological protection, using technology to monitor and protect wildlife has become an industry trend. Traditional monitoring methods mainly rely on manual patrols and the deployment of infrared trigger cameras, which face the challenges of inefficiency and slow response in vast protected areas. In recent years, artificial intelligence technology has been introduced into this field. By training deep learning models on massive amounts of image data, the automatic identification of specific wild animals has been achieved, making automated monitoring possible.

[0003] However, existing AI-based monitoring solutions suffer from a significant technical flaw: once deployed, their recognition models tend to become rigid in performance, making them ill-suited to the complex and ever-changing real-world environment. Specifically, when encountering animal postures not fully covered in the training data, new environmental backgrounds, lighting conditions, or rare species, the model's recognition accuracy drops significantly, and the system lacks effective self-optimization capabilities to overcome these issues. This leads to high false alarm and false negative rates, severely compromising the reliability of the monitoring system and failing to meet the needs of high-precision early warning and long-term scientific research monitoring. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a wildlife AI monitoring and early warning system and method, aiming to improve the high false alarm and false negative rates in the existing technology, which greatly reduces the reliability of the monitoring system and fails to meet the needs of high-precision early warning and long-term scientific research monitoring.

[0005] In a first aspect, the present invention provides the following technical solution: a wildlife AI monitoring and early warning system, comprising the following modules:

[0006] The data acquisition and edge detection module is used to continuously collect environmental data through monitoring equipment deployed in the field, and to use the first AI model deployed on the edge to analyze the environmental data in real time to detect whether there are potential targets in the data. When a potential target is identified, the data segment containing the potential target is marked and uploaded to the cloud server.

[0007] The cloud-based identification and desensitization module is used to receive data fragments from the edge via a cloud server, and to use a second AI model deployed in the cloud to perform fine-grained identification on the data fragments. It also outputs the confidence level of the fine-grained identification results and performs automatic and irreversible blurring of the human biometric information and vehicle license plate information identified in the data fragments.

[0008] The manual verification and data management module is used to construct a central training dataset. Data segments with a confidence level lower than a preset threshold for fine-grained recognition by the second AI model or data segments that have been blurred are pushed to the manual review queue. Authorized personnel verify and correct the data segments in the manual review queue, generate standard data that has been manually verified, and add the standard data to the central training dataset.

[0009] The model training module is used to periodically perform incremental training and fine-tuning on the second AI model using the updated central training dataset to generate an updated version of the second AI model.

[0010] The model deployment and early warning module is used to perform lightweight processing on the updated version of the second AI model obtained after fine-tuning, and send it to the edge to update the first AI model. At the same time, based on the fine-grained recognition results of the second AI model or the confirmation results of the manual review queue, it generates and issues early warning information of the corresponding level.

[0011] Preferably, the potential target identification process includes:

[0012] By deploying at least one of infrared cameras, visible light cameras, and sound sensors in the field, environmental data including video and audio streams is continuously collected.

[0013] The collected environmental data is preprocessed, including at least noise reduction, format standardization, and keyframe extraction.

[0014] The preprocessed environmental data is input into the first AI model deployed at the edge, and the first AI model analyzes the data to detect whether it contains potential targets that meet the preset characteristics of animals, humans or vehicles.

[0015] When the first AI model detects a potential target, it extracts a data segment from the environmental data that contains the time period in which the potential target appeared, and then marks and compresses the data segment.

[0016] Upload the tagged and compressed data fragments to the cloud server.

[0017] Preferably, the process for fine-grained identification and the identification of the confidence level of the fine-grained identification results includes:

[0018] The cloud server receives data fragments from the edge and decompresses and parses the data fragments to obtain the data to be analyzed.

[0019] The data to be analyzed is input into the second AI model deployed in the cloud;

[0020] The second AI model extracts features from the data to be analyzed to obtain a multidimensional feature representation of the data;

[0021] Based on the multidimensional feature representation, the second AI model performs at least one fine-grained identification task among species-level classification, individual identification, or abnormal behavior analysis;

[0022] The second AI model outputs the recognition results and corresponding confidence scores for the fine-grained recognition task.

[0023] Preferably, the blurring process includes:

[0024] A preset privacy filtering model is invoked to scan the data to be analyzed or the data fragments after fine-grained identification;

[0025] The privacy filtering model locates and selects all human facial regions, human body contour regions, and vehicle license plate regions in the data to be analyzed.

[0026] For each selected area, a pixelation or mosaic algorithm is applied to blur the rendering. The rendering process permanently destroys the original image details of the area, making them irreversible.

[0027] The data segments that have completed the blurring rendering are marked as de-identified data and correlated with the original recognition results and confidence levels.

[0028] Preferably, the process of constructing the central training dataset includes:

[0029] Create an initially empty database on the cloud server, which is defined as the central training dataset;

[0030] Initialize the version number of the central training dataset and create a data table for storing data entries, the data table having predefined fields for recording metadata;

[0031] Receive the standard data, which has been verified and corrected, from the human review queue;

[0032] Generate corresponding metadata for each piece of standard data. The metadata includes at least the data source, collection time, labeled species information, labeler identification, and corresponding original AI recognition results.

[0033] The standard data and its metadata are added to the central training dataset as a complete data entry, and the dataset version number is updated.

[0034] Regularly perform deduplication, quality verification, and class balance analysis on the data in the central training dataset to maintain the quality of the dataset.

[0035] Preferably, the process for generating the manually verified standard data includes:

[0036] The cloud server filters out data segments whose confidence level is lower than the preset confidence threshold based on the fine-grained recognition of the second AI model, and stores them together with the data segments that have been blurred into a preset data pool to be reviewed.

[0037] Data fragments are extracted sequentially from the data pool to be reviewed and assigned to the review workbench of an authorized person.

[0038] The data fragments are displayed to authorized personnel on the review workbench, along with the preliminary identification results and confidence levels provided by the second AI model.

[0039] Authorized personnel shall confirm, correct or supplement the preliminary identification results based on observation, and the annotation information shall include at least the correct species category, individual identifier or behavioral description;

[0040] The system receives annotation information submitted by authorized personnel, binds the annotation information with the corresponding data segment, generates a standard data that has been manually verified, and marks the standard data as available.

[0041] Preferably, the version update process of the second AI model includes:

[0042] The version number of the central training dataset is checked at a preset period. If it is determined that the version number has been updated, all the standard data newly added since the previous training period are obtained and divided into an incremental training set and an incremental validation set according to a preset ratio.

[0043] Load the weight parameters of the currently deployed second AI model as the initial model for incremental training;

[0044] The initial model is trained using the incremental training set. During the training process, a small learning rate is used and some of the lower network layers of the model are frozen. Only the weights of some of the higher network layers of the model are updated to obtain a candidate model.

[0045] The incremental validation set is used to evaluate the performance of the candidate model. When the performance improvement reaches a preset standard, the candidate model is judged to be qualified.

[0046] The candidate models that are deemed qualified are saved as updated versions of the second AI model and bound to the new version identifier, while the old model versions are archived and stored.

[0047] Preferably, the update process of the first AI model includes:

[0048] The updated version of the second AI model that is deemed qualified is subjected to lightweight processing, which includes at least one of model pruning, quantization, or knowledge distillation, to generate a lightweight model file.

[0049] The lightweight model file and model version metadata are packaged together to generate a model update package. The model version metadata includes at least the model version number, the applicable edge device type, and performance indicators.

[0050] The model update package is downloaded from the cloud server to the target edge device;

[0051] The edge device receives and parses the model update package, and replaces the original first AI model with the lightweight model file in the package;

[0052] After the replacement is completed, the edge device performs integrity verification and performance testing on the new first AI model. Once the test is passed, it sends a confirmation of successful update to the cloud server.

[0053] Preferably, the process for generating and issuing the corresponding level of early warning information includes:

[0054] Receive fine-grained recognition results and their confidence levels from the second AI model from the cloud server, or receive the confirmed final results from the manual review queue;

[0055] Based on a pre-set early warning rule base, the received results are parsed to determine the type and level of the early warning event. The types of early warning events include at least poaching, appearance of specific protected animals, abnormal animal gatherings, or human-wildlife conflict risks. The early warning levels include at least emergency, important, and general.

[0056] Based on the determined warning event type and warning level, the corresponding warning information template is called from the preset warning template library, and the event time, location, target species, quantity and confidence information are automatically filled in to generate structured warning information;

[0057] Based on the warning level, at least one preset warning channel is selected to send the structured warning information;

[0058] After the warning information is issued, the system records the time of issuance, content, recipient and channel of the warning, and waits for feedback information from the patrol personnel.

[0059] Secondly, the present invention provides the following technical solution: a method for AI-based monitoring and early warning of wild animals, the method comprising the following steps:

[0060] Environmental data is continuously collected by monitoring equipment deployed in the field, and the environmental data is analyzed in real time using a first AI model deployed at the edge to detect whether there are potential targets in the data. When a potential target is identified, the data segment containing the potential target is marked and uploaded to the cloud server.

[0061] The system receives data segments from the edge via a cloud server and uses a second AI model deployed in the cloud to perform fine-grained identification on the data segments. It also outputs the confidence level of the fine-grained identification results and performs automatic and irreversible blurring on the human biometric information and vehicle license plate information identified in the data segments.

[0062] A central training dataset is constructed. Data segments with a confidence level lower than a preset threshold or data segments that have been blurred by the fine-grained recognition of the second AI model are pushed to a manual review queue. Authorized personnel verify and correct the data segments in the manual review queue, generate standard data that has been manually verified, and add the standard data to the central training dataset.

[0063] The second AI model is incrementally trained and fine-tuned periodically using the updated central training dataset to generate an updated version of the second AI model;

[0064] The updated version of the second AI model, after fine-tuning, is lightweighted and sent to the edge to update the first AI model. At the same time, based on the fine-grained recognition results of the second AI model or the confirmation results of the manual review queue, a warning message of the corresponding level is generated and issued.

[0065] The present invention has the following beneficial effects:

[0066] 1. In this invention, by constructing a technical closed loop of edge preliminary screening, cloud-based fine-grained identification, manual verification feedback, and incremental model training, the system can continuously optimize the core AI model using the difficult data generated in actual operation. This effectively overcomes the technical defects of existing system models being rigid and difficult to adapt to complex field environments, and significantly improves the accuracy and adaptability of long-term operation.

[0067] 2. In this invention, by dividing the work and cooperating between the edge side and the cloud, edge computing is used to realize the rapid perception and data filtering of potential targets, reducing bandwidth pressure. At the same time, the powerful computing power of the cloud is used to complete high-precision fine-grained identification and multi-source information fusion. Finally, based on the confidence level and the results of manual confirmation, a graded warning is given, thereby greatly reducing false alarms and missed alarms while ensuring low latency.

[0068] 3. In this invention, by pushing low-confidence identification results and desensitized data to a manual review queue for annotation, and then integrating the generated standard data into a central training dataset, the system automatically completes the process of high-quality data collection, cleaning, and annotation, providing a continuous source of high-quality information for model iteration, and also accumulating digital assets for wildlife research.

[0069] 4. In this invention, by integrating an automated privacy filtering module in the cloud, the face, body and license plate information in the image are irreversibly blurred. While giving full play to the value of data for model training and early warning, personal privacy is strictly protected, and the data ethics problem faced by AI technology in the application of protected areas is solved. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the architecture of a wildlife AI monitoring and early warning system proposed in this invention;

[0071] Figure 2 This is a flowchart illustrating a wildlife AI monitoring and early warning method proposed in this invention. Detailed Implementation

[0072] The technical solutions in the embodiments 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Example 1:

[0074] In a first embodiment of the present invention, the present invention provides a wildlife AI monitoring and early warning system, such as... Figure 1 As shown, it includes the following modules:

[0075] The data acquisition and edge detection module is used to continuously collect environmental data through monitoring equipment deployed in the field, and to use the first AI model deployed at the edge to perform real-time analysis of the environmental data to detect whether there are potential targets in the data. When a potential target is identified, the data segment containing the potential target is marked and uploaded to the cloud server.

[0076] Furthermore, the potential target identification process includes:

[0077] By deploying at least one of infrared cameras, visible light cameras, and sound sensors in the field, environmental data including video and audio streams is continuously collected.

[0078] The collected environmental data is preprocessed, including at least noise reduction, format standardization, and keyframe extraction.

[0079] The preprocessed environmental data is input into the first AI model deployed at the edge, which analyzes the data to detect whether it contains potential targets that match the preset characteristics of animals, humans or vehicles.

[0080] When the first AI model detects a potential target, it extracts a data segment from the environmental data that contains the time period in which the potential target appeared, and then labels and compresses the data segment.

[0081] Upload the tagged and compressed data fragments to the cloud server.

[0082] Specifically, in the field environment, various sensors are deployed at predetermined locations based on the terrain and vegetation cover of the monitoring area. These sensors include, but are not limited to, infrared thermal imaging cameras for nighttime monitoring, visible light cameras for daytime high-definition photography, and high-sensitivity microphone arrays for capturing audio signals such as animal calls and vehicle engine sounds. These devices are powered by solar panels and battery systems to ensure uninterrupted operation 24 / 7. They continuously collect raw video streams (such as H.264 / H.265 format) and audio streams (such as PCM format) to form initial environmental data. The collected raw data is first processed by Wiener filtering on the audio data. Wave or spectral subtraction is used to reduce the interference of environmental noise such as wind and rain. Gaussian filtering or median filtering is applied to video data to eliminate image sensor noise. Video streams with different resolutions and frame rates are uniformly converted to a fixed resolution (e.g., 640x480 pixels) and frame rate (e.g., 15fps). Audio streams are uniformly resampled to the same sampling rate (e.g., 16kHz) to ensure the stability of subsequent AI model input. To avoid processing a large number of redundant frames, the video stream will use inter-frame difference or motion detection algorithms to extract only key frames (I-Frames) where the image content changes significantly for subsequent analysis, which can greatly reduce the computational load on the edge side.

[0083] The preprocessed data (mainly keyframe images) is input into a first AI model deployed on an edge device. This model is a lightweight object detection neural network, such as a cropped and quantized YOLO-V5s or SSD-MobileNet model. The model analyzes the image for potential targets that meet preset features, defined as three main categories: wildlife (such as whole or salient parts of tigers, leopards, bears, deer, etc.), humans, and vehicles. The model outputs the bounding box location of the target in the image and a preliminary, comprehensive presence confidence score. The first AI model does not perform fine-grained species classification; it does not miss any possible targets to ensure valuable data is uploaded. Once the first AI model detects a potential target (i.e., its presence confidence score is determined), it performs a full search. When the confidence level exceeds a low baseline threshold (e.g., 0.3), the system triggers the following actions: it extracts a short video clip (in MP4 format) from the original video stream, from 5 seconds before to 10 seconds after the target appears, and generates a unique metadata tag for it. This tag includes at least the device ID, GPS coordinates, timestamp, and the target type and confidence level detected by the first AI model. To save valuable wireless transmission bandwidth (e.g., 4G / 5G networks), the extracted video clip is encoded with a high compression rate using H.265 encoding. Finally, the tag information and the compressed data clip are packaged and uploaded to the cloud server via the wireless network module. Most of the data that does not detect potential targets is directly discarded at the edge, thus achieving large-scale data filtering and effective uploading.

[0084] Through the above steps, real-time analysis and filtering can be performed at the data source, achieving efficient transformation from a large amount of raw data to effective target segments. This greatly reduces the network and cloud load while ensuring that key information is uploaded in real time without any omissions.

[0085] The cloud-based identification and desensitization module is used to receive data fragments from the edge via a cloud server, and to perform fine-grained identification on the data fragments using a second AI model deployed in the cloud. It also outputs the confidence level of the fine-grained identification results and performs automatic and irreversible blurring of the human biometric information and vehicle license plate information identified in the data fragments.

[0086] Furthermore, the process for fine-grained identification and the identification of the confidence level of fine-grained identification results includes:

[0087] The cloud server receives data fragments from the edge and decompresses and parses the data fragments to obtain the data to be analyzed.

[0088] The data to be analyzed is input into a second AI model deployed in the cloud;

[0089] The second AI model extracts features from the data to be analyzed, obtaining a multidimensional feature representation of the data.

[0090] Based on multidimensional feature representation, the second AI model performs at least one fine-grained identification task among species-level classification, individual identification, or abnormal behavior analysis;

[0091] The second AI model outputs the recognition results and corresponding confidence scores for the fine-grained recognition task.

[0092] Furthermore, the blurring process includes:

[0093] A preset privacy filtering model is invoked to scan the data to be analyzed or the data fragments after fine-grained identification;

[0094] The privacy filtering model locates and selects all human facial regions, human body contour regions, and vehicle license plate regions in the data to be analyzed.

[0095] For each selected area, a pixelation or mosaic algorithm is applied to blur the rendering. The rendering process permanently destroys the original image details of the area, making them irreversible.

[0096] The data segments that have completed the blurring rendering are marked as de-identified data and correlated with the original recognition results and confidence levels.

[0097] Specifically, the cloud server continuously receives data packets uploaded from various edge devices via message queues (such as RabbitMQ or Kafka). Upon receiving the data packets, the service first decompresses them to restore the original MP4 video segments and their metadata tags. Then, it parses the video segments, decoding them into a continuous sequence of image frames. These image frames constitute the data to be analyzed later. The parsed image frames are then fed into a second AI model deployed in the cloud. This model is a large, deep convolutional neural network, such as ResNet-152, Inception-V4, or a Custom CNN optimized for wildlife. Its computational accuracy and complexity are far higher than the first AI model on the edge side. The model first extracts features from the input image, transforming it into a high-dimensional, multi-dimensional feature vector (e.g., a 2048-dimensional vector) containing rich semantic information through multiple convolutional and pooling layers. Subsequently, based on this feature vector, different classification heads at the top of the model... The Head performs one or more of the following fine-grained recognition tasks in parallel: species-level classification, individual identification, and anomalous behavior analysis. Species-level classification determines the specific species of an animal in an image, such as accurately distinguishing between a Siberian tiger and a South China tiger, or a leopard and a snow leopard. Individual identification identifies individuals with unique markings (such as tiger stripes or leopard spots) by comparing them to a pre-stored individual feature database to determine their specific individual number (e.g., Siberian tiger-male-number 003). Anomalous behavior analysis analyzes animal behavior patterns, such as identifying lameness, group mortality, aggressive behavior, or other abnormal behaviors. In cases of unusual situations such as trespassing into residential areas, for each identification task, the second AI model outputs a specific identification result and its corresponding confidence score (a probability value between 0 and 1). In this embodiment, the confidence score is the inherent probability estimate of the reliability of the fine-grained identification results by the second AI model (a deep convolutional neural network). Its core originates from the Softmax function of the model's output layer. Specifically, for a classification task (such as species identification), the last fully connected layer of the model outputs an unnormalized score for each possible category (N categories in total), called Logits, denoted as z. i (where i = 1, 2, ..., N), the Softmax function transforms these Logits into a probability distribution. The confidence level P for a certain class i... i This is the output value of the category after Softmax activation, and its calculation formula is:

[0098]

[0099] Among them, P i This refers to the confidence level of the model predicting that the target belongs to category i, z iIt is the Logit value output by the model for class i, where e is the natural constant. This is an exponential summation of Logits for all categories (from j=1 to j=N), ensuring that the sum of the confidence scores for all categories is 1. This function maps the model's original output to a (0,1) interval, with the sum of the probabilities for all categories being 1. The confidence score P... i The higher the confidence level, the more confident the model is that its prediction result is category i. This confidence level is the key basis for subsequent judgment of information reliability and triggering of workflow.

[0100] After the second AI model performs fine-grained recognition (or in another parallel thread, before recognition of the original data to be analyzed), the system calls a preset privacy filtering model. This model is a combination of an efficient face detection model (such as MTCNN) and a general object detection model (such as YOLOv4). This model scans each frame of the data segment, specifically to locate and bound all human facial regions, complete human body contour regions, and vehicle license plate regions. For each bounded privacy region, the system applies image processing algorithms for rendering; specifically, it uses pixelation. The pixelation process, which divides the target area into several large pixel blocks, fills each pixel block with the average color of the color in that area. The mosaic process, on the other hand, uses a fixed square pattern to cover the area. Both of these processing methods permanently destroy the original image details of the area, making the processed information such as faces and license plates computationally irreversible, thus achieving irreversible blurring. After the blurring rendering is completed, the entire data segment is marked as desensitized data. The system associates and binds this desensitized data with the original recognition results (species, individuals, behaviors, etc.) output by the second AI model and their confidence levels, and stores it in the database. This ensures that the data used for subsequent display, review and training has been desensitized, and at the same time, it can accurately correspond to the original analysis results when traceability is required.

[0101] Through the above steps, cloud computing power can be used to achieve fine-grained identification down to the species and individual level, while simultaneously completing irreversible desensitization of privacy information. This ensures data security and human privacy while deeply mining the value of data.

[0102] The manual verification and data management module is used to build a central training dataset. Data segments with a confidence level lower than a preset threshold for fine-grained identification by the second AI model, or data segments that have been blurred, are pushed to the manual review queue. Authorized personnel verify and correct the data segments in the manual review queue, generate standard data that has been manually verified, and add the standard data to the central training dataset.

[0103] Furthermore, the process of constructing the central training dataset includes:

[0104] Create an initially empty database on the cloud server and define it as the central training dataset;

[0105] Initialize the version number of the central training dataset and create a data table to store data entries. The data table has predefined fields for recording metadata.

[0106] Receives verified and corrected standard data from the human review queue;

[0107] Generate corresponding metadata for each piece of standard data. The metadata should include at least the data source, collection time, labeled species information, labeler identification, and the corresponding original AI recognition result.

[0108] The standard data and its metadata are added to the central training dataset as a complete data entry, and the dataset version number is updated.

[0109] Regularly perform deduplication, quality checks, and class balance analysis on the data in the central training dataset to maintain the quality of the dataset.

[0110] Furthermore, the process for generating standard data that has undergone manual verification includes:

[0111] The cloud server filters out data segments with confidence levels below a preset confidence threshold based on the fine-grained identification of the second AI model, and stores them together with the data segments that have been blurred into a preset data pool to be reviewed.

[0112] Extract data fragments sequentially from the data pool to be reviewed and assign them to the review workbench of an authorized person.

[0113] The data snippets are presented to authorized personnel at the review workbench, along with the preliminary identification results and confidence levels provided by the second AI model.

[0114] Authorized personnel shall confirm, correct or supplement the preliminary identification results based on observation, and the annotation information shall include at least the correct species category, individual identifier or behavioral description;

[0115] The system receives annotation information submitted by authorized personnel, binds the annotation information with the corresponding data segment, generates a standard data that has been manually verified, and marks the standard data as available.

[0116] Specifically, the cloud server sets a preset confidence threshold (e.g., 0.85). The system automatically filters out data fragments with a confidence level below this threshold that are identified by the second AI model in a fine-grained manner. These are usually blurry images, rare species, or uncommon behaviors that the model has difficulty judging. At the same time, all data fragments that have been blurred (because they involve people and vehicles and have special audit value) are also included. The above two types of data are stored together in a preset data pool to be audited. Technically, this data pool can be a database table (such as a MySQL table) or a task queue (such as a Redis List). Its structure includes fields such as data fragment storage path and metadata pointer.

[0117] The system extracts data fragments from the pending data pool sequentially according to a first-in, first-out (FIFO) or other priority strategy. These fragments are then distributed to an authorized personnel's review workbench via WebSocket or polling API. This workbench is a web-based visual interface that displays the data fragments (videos) along with the preliminary identification results and confidence levels provided by a second AI model. For example, the interface might display: AI identification: Leopard, confidence level 72%. The authorized personnel (typically experienced rangers or biologists) then confirm, correct, or supplement the AI's preliminary results based on their expertise and observations. Operations are performed through dropdown menus, text boxes, and selection tools on the interface. Labeling information must include at least the correct species category (e.g., correcting a feline to a leopard cat), individual identification (e.g., designating individual F-02 from a leopard spot database), and behavioral description (e.g., labeling as foraging or injured).

[0118] Once authorized personnel submit annotation information, the system's backend service receives this information and binds it to the corresponding original data fragment. After binding, the system generates a standard data set that has been manually verified. Logically, this standard data set is a structured object containing data fragment pointers and precise annotation information, and it is marked as available, indicating that it is ready to be used in subsequent processes. On a cloud server (e.g., using a MySQL or PostgreSQL database), an initially empty database is created and defined as the central training dataset. The system initializes the version number of this dataset (e.g., V1.0.0) and creates a data table to store data entries. This data table has predefined fields for recording metadata, including but not limited to: data_id (data ID), file_path (file path), species_label (species label), annotator_id (annotator ID), original_ai_result (original AI result), etc.

[0119] The system receives verified and calibrated standard data from a manually reviewed queue. For each piece of standard data, the system automatically generates corresponding metadata and fills it into the relevant fields of the data table. The metadata includes at least the data source (collection device ID), collection time, labeled species information, annotator identification, and the corresponding original AI recognition result and confidence level. Subsequently, the standard data file (such as a video clip) and its complete metadata are added to the central training dataset as a complete data entry. Each time a new batch of data is added, the system automatically updates the dataset version number (e.g., upgrading from V1.0.0 to V1.1.0) to ensure consistency. Currently, strict version control is implemented, and the system regularly (e.g., weekly) uses an independent maintenance script to automatically maintain the data in the central training dataset. This includes deduplication, quality verification, and class balance analysis. Specifically, deduplication involves removing duplicate or highly similar data based on the hash value or feature similarity of video content to prevent model overfitting. Quality verification involves checking the integrity of data files and the standardization of annotation information. Class balance analysis involves counting the amount of data for each species label and marking rare species with insufficient data to prompt attention in subsequent sampling or active monitoring, thereby maintaining the quality of the dataset and mitigating data bias.

[0120] Through the above steps, the uncertain results of machines can be accurately verified by introducing human expert experience, and high-quality datasets can be systematically built and managed, providing a reliable source of knowledge and updated accuracy for the continuous optimization of the entire system.

[0121] The model training module is used to periodically perform incremental training and fine-tuning of the second AI model using the updated central training dataset to generate an updated version of the second AI model.

[0122] Furthermore, the version update process for the second AI model includes:

[0123] The version number of the central training dataset is checked at a preset period. If the version number is determined to be updated, all the standard data newly added since the previous training period are obtained and divided into an incremental training set and an incremental validation set according to a preset ratio.

[0124] Load the weight parameters of the currently deployed second AI model as the initial model for incremental training;

[0125] The initial model is trained using an incremental training set. During training, a small learning rate is used and some of the lower network layers of the model are frozen. Only the weights of some of the higher network layers of the model are updated to obtain a candidate model.

[0126] The performance of candidate models is evaluated using an incremental validation set. When the performance improvement reaches a preset standard, the candidate model is deemed qualified.

[0127] The qualified candidate models are saved as updated versions of the second AI model and bound to the new version identifier, while the old model versions are archived and stored.

[0128] Specifically, the system automatically executes training tasks according to a preset cycle (e.g., once every two weeks). After the task starts, it first checks the version number of the central training dataset and compares it with the version number recorded during the last successful training. If it is determined that the version number has been updated, it indicates that there is new content in the dataset, and the incremental training process is triggered. The system obtains all the standard data newly added since the last training cycle, which constitutes the original data pool for this training. Subsequently, the system divides this batch of data into an incremental training set and an incremental validation set according to a preset ratio (e.g., 80% for training and 20% for validation). The validation set here is specifically used to evaluate the model's mastery of new knowledge, ensuring the timeliness and fairness of the evaluation.

[0129] The system loads the weight parameters of the currently deployed second AI model and uses it as the initial model for this incremental training. During training, a smaller learning rate is used, specifically, a learning rate of an order of magnitude smaller than that used during the initial training (e.g., an initial learning rate of 1e-3 and an incremental learning rate of 1e-5). This allows for fine-tuning of the model weights to adapt to new data without destroying the original features. Some of the model's lower-level network layers are frozen, meaning their weight parameters are fixed and they do not participate in this backpropagation update. This is because these lower-level layers are typically responsible for extracting general features (such as edges and textures). Only the weights of some higher-level network layers are updated. The system concentrates its computational resources on updating the weights of the higher-level layers (usually fully connected layers or the last convolutional blocks). These higher-level layers are responsible for making decisions related to specific tasks and are key to adapting to new species and scenarios.

[0130] After completing the above training, a candidate model is obtained. The system then uses an incremental validation set to evaluate the performance of the candidate model. The evaluation metric is usually the macro-F1 score, which can better measure the overall performance of the model in various categories (especially rare species with limited data). When the performance improvement reaches a preset standard (for example, the macro-F1 score on the validation set is improved by more than 0.5 percentage points compared to the old model), the system determines that the candidate model is qualified. For qualified models, the system performs version management, saves it as the updated version of the second AI model, and binds it with the new version identifier (for example, the version number is upgraded from v2.1.0 to v2.2.0), while the old model version is archived and stored.

[0131] Through the above steps, the periodic incremental training and fine-tuning strategy enables the core AI model to continuously absorb knowledge from new data, thereby achieving self-iteration and stable performance improvement, and preventing the model from degrading due to environmental changes.

[0132] The model deployment and early warning module is used to perform lightweight processing on the updated version of the second AI model obtained after fine-tuning, and distribute it to the edge to update the first AI model. At the same time, based on the fine-grained recognition results of the second AI model or the confirmation results of the manual review queue, it generates and issues early warning information of the corresponding level.

[0133] Furthermore, the update process for the first AI model includes:

[0134] The second AI model, which is deemed to be qualified, is subjected to lightweight processing, which includes at least one of model pruning, quantization, or knowledge distillation, to generate a lightweight model file.

[0135] Package the lightweight model file and model version metadata to generate a model update package. The model version metadata includes at least the model version number, applicable edge device type, and performance metrics.

[0136] The model update package is downloaded from the cloud server to the target edge device;

[0137] The edge device receives and parses the model update package, and replaces the original first AI model with the lightweight model file in the package;

[0138] After the replacement is completed, the edge device performs integrity verification and performance testing on the new first AI model. Once the test is passed, it sends a confirmation of successful update to the cloud server.

[0139] Furthermore, the process for generating and issuing warning information at the corresponding level includes:

[0140] Receive fine-grained recognition results and their confidence levels from a second AI model on a cloud server, or receive confirmed final results from a human review queue;

[0141] Based on the preset early warning rule base, the received results are parsed to determine the type and level of the early warning event. The early warning event type includes at least poaching, appearance of specific protected animals, abnormal animal gathering or risk of human-wildlife conflict, and the early warning level includes at least emergency, important and general.

[0142] Based on the determined warning event type and warning level, the corresponding warning information template is called from the preset warning template library, and the event time, location, target species, quantity and confidence information are automatically filled in to generate structured warning information;

[0143] Based on the warning level, select at least one preset warning channel to send structured warning information;

[0144] After the warning information is issued, the system records the time of issuance, content, recipient and channel of the warning, and waits for feedback information from the patrol personnel.

[0145] Specifically, for the updated version of the second AI model that has been deemed qualified by the above criteria, the system first performs lightweight processing to adapt to the limited computing resources of edge devices. The processing includes at least one of the following techniques, including model pruning, quantization, and knowledge distillation. Model pruning involves removing redundant weights or neural connections in the model to generate a sparser and smaller network structure. Quantization involves converting the model weights from 32-bit floating-point numbers to 8-bit integers to significantly reduce the model size and memory usage. Knowledge distillation involves using the large second AI model as a teacher model to guide a pre-designed small student model to learn, and finally deploying the small model. After processing, a lightweight model file is generated, which can be in the format of TFLite, ONNX, or OpenVINOIR, so as to enable efficient inference on edge devices.

[0146] The system packages the lightweight model file and model version metadata into a model update package. The metadata includes at least the model version number, applicable edge device type, and performance metrics. Then, the model update package is sent from the cloud server to the target edge device via a secure network connection (such as HTTPS or MQTTS). The sending strategy can be silent push or the edge device can pull the package periodically. The target edge device receives and parses the model update package. Before updating, the original first AI model is usually backed up. After parsing, the original first AI model is replaced with the lightweight model file in the package. After the replacement, the edge device does not immediately enable the new model. Instead, it performs integrity verification (such as calculating the file hash value) and performance testing (using a set of standard test images for forward inference to check for crashes and whether the speed meets the standards) on the new first AI model. After the test is passed, the edge device officially enables the new model and returns a confirmation of successful update to the cloud server, completing this iteration deployment.

[0147] The pre-set early warning engine receives fine-grained recognition results and their confidence levels from a second AI model on a cloud server, as well as confirmed final results from a human review queue. Based on a pre-set early warning rule base, the system parses and logically judges the received results to determine the type and level of the early warning event. Early warning event types include at least poaching (identification of humans carrying guns), appearance of specific protected animals (e.g., Siberian tigers), and abnormal animal gatherings or risks of human-wildlife conflict (e.g., elephant herds approaching villages). Early warning levels include at least emergency (e.g., poaching, wild animals entering residential areas), important (e.g., appearance of protected animals), and general (e.g., common animal activities). Based on the determined early warning event type and level, the system retrieves the corresponding early warning information template from a pre-set early warning template library. The template is a predefined text structure, such as "[Emergency] [Target species] discovered at [Location], please immediately..." In response, the engine automatically populates variables in the template, including event time, location, target species, quantity, and confidence level information, to generate structured early warning information. Subsequently, based on the early warning level, it selects at least one preset early warning channel to issue structured early warning information. For emergency levels, it will issue the information simultaneously via AppPush, SMS, and the dispatch system (directly dispatched to the patrol vehicle terminal) to ensure the highest delivery rate. For important levels, it will issue the information via AppPush and SMS. For general levels, it will issue the information only via AppPush to avoid excessive interference. After the early warning information is issued, the system records the issuance time, content, recipient, and channel of the early warning in the log database. At the same time, the system will wait for feedback information from patrol personnel (such as whether it has been handled, whether it is a false alarm, or whether support is needed). This feedback is recorded and associated with the original early warning to form an early warning closed loop, which is used to evaluate the effectiveness of the system and optimize the early warning rules.

[0148] Through the above steps, the cloud-based evolution model can be lightweightly deployed to the edge and trigger tiered early warnings, realizing a closed loop of system intelligence from cloud-based decision-making to edge execution and precise intervention, ultimately transforming the analysis results into practical protection actions.

[0149] Example 2:

[0150] When encountering animal postures not fully covered in the training data, new environmental backgrounds, lighting conditions, or rare species, the model's recognition accuracy drops significantly, and the system lacks effective self-optimization capabilities to overcome these problems. This leads to high false positive and false negative rates, greatly reducing the reliability of the monitoring system and failing to meet the needs of high-precision early warning and long-term scientific research monitoring. To address these issues, this invention provides a wildlife AI monitoring and early warning method, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this method is as follows:

[0151] By continuously collecting environmental data through monitoring equipment deployed in the field, and using the first AI model deployed on the edge to analyze the environmental data in real time, the system detects whether there are potential targets in the data. When a potential target is identified, the data segment containing the potential target is marked and uploaded to the cloud server.

[0152] The system receives data segments from the edge via a cloud server and uses a second AI model deployed in the cloud to perform fine-grained identification on the data segments. It also outputs the confidence level of the fine-grained identification results and performs automatic and irreversible blurring on the human biometric information and vehicle license plate information identified in the data segments.

[0153] A central training dataset is constructed. Data segments with confidence levels below a preset threshold or data segments that have been blurred by the fine-grained identification of the second AI model are pushed to a manual review queue. Authorized personnel verify and correct the data segments in the manual review queue, generate standard data that has been manually verified, and add the standard data to the central training dataset.

[0154] The second AI model is incrementally trained and fine-tuned regularly using the updated central training dataset to generate an updated version of the second AI model;

[0155] The updated version of the second AI model, after fine-tuning, is lightweighted and sent to the edge to update the first AI model. At the same time, based on the fine-grained recognition results of the second AI model or the confirmation results of the manual review queue, a warning message of the corresponding level is generated and issued.

[0156] Specifically, by having monitoring equipment deployed in the field work in collaboration with the first AI model on the edge, real-time analysis of environmental data and preliminary screening of potential targets are achieved. Valid data fragments are then uploaded to the cloud, where a more powerful second AI model performs fine-grained identification, simultaneously outputting the confidence level of the identification results. Irreversible blurring of human biometric features and vehicle license plate information is applied to protect privacy. The system constructs a central training dataset, pushing low-confidence identification results and de-identified data to a manual review queue. After verification and annotation by authorized personnel, standard data is generated to enrich the dataset. Subsequently, the updated dataset is used periodically for incremental training and fine-tuning of the second AI model, generating an optimized new version model. Finally, the lightweight new model is deployed to the edge to update the first AI model. Based on the fine-grained identification results or manual confirmation results, tiered early warning information is generated and issued, thus forming an autonomously evolving intelligent monitoring and early warning system integrating perception, analysis, learning, decision-making, and intervention.

[0157] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wildlife AI monitoring and early warning system, characterized in that, Includes the following modules: The data acquisition and edge detection module is used to continuously collect environmental data through monitoring equipment deployed in the field, and to use the first AI model deployed on the edge to analyze the environmental data in real time to detect whether there are potential targets in the data. When a potential target is identified, the data segment containing the potential target is marked and uploaded to the cloud server. The cloud-based identification and desensitization module is used to receive data fragments from the edge via a cloud server, and to use a second AI model deployed in the cloud to perform fine-grained identification on the data fragments. It also outputs the confidence level of the fine-grained identification results and performs automatic and irreversible blurring of the human biometric information and vehicle license plate information identified in the data fragments. The manual verification and data management module is used to construct a central training dataset. Data segments with a confidence level lower than a preset threshold for fine-grained recognition by the second AI model or data segments that have been blurred are pushed to the manual review queue. Authorized personnel verify and correct the data segments in the manual review queue, generate standard data that has been manually verified, and add the standard data to the central training dataset. The model training module is used to periodically perform incremental training and fine-tuning on the second AI model using the updated central training dataset to generate an updated version of the second AI model. The model deployment and early warning module is used to perform lightweight processing on the updated version of the second AI model obtained after fine-tuning, and send it to the edge to update the first AI model. At the same time, based on the fine-grained recognition results of the second AI model or the confirmation results of the manual review queue, it generates and issues early warning information of the corresponding level.

2. The wildlife AI monitoring and early warning system according to claim 1, characterized in that, The process for identifying potential targets includes: By deploying at least one of infrared cameras, visible light cameras, and sound sensors in the field, environmental data including video and audio streams is continuously collected. The collected environmental data is preprocessed, including at least noise reduction, format standardization, and keyframe extraction. The preprocessed environmental data is input into the first AI model deployed at the edge, and the first AI model analyzes the data to detect whether it contains potential targets that meet the preset characteristics of animals, humans or vehicles. When the first AI model detects a potential target, it extracts a data segment from the environmental data that contains the time period in which the potential target appeared, and then marks and compresses the data segment. Upload the tagged and compressed data fragments to the cloud server.

3. The wildlife AI monitoring and early warning system according to claim 1, characterized in that, The process for determining fine-grained identification and the confidence level of the fine-grained identification results includes: The cloud server receives data segments from the edge and decompresses and parses the data segments to obtain the data to be analyzed. The data to be analyzed is input into the second AI model deployed in the cloud; The second AI model extracts features from the data to be analyzed, obtaining a multidimensional feature representation of the data; Based on the multidimensional feature representation, the second AI model performs at least one fine-grained identification task among species-level classification, individual identification, or abnormal behavior analysis; The second AI model outputs the recognition results and corresponding confidence scores for the fine-grained recognition task.

4. The wildlife AI monitoring and early warning system according to claim 3, characterized in that, The blurring process includes: A preset privacy filtering model is invoked to scan the data to be analyzed or the data fragments after fine-grained identification; The privacy filtering model locates and selects all human facial regions, human body contour regions, and vehicle license plate regions in the data to be analyzed. For each selected area, a pixelation or mosaic algorithm is applied to blur the rendering. The rendering process permanently destroys the original image details of the area, making them irreversible. The data segments that have completed the blurring rendering are marked as de-identified data and correlated with the original recognition results and confidence levels.

5. A wildlife AI monitoring and early warning system according to claim 1, characterized in that, The process of constructing the central training dataset includes: Create an initially empty database on the cloud server, which is defined as the central training dataset; Initialize the version number of the central training dataset and create a data table for storing data entries, the data table having predefined fields for recording metadata; Receive the standard data, which has been verified and corrected, from the human review queue; Generate corresponding metadata for each piece of standard data. The metadata includes at least the data source, collection time, labeled species information, labeler identification, and corresponding original AI recognition results. The standard data and its metadata are added to the central training dataset as a complete data entry, and the dataset version number is updated. Regularly perform deduplication, quality verification, and class balance analysis on the data in the central training dataset to maintain the quality of the dataset.

6. The wildlife AI monitoring and early warning system according to claim 1, characterized in that, The process for generating the manually verified standard data includes: The cloud server filters out data segments whose confidence level is lower than the preset confidence threshold based on the fine-grained recognition of the second AI model, and stores them together with the data segments that have been blurred into a preset data pool to be reviewed. Data fragments are extracted sequentially from the data pool to be reviewed and assigned to the review workbench of an authorized person. The data fragments are displayed to authorized personnel on the review workbench, along with the preliminary identification results and confidence levels provided by the second AI model. Authorized personnel shall confirm, correct or supplement the preliminary identification results based on observation, and the annotation information shall include at least the correct species category, individual identifier or behavioral description; The system receives annotation information submitted by authorized personnel, binds the annotation information with the corresponding data segment, generates a standard data that has been manually verified, and marks the standard data as available.

7. The wildlife AI monitoring and early warning system according to claim 1, characterized in that, The version update process for the second AI model includes: The version number of the central training dataset is checked at a preset period. If it is determined that the version number has been updated, all the standard data newly added since the previous training period are obtained and divided into an incremental training set and an incremental validation set according to a preset ratio. Load the weight parameters of the currently deployed second AI model as the initial model for incremental training; The initial model is trained using the incremental training set. During the training process, a small learning rate is used and some of the lower network layers of the model are frozen. Only the weights of some of the higher network layers of the model are updated to obtain a candidate model. The incremental validation set is used to evaluate the performance of the candidate model. When the performance improvement reaches a preset standard, the candidate model is judged to be qualified. The candidate models that are deemed qualified are saved as updated versions of the second AI model and bound to the new version identifier, while the old model versions are archived and stored.

8. A wildlife AI monitoring and early warning system according to claim 7, characterized in that, The update process for the first AI model includes: The updated version of the second AI model that is deemed qualified is subjected to lightweight processing, which includes at least one of model pruning, quantization, or knowledge distillation, to generate a lightweight model file. The lightweight model file and model version metadata are packaged together to generate a model update package. The model version metadata includes at least the model version number, the applicable edge device type, and performance indicators. The model update package is downloaded from the cloud server to the target edge device; The edge device receives and parses the model update package, and replaces the original first AI model with the lightweight model file in the package; After the replacement is completed, the edge device performs integrity verification and performance testing on the new first AI model. Once the test is passed, it sends a confirmation of successful update to the cloud server.

9. A wildlife AI monitoring and early warning system according to claim 1, characterized in that, The process for generating and issuing the corresponding level of early warning information includes: Receive fine-grained recognition results and their confidence levels from the second AI model from the cloud server, or receive the confirmed final results from the manual review queue; Based on a pre-set early warning rule base, the received results are parsed to determine the type and level of the early warning event. The types of early warning events include at least poaching, appearance of specific protected animals, abnormal animal gatherings, or human-wildlife conflict risks. The early warning levels include at least emergency, important, and general. Based on the determined warning event type and warning level, the corresponding warning information template is called from the preset warning template library, and the event time, location, target species, quantity and confidence information are automatically filled in to generate structured warning information; Based on the warning level, at least one preset warning channel is selected to send the structured warning information; After the warning information is issued, the system records the time of issuance, content, recipient and channel of the warning, and waits for feedback information from the patrol personnel.

10. A method for AI-based monitoring and early warning of wild animals, characterized in that, The method for a wildlife AI monitoring and early warning system according to any one of claims 1-9 includes the following steps: Environmental data is continuously collected by monitoring equipment deployed in the field, and the environmental data is analyzed in real time using a first AI model deployed at the edge to detect whether there are potential targets in the data. When a potential target is identified, the data segment containing the potential target is marked and uploaded to the cloud server. The system receives data segments from the edge via a cloud server and uses a second AI model deployed in the cloud to perform fine-grained identification on the data segments. It also outputs the confidence level of the fine-grained identification results and performs automatic and irreversible blurring on the human biometric information and vehicle license plate information identified in the data segments. A central training dataset is constructed. Data segments with a confidence level lower than a preset threshold or data segments that have been blurred by the fine-grained recognition of the second AI model are pushed to a manual review queue. Authorized personnel verify and correct the data segments in the manual review queue, generate standard data that has been manually verified, and add the standard data to the central training dataset. The second AI model is incrementally trained and fine-tuned periodically using the updated central training dataset to generate an updated version of the second AI model; The updated version of the second AI model, after fine-tuning, is lightweighted and sent to the edge to update the first AI model. At the same time, based on the fine-grained recognition results of the second AI model or the confirmation results of the manual review queue, a warning message of the corresponding level is generated and issued.

Images