Tracking monitoring system based on artificial intelligence
Through a tracking and monitoring system based on artificial intelligence, GAN generates high-quality data and deep learning models, combining data augmentation and manual auditing, the problems of insufficient labeled data and inaccurate behavioral analysis in traditional monitoring systems are solved, and animal identification and behavior are improved Accuracy of the analysis.
Patent Information
- Application Number
- CN202411051181.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional monitoring systems realize animal recognition through animal face shape data annotation and training, but the accuracy is affected by the quantity and quality of the labeled data, and behavioral analysis is affected by factors such as environment, health and psychological state, resulting in inaccurate identification and analysis.
Adopting a tracking and monitoring system based on artificial intelligence, including data generation and enhancement module, automated data annotation module, behavior feature extraction module, behavior trajectory analysis and pattern recognition module, real-time data processing and monitoring module, and visualization and interactive interface module, GAN generation High-quality data, deep learning models and pre-trained classification models, combined with data augmentation and manual auditing, optimize feature extraction and behavioral analysis.
It improves the accuracy of animal face shape recognition and the accuracy of behavioral analysis, enhances the generalization ability of the model, can handle diverse animal breeds and behavior patterns, and achieves more accurate animal behavior monitoring and analysis.
Smart Images

Figure CN119028013A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to a tracking and monitoring system based on artificial intelligence. Background Art
[0002] Animal behavior refers to a series of movements and postures that animals display under specific environmental conditions in order to meet their physiological needs (such as feeding, excretion, temperature regulation, survival and reproduction, etc.). With the development of science and technology, especially the progress of artificial intelligence (AI) and machine learning (ML), scientists have been able to develop AI-based tracking and monitoring systems to study animal behavior more accurately and efficiently.
[0003] Although traditional monitoring systems have achieved accurate identification of animals through animal facial data annotation and training, the accuracy of this method is limited by the quantity and quality of the annotation data. If the annotation data is insufficient or biased, the recognition accuracy may decrease. In addition, the behavior analysis function of the traditional monitoring system relies on the monitoring of animal behavior trajectories and the identification of activity areas. Animal behavior is affected by many factors, such as environment, health status, psychological state, etc. These factors may lead to inaccurate behavior analysis. Therefore, a tracking and monitoring system based on artificial intelligence is proposed. Summary of the invention
[0004] The purpose of the present invention is to solve the problems existing in the prior art. Traditional monitoring systems achieve accurate identification of animals through animal facial data annotation and training, but the accuracy of this method is limited by the quantity and quality of the annotation data. If the annotation data is insufficient or there are deviations, it may lead to a decrease in recognition accuracy. In addition, the behavior analysis function of the traditional monitoring system relies on the monitoring of animal behavior trajectories and the identification of activity areas. The behavior of animals is affected by many factors, such as environment, health status, psychological state, etc. These factors may lead to inaccurate behavior analysis. A tracking and monitoring system based on artificial intelligence is proposed.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] A tracking and monitoring system based on artificial intelligence, comprising:
[0007] Data generation and enhancement module: responsible for using the GAN generator network to generate high-quality animal face images and images or video sequences that simulate animal behaviors, expand the training data set, and apply data enhancement technology to increase data diversity;
[0008] Automatic data annotation module: responsible for the preliminary annotation of the generated images using the pre-trained classification model, and manually reviewing and correcting them to ensure the accuracy of data annotation;
[0009] Behavioral feature extraction module: responsible for extracting the behavioral features of animals using deep learning models, and optimizing the robustness and generalization ability of feature extraction models using simulated data generated by GAN;
[0010] Behavior trajectory analysis and pattern recognition module: responsible for modeling and analyzing the animal's trajectory data to capture its behavior pattern. At the same time, it uses the GAN discriminator network to detect abnormal trajectories and realize abnormal behavior detection and analysis.
[0011] Real-time data processing and monitoring module: responsible for receiving and processing images, videos and sensor data from the farm in real time;
[0012] Visualization and interactive interface module: responsible for providing users with a friendly visualization interface to display animal identification results, behavior analysis data and real-time monitoring images. At the same time, it provides interactive functions such as data query and report generation.
[0013] The image generated by the data generation and enhancement module is passed to the automatic data annotation module for annotation. The annotation results are manually reviewed and corrected and then returned to the data generation and enhancement module for further optimizing the quality of data generation. The data annotated by the automatic data annotation module is used as training data for the behavior feature extraction module. The features extracted by the behavior feature extraction module are passed to the behavior trajectory analysis and pattern recognition module for analysis. The data processed by the real-time data processing and monitoring module is passed to the visualization and interactive interface module for display.
[0014] The above technical solution further includes:
[0015] Furthermore, the data generation and enhancement module includes a GAN generator and a data enhancement unit, wherein the GAN generator is responsible for receiving random noise or conditional vectors as input to generate realistic animal face images or image / video sequences simulating animal behaviors, and the data enhancement unit is responsible for performing various transformations on the original data or the data generated by the GAN, such as rotation, scaling, translation, flipping, etc., to increase the diversity of the data, and the random noise or conditional vectors are used as inputs of the GAN generator, and the image / video sequences generated by the GAN generator are passed to the data enhancement unit, and the data processed by the data enhancement unit is added to the training data set for subsequent tasks such as animal recognition and behavior analysis. When generating images, the GAN generator can adjust the network structure and parameters according to different needs to generate images with specific attributes (such as animal faces at different angles and lighting conditions).
[0016] Furthermore, the automatic data annotation module includes an image input unit, a pre-trained classification model unit, a preliminary annotation output unit, a manual review unit and a final annotation output unit. The image input unit is responsible for receiving image data to be annotated, which may come from a GAN generator or other data sources. The pre-trained classification model unit contains one or more pre-trained classification models for preliminary annotation of input images. These models have been trained on a large amount of data and have certain classification capabilities. The preliminary annotation output unit outputs the annotation results of the pre-trained classification model unit for subsequent manual review. The manual review unit is composed of professional data annotators, who are responsible for checking the accuracy of the preliminary annotation results and making necessary corrections. The final annotation output unit outputs the annotation results that have been manually reviewed and corrected as training data. The image input unit passes the image data to be annotated to the pre-trained classification model unit. The pre-trained classification model unit passes the preliminary annotation results to the preliminary annotation output unit. The manual review unit receives the annotation results output by the preliminary annotation output unit, and the reviewer passes the corrected annotation results to the final annotation output unit.
[0017] Furthermore, the behavior feature extraction module includes a data input unit, a deep learning model unit, a feature extraction unit and a model optimization unit. The data input unit is responsible for receiving original image or video data, which may come from real-time monitoring, historical records or simulated data generated by GAN. The deep learning model unit contains one or more pre-trained or customized deep learning models for extracting animal behavior features from input data. The feature extraction unit uses the deep learning model to extract key behavior features from the input data, such as action type, speed, frequency, etc. The model optimization unit iterates and optimizes the deep learning model according to the performance of the simulated data generated by GAN and the actual data to improve its performance. The data input unit transmits the original image or video data to the deep learning model. A deep learning model unit, after receiving the data, the deep learning model unit uses the convolutional neural network and other structures therein to extract features in the image or video, and the deep learning model unit passes the extracted feature information to the feature extraction unit, and the feature extraction unit further processes and analyzes these features to extract key behavioral features. The feature extraction unit passes the extracted behavioral features and the performance information of the model (such as accuracy, recall rate, etc.) to the model optimization unit, and the model optimization unit iterates and optimizes the deep learning model according to this information. The model optimization unit returns the optimized model parameters or structural information to the deep learning model unit, and the deep learning model unit is updated according to this information to better extract behavioral features in the next iteration.
[0018] Furthermore, the feature extraction unit uses a deep learning model to extract key behavioral features from the input data, and the specific steps are:
[0019] Feature reception: Receive extracted feature information from the deep learning model unit, which may include feature maps output by the convolutional layer, output of the fully connected layer, etc.
[0020] Feature processing: further processing and analysis of the received features, which may include operations such as feature selection and feature dimensionality reduction, to extract key behavioral features and use the attention mechanism to determine which features are important for subsequent analysis modules or application scenarios;
[0021] Feature encoding: Convert selected features into a format suitable for subsequent processing and analysis, such as a numerical feature vector. Dimension scaling (such as standardization and normalization) and encoding (such as one-hot encoding, word embedding, etc.) may be required to better adapt to subsequent algorithms or models.
[0022] Feature output: The processed features are output to subsequent analysis modules or application scenarios, such as animal behavior classification, anomaly detection, etc.
[0023] Furthermore, the feature extraction unit utilizes an attention mechanism, comprising the following steps:
[0024] Introduction of attention mechanism: In the feature extraction unit, the attention mechanism is introduced. The attention mechanism can be regarded as a resource allocation strategy, which allows the model to allocate more computing resources to the important parts of the input data;
[0025] Calculate attention weights: For each feature or feature combination, use the attention mechanism to calculate its corresponding attention weight. The attention score reflects the importance of the feature to the current task. The attention weight can be calculated based on a variety of methods, such as sequence-based attention models (such as RNN, LSTM output), convolution-based attention models (such as self-attention mechanism), or through multi-layer perceptron (MLP) learning;
[0026] Feature weighting: Use the calculated attention weights to weight the original features. In this way, the model can pay more attention to those features that are important for subsequent analysis modules or application scenarios.
[0027] Feature integration: The weighted features are integrated to obtain the final feature representation, which can be a feature vector that contains the information of all features, but the importance of different features has been distinguished through the attention mechanism;
[0028] Feature output: The integrated features are output to subsequent analysis modules or application scenarios. Since the features have been weighted according to their importance, subsequent modules can use these features more effectively to perform tasks.
[0029] Model training and optimization: Throughout the process, the parameters of the model (including the deep learning model and the attention mechanism) are optimized through training data, and the model learns how to allocate attention weights to better adapt to the needs of subsequent analysis modules or application scenarios.
[0030] Furthermore, the behavior trajectory analysis and pattern recognition module includes a trajectory data input unit, a time series analysis unit, a graph model construction unit, a GAN anomaly detection unit and a pattern recognition and output unit. The trajectory data input unit is responsible for receiving and processing animal trajectory data from various sensors, GPS devices or other data sources. The time series analysis unit is responsible for performing time series analysis on the input trajectory data to identify the behavior patterns of animals at different time points. The graph model construction unit is responsible for constructing a graph model to describe the behavior patterns of animals and the associations between them based on the results of the time series analysis. The GAN anomaly detection unit is responsible for performing anomaly detection on the trajectory data using the discriminator network of the GAN to identify abnormal trajectories that are significantly different from normal behavior patterns. The pattern recognition and output unit The unit is responsible for identifying the animal's behavior pattern based on the results of time series analysis, graph model construction and GAN anomaly detection, and outputting the results for subsequent analysis or application. The trajectory data input unit passes the collected animal trajectory data to the time series analysis unit, and the time series analysis unit processes and analyzes these data to extract time series features related to the animal behavior. The time series analysis unit passes the extracted time series features to the graph model construction unit, and the graph model construction unit constructs a graph model based on these features to describe the animal's behavior patterns and the associations between them. The trajectory data input unit also passes the data to the GAN anomaly detection unit, and the GAN anomaly detection unit uses the discriminator network of GAN to perform anomaly detection on the trajectory data and identify abnormal trajectories.
[0031] Furthermore, the time series analysis unit data uses an autoregressive moving average model to perform time series analysis on the input trajectory data, and the specific steps are as follows:
[0032] Data preparation: Determine whether the input trajectory data is a stationary non-pure random sequence. Stationarity means that the statistical characteristics of the time series (such as mean and variance) do not change over time. If the data is not stationary, differential processing is required to make it stationary.
[0033] Model identification and order determination: Use the autocorrelation function (ACF) and partial autocorrelation function (PACF) to analyze the processed stationary series. The autocorrelation function (ACF) describes the correlation between the time series and its lagged version. The partial autocorrelation function (PACF) is the correlation between the time series and a specific lagged version when the early data is known. By observing the graphs of the autocorrelation function (ACF) and the partial autocorrelation function (PACF), the order (p and q) of the autoregressive moving average model can be preliminarily determined. By minimizing AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion), the optimal model order (p and q) can be determined. AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) are statistics used to compare the goodness of fit of different models.
[0034] Autoregressive moving average model establishment: Autoregressive moving average model expression: X t =c+φ1X t-1 +φ2X t-2 +...φ p X t-p +ε1+θ1ε t-1 +θ2ε t-2 +...+θ q ε t-q , where X t are the observations at the current time step (e.g., the coordinates of a location), which describe the relationship between the current value of the time series and its past values, φ1, φ2, ..., φ p are the parameters of the autoregressive model, θ1,θ2,...,θ q are the parameters of the moving average model, which describe the relationship between the current value of the time series and the past errors (or residuals, disturbance terms), c is a constant term, ε t is the white noise error term, which is usually assumed to be independent and identically distributed (IID) with a mean of 0, p is the order of the autoregressive part, that is, the number of past observations used in the model, q is the order of the moving average part, that is, the number of past error terms used in the model, and the parameters c, φ1, φ2, ... φ of the model are estimated using methods such as maximum likelihood estimation and least squares. p ,θ1,θ2,...,θ q ;
[0035] Model diagnosis and testing:
[0036] Residual analysis: Check whether the residuals of the fitted model meet the white noise assumption, that is, whether the residuals are independent, have a mean of zero, and a constant variance. Commonly used test methods include the Ljung-Box test, etc.
[0037] Adaptability test: If the residual test passes, it means that the model has good adaptability and can proceed to the next step. If the test fails, it is necessary to re-identify the model and determine the order until a model with good adaptability is found.
[0038] Prediction and result interpretation:
[0039] Prediction: Predict the input trajectory data and predict the trajectory value at a future time point.
[0040] Result interpretation: Based on the prediction results, analyze the future change trend of the trajectory data and explain the possible reasons behind it.
[0041] Furthermore, the real-time data processing and monitoring module includes a data receiving unit, an image recognition unit, a video analysis unit, a sensor data processing unit, a data synchronization and storage unit and a user interface unit. The data receiving unit is responsible for receiving images, videos and sensor data from the farm in real time. The image recognition unit is responsible for processing the received image data and identifying the animal species and individuals in the image through image recognition technology. The video analysis unit is responsible for real-time analysis of the video stream and detecting the animal's behavior patterns, such as walking, foraging, resting, etc., through a behavior analysis algorithm. The sensor data processing unit is responsible for parsing and processing the received sensor data and extracting information related to animal health and environmental status, such as temperature, humidity, etc. The data synchronization and storage unit is responsible for synchronizing the results of image recognition, video analysis and sensor data processing, and storing the data in an appropriate storage system for subsequent query and analysis. The user interface unit provides users with a real-time monitoring interface and data analysis tools so that users can intuitively understand the animal behavior and environmental status of the farm. The data receiving unit transmits the received images, videos and sensor data to the image recognition unit, the video analysis unit and the sensor data processing unit respectively. The image recognition unit, the video analysis unit and the sensor data processing unit transmit their respective processing results to the data synchronization and storage unit, and the data synchronization and storage unit transmits the synchronized data to the user interface unit.
[0042] The present invention has the following beneficial effects:
[0043] 1. In the present invention, GAN can enhance the training data set by generating fictitious data similar to real data. These fictitious data can cover different behavior types and environmental conditions, make up for the problem of insufficient labeled data, and improve the model's recognition accuracy of animal faces. The GAN generator can learn the statistical characteristics of real data and generate realistic samples, which helps to enhance the generalization ability of the model and enable it to handle more diverse animal breeds and age groups.
[0044] 2. In the present invention, GAN can generate animal samples under different behavioral patterns, and use the generated simulated data and real data to jointly train the behavioral feature extraction model, so that it can learn more robust and generalized behavioral features, which helps the model to more accurately identify and analyze the behavioral trajectories and types of animals. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a system block diagram of an artificial intelligence-based tracking and monitoring system proposed by the present invention. DETAILED DESCRIPTION
[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0047] See also Figure 1 As shown, the present invention is a tracking and monitoring system based on artificial intelligence, comprising:
[0048] Data generation and enhancement module: responsible for using the GAN generator network to generate high-quality animal face images and images or video sequences that simulate animal behaviors, expand the training data set, and apply data enhancement technology to increase data diversity;
[0049] Automatic data annotation module: responsible for the preliminary annotation of the generated images using the pre-trained classification model, and manually reviewing and correcting them to ensure the accuracy of data annotation;
[0050] Behavioral feature extraction module: responsible for extracting the behavioral features of animals using deep learning models, and optimizing the robustness and generalization ability of feature extraction models using simulated data generated by GAN;
[0051] Behavior trajectory analysis and pattern recognition module: responsible for modeling and analyzing the animal's trajectory data to capture its behavior pattern. At the same time, it uses the GAN discriminator network to detect abnormal trajectories and realize abnormal behavior detection and analysis.
[0052] Real-time data processing and monitoring module: responsible for receiving and processing images, videos and sensor data from the farm in real time;
[0053] Visualization and interactive interface module: responsible for providing users with a friendly visualization interface to display animal identification results, behavior analysis data and real-time monitoring images. At the same time, it provides interactive functions such as data query and report generation.
[0054] The images generated by the data generation and enhancement module are passed to the automatic data annotation module for annotation. After manual review and correction, the annotation results are returned to the data generation and enhancement module to further optimize the quality of data generation. The data annotated by the automatic data annotation module is used as training data for the behavior feature extraction module. The features extracted by the behavior feature extraction module are passed to the behavior trajectory analysis and pattern recognition module for analysis. The data processed by the real-time data processing and monitoring module is passed to the visualization and interactive interface module for display.
[0055] The working principle of the artificial intelligence-based tracking and monitoring system proposed in the present invention is to use the generator network of GAN to learn the data distribution from the existing data set and generate new, high-quality animal face images or image / video sequences simulating animal behaviors, and to perform diversity enhancement on the generated images through data enhancement techniques (such as rotation, scaling, flipping, etc.) to simulate various conditions in real scenes;
[0056] The generated and enhanced images are passed to the automated data annotation module, and the pre-trained classification model is used to preliminarily annotate the images. The preliminary annotation results are manually reviewed and corrected to ensure the accuracy of the data annotations. The corrected annotation results are returned to the data generation and enhancement module as feedback to further optimize the quality of data generation by adjusting the parameters of the GAN or improving the data enhancement strategy.
[0057] The data annotated by the automatic data annotation module is used as training data for the behavior feature extraction module. The behavior features of animals are extracted from images or videos using a convolutional neural network. The features extracted by the behavior feature extraction module are passed to the behavior trajectory analysis and pattern recognition module to model and analyze the animal's trajectory data to capture its behavior pattern. The GAN discriminator network or other anomaly detection technologies are used to detect abnormal trajectories, thereby realizing abnormal behavior detection and analysis.
[0058] Receive images, videos and sensor data from farms in real time, pre-process and analyze the received data, extract key information, and use trained models (such as behavior feature extraction models and behavior trajectory analysis models) to process real-time data to obtain animal identification results, behavior analysis, etc.
[0059] The data and results processed by the real-time data processing and monitoring module are passed to the visualization and interactive interface module, and the animal identification results, behavior analysis data, real-time monitoring images and other information are displayed on the interface.
[0060] In one embodiment, for the above-mentioned data generation and enhancement module, the data generation and enhancement module includes a GAN generator and a data enhancement unit. The GAN generator is responsible for receiving random noise or conditional vectors as input to generate realistic animal face images or image / video sequences that simulate animal behaviors. The data enhancement unit is responsible for performing various transformations on the original data or the data generated by the GAN, such as rotation, scaling, translation, flipping, etc., to increase the diversity of the data. The random noise or conditional vector is used as the input of the GAN generator, and the image / video sequence generated by the GAN generator is passed to the data enhancement unit. The data processed by the data enhancement unit is added to the training data set for subsequent tasks such as animal recognition and behavior analysis. When generating images, the GAN generator can adjust the network structure and parameters according to different needs to generate images with specific attributes (such as animal faces at different angles and lighting conditions).
[0061] In one embodiment, for the above-mentioned automatic data labeling module, the automatic data labeling module includes an image input unit, a pre-trained classification model unit, a preliminary labeling output unit, a manual review unit and a final labeling output unit. The image input unit is responsible for receiving the image data to be labeled, which may come from a GAN generator or other data sources. The pre-trained classification model unit contains one or more pre-trained classification models for preliminary labeling of the input image. These models have been trained on a large amount of data and have certain classification capabilities. The preliminary labeling output unit outputs the labeling results of the pre-trained classification model unit for subsequent manual review. The manual review unit is composed of professional data labelers who are responsible for checking the accuracy of the preliminary labeling results and making necessary corrections. The final labeling output unit outputs the labeling results that have been manually reviewed and corrected as training data. The image input unit passes the image data to be labeled to the pre-trained classification model unit, and the pre-trained classification model unit passes the preliminary labeling results to the preliminary labeling output unit. The manual review unit receives the labeling results output by the preliminary labeling output unit, and the reviewer passes the corrected labeling results to the final labeling output unit.
[0062] In one embodiment, for the above-mentioned behavior feature extraction module block, the behavior feature extraction module includes a data input unit, a deep learning model unit, a feature extraction unit and a model optimization unit. The data input unit is responsible for receiving original image or video data, which may come from real-time monitoring, historical records or simulated data generated by GAN. The deep learning model unit contains one or more pre-trained or customized deep learning models for extracting animal behavior characteristics from the input data. The feature extraction unit uses the deep learning model to extract key behavior characteristics from the input data, such as action type, speed, frequency, etc. The model optimization unit iterates and optimizes the deep learning model according to the performance of the simulated data generated by GAN and the actual data to improve its performance. The data input unit converts the original image or video data into a pre-trained or customized deep learning model. The initial image or video data is passed to the deep learning model unit. After receiving the data, the deep learning model unit uses structures such as convolutional neural networks to extract features in the image or video. The deep learning model unit passes the extracted feature information to the feature extraction unit. The feature extraction unit further processes and analyzes these features to extract key behavioral features. The feature extraction unit passes the extracted behavioral features and the performance information of the model (such as accuracy, recall rate, etc.) to the model optimization unit. The model optimization unit iterates and optimizes the deep learning model based on this information. The model optimization unit returns the optimized model parameters or structural information to the deep learning model unit. The deep learning model unit updates according to this information to better extract behavioral features in the next iteration.
[0063] In one embodiment, for the feature extraction unit, the feature extraction unit extracts key behavior features from input data using a deep learning model, and the specific steps are:
[0064] Feature reception: Receive extracted feature information from the deep learning model unit, which may include feature maps output by the convolutional layer, output of the fully connected layer, etc.
[0065] Feature processing: further processing and analysis of the received features, which may include operations such as feature selection and feature dimensionality reduction, to extract key behavioral features and use the attention mechanism to determine which features are important for subsequent analysis modules or application scenarios;
[0066] Feature encoding: Convert selected features into a format suitable for subsequent processing and analysis, such as a numerical feature vector. Dimension scaling (such as standardization and normalization) and encoding (such as one-hot encoding, word embedding, etc.) may be required to better adapt to subsequent algorithms or models.
[0067] Feature output: The processed features are output to subsequent analysis modules or application scenarios, such as animal behavior classification, anomaly detection, etc.
[0068] In one embodiment, for the feature extraction unit, the feature extraction unit uses an attention mechanism, including the following steps:
[0069] Introduction of attention mechanism: In the feature extraction unit, the attention mechanism is introduced. The attention mechanism can be regarded as a resource allocation strategy, which allows the model to allocate more computing resources to the important parts of the input data;
[0070] Calculate attention weights: For each feature or feature combination, use the attention mechanism to calculate its corresponding attention weight. The attention score reflects the importance of the feature to the current task. The attention weight can be calculated based on a variety of methods, such as sequence-based attention models (such as RNN, LSTM output), convolution-based attention models (such as self-attention mechanism), or through multi-layer perceptron (MLP) learning;
[0071] Feature weighting: Use the calculated attention weights to weight the original features. In this way, the model can pay more attention to those features that are important for subsequent analysis modules or application scenarios.
[0072] Feature integration: The weighted features are integrated to obtain the final feature representation, which can be a feature vector that contains the information of all features, but the importance of different features has been distinguished through the attention mechanism;
[0073] Feature output: The integrated features are output to subsequent analysis modules or application scenarios. Since the features have been weighted according to their importance, subsequent modules can use these features more effectively to perform tasks.
[0074] Model training and optimization: Throughout the process, the parameters of the model (including the deep learning model and the attention mechanism) are optimized through training data, and the model learns how to allocate attention weights to better adapt to the needs of subsequent analysis modules or application scenarios.
[0075] In one embodiment, for the above-mentioned behavior trajectory analysis and pattern recognition module, the behavior trajectory analysis and pattern recognition module includes a trajectory data input unit, a time series analysis unit, a graph model construction unit, a GAN anomaly detection unit and a pattern recognition and output unit. The trajectory data input unit is responsible for receiving and processing animal trajectory data from various sensors, GPS devices or other data sources. The time series analysis unit is responsible for performing time series analysis on the input trajectory data to identify the behavior patterns of animals at different time points. The graph model construction unit is responsible for constructing a graph model to describe the behavior patterns of animals and the associations between them based on the results of the time series analysis. The GAN anomaly detection unit is responsible for using the discriminator network of the GAN to perform anomaly detection on the trajectory data and identify anomalies that are significantly different from normal behavior patterns. The common trajectory, pattern recognition and output unit is responsible for identifying the animal's behavior pattern according to the results of time series analysis, graph model construction and GAN anomaly detection, and outputs the results for subsequent analysis or application. The trajectory data input unit passes the collected animal trajectory data to the time series analysis unit, which processes and analyzes these data and extracts time series features related to animal behavior. The time series analysis unit passes the extracted time series features to the graph model construction unit, which constructs a graph model based on these features to describe the animal's behavior patterns and the associations between them. The trajectory data input unit also passes the data to the GAN anomaly detection unit, which uses the discriminator network of GAN to perform anomaly detection on the trajectory data and identify abnormal trajectories.
[0076] In one embodiment, for the above-mentioned time series analysis unit, the time series analysis unit data uses an autoregressive moving average model to perform time series analysis on the input trajectory data, and the specific steps are:
[0077] Data preparation: Determine whether the input trajectory data is a stationary non-pure random sequence. Stationarity means that the statistical characteristics of the time series (such as mean and variance) do not change over time. If the data is not stationary, differential processing is required to make it stationary.
[0078] Model identification and order determination: Use the autocorrelation function (ACF) and partial autocorrelation function (PACF) to analyze the processed stationary series. The autocorrelation function (ACF) describes the correlation between the time series and its lagged version. The partial autocorrelation function (PACF) is the correlation between the time series and a specific lagged version when the early data is known. By observing the graphs of the autocorrelation function (ACF) and the partial autocorrelation function (PACF), the order (p and q) of the autoregressive moving average model can be preliminarily determined. By minimizing AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion), the optimal model order (p and q) can be determined. AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) are statistics used to compare the goodness of fit of different models.
[0079] Autoregressive moving average model establishment: Autoregressive moving average model expression: X t =c+φ1X t-1 +φ2X t-2 +...φ p X t-p +ε1+θ1ε t-1 +θ2ε t-2 +...+θ q ε t-q , where X t are the observations at the current time step (e.g., the coordinates of a location), which describe the relationship between the current value of the time series and its past values, φ1, φ2, ..., φ p are the parameters of the autoregressive model, θ1,θ2,...,θ q are the parameters of the moving average model, which describe the relationship between the current value of the time series and the past errors (or residuals, disturbance terms), c is a constant term, ε t is the white noise error term, which is usually assumed to be independent and identically distributed (IID) with a mean of 0, p is the order of the autoregressive part, that is, the number of past observations used in the model, q is the order of the moving average part, that is, the number of past error terms used in the model, and the parameters c, φ1, φ2, ... φ of the model are estimated using methods such as maximum likelihood estimation and least squares. p ,θ1,θ2,...,θ q ;
[0080] Model diagnosis and testing:
[0081] Residual analysis: Check whether the residuals of the fitted model meet the white noise assumption, that is, whether the residuals are independent, have a mean of zero, and a constant variance. Commonly used test methods include the Ljung-Box test, etc.
[0082] Adaptability test: If the residual test passes, it means that the model has good adaptability and can proceed to the next step. If the test fails, it is necessary to re-identify the model and determine the order until a model with good adaptability is found.
[0083] Prediction and result interpretation:
[0084] Prediction: Predict the input trajectory data and predict the trajectory value at a future time point.
[0085] Result interpretation: Based on the prediction results, analyze the future change trend of the trajectory data and explain the possible reasons behind it.
[0086] In one embodiment, for the above-mentioned real-time data processing and monitoring module, the real-time data processing and monitoring module includes a data receiving unit, an image recognition unit, a video analysis unit, a sensor data processing unit, a data synchronization and storage unit and a user interface unit. The data receiving unit is responsible for receiving images, videos and sensor data from the farm in real time. The image recognition unit is responsible for processing the received image data and identifying the animal species and individuals in the image through image recognition technology. The video analysis unit is responsible for real-time analysis of the video stream and detecting the animal's behavior pattern, such as walking, foraging, resting, etc., through a behavior analysis algorithm. The sensor data processing unit is responsible for parsing and processing the received sensor data and extracting information related to the animal's health and environmental status. The data synchronization and storage unit is responsible for synchronizing the results of image recognition, video analysis and sensor data processing, and storing the data in an appropriate storage system for subsequent query and analysis. The user interface unit provides users with a real-time monitoring interface and data analysis tools so that users can intuitively understand the animal behavior and environmental status of the farm. The data receiving unit transmits the received images, videos and sensor data to the image recognition unit, video analysis unit and sensor data processing unit respectively. The image recognition unit, video analysis unit and sensor data processing unit transmit their respective processing results to the data synchronization and storage unit. The data synchronization and storage unit transmits the synchronized data to the user interface unit.
[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A tracking and monitoring system based on artificial intelligence, characterized in that: include: Data generation and enhancement module: responsible for generating animal face images and images or video sequences simulating animal behaviors, expanding the training data set, and applying data enhancement technology to increase data diversity; Automatic data annotation module: responsible for preliminary annotation of generated images using pre-trained classification models; Behavioral feature extraction module: responsible for extracting animal behavioral features using deep learning models; Behavior trajectory analysis and pattern recognition module: responsible for modeling and analyzing the animal's trajectory data to capture its behavior patterns; Real-time data processing and monitoring module: responsible for receiving and processing images, videos and sensor data from the farm in real time; Visualization and interactive interface module: responsible for providing users with a friendly visualization interface; The image generated by the data generation and enhancement module is passed to the automatic data annotation module for annotation. The data annotated by the automatic data annotation module is used as training data for the behavior feature extraction module. The features extracted by the behavior feature extraction module are passed to the behavior trajectory analysis and pattern recognition module for analysis. The data processed by the real-time data processing and monitoring module is passed to the visualization and interactive interface module for display.
2. The tracking and monitoring system based on artificial intelligence according to claim 1, characterized in that: The data generation and enhancement module includes a GAN generator and a data enhancement unit. The GAN generator is responsible for receiving random noise or conditional vectors as input to generate realistic animal face images or image / video sequences that simulate animal behaviors. The data enhancement unit is responsible for performing various transformations on the original data or the data generated by the GAN, such as rotation, scaling, translation, flipping, etc., to increase the diversity of the data. The random noise or conditional vector is used as the input of the GAN generator, and the image / video sequence generated by the GAN generator is passed to the data enhancement unit. The data processed by the data enhancement unit is added to the training data set.
3. The tracking and monitoring system based on artificial intelligence according to claim 1, characterized in that: The automatic data annotation module includes an image input unit, a pre-trained classification model unit, a preliminary annotation output unit, a manual review unit and a final annotation output unit. The image input unit is responsible for receiving the image data to be annotated. The pre-trained classification model unit includes one or more pre-trained classification models for preliminary annotation of the input image. The preliminary annotation output unit outputs the annotation result of the pre-trained classification model unit for subsequent manual review. The manual review unit is responsible for checking the accuracy of the preliminary annotation result. The final annotation output unit outputs the annotation result that has been manually reviewed and corrected. The image input unit passes the image data to be annotated to the pre-trained classification model unit. The pre-trained classification model unit passes the preliminary annotation result to the preliminary annotation output unit. The manual review unit receives the annotation result output by the preliminary annotation output unit. The reviewer passes the corrected annotation result to the final annotation output unit.
4. The tracking and monitoring system based on artificial intelligence according to claim 2, characterized in that: The behavior feature extraction module includes a data input unit, a deep learning model unit, a feature extraction unit and a model optimization unit. The data input unit is responsible for receiving original image or video data. The deep learning model unit contains one or more pre-trained or customized deep learning models for extracting animal behavior features from input data. The feature extraction unit uses the deep learning model to extract key behavior features from the input data. The model optimization unit iterates and optimizes the deep learning model according to the performance of the simulated data and actual data generated by GAN. The data input unit transmits the original image or video data to the deep learning model unit, and the deep learning model unit transmits the extracted feature information to the feature extraction unit. The feature extraction unit transmits the extracted behavior features and the performance information of the model to the model optimization unit. The model optimization unit returns the optimized model parameters or structure information to the deep learning model unit.
5. The tracking and monitoring system based on artificial intelligence according to claim 4, characterized in that: The feature extraction unit uses a deep learning model to extract key behavior features from input data. The specific steps are: Feature reception: receiving extracted feature information from the deep learning model unit; Feature processing: further process and analyze the received features, and use the attention mechanism to determine which features are important for subsequent analysis modules or application scenarios; Feature encoding: converting selected features into a format suitable for subsequent processing and analysis; Feature output: Output the processed features to subsequent analysis modules or application scenarios.
6. The tracking and monitoring system based on artificial intelligence according to claim 5, characterized in that: The feature extraction unit uses an attention mechanism and includes the following steps: Introduction of attention mechanism: In the feature extraction unit, the attention mechanism is introduced; Calculate attention weight: For each feature or feature combination, use the attention mechanism to calculate its corresponding attention weight; Feature weighting: Use the calculated attention weights to weight the original features; Feature integration: Integrate the weighted features to obtain the final feature representation; Feature output: Output the integrated features to subsequent analysis modules or application scenarios.
7. The tracking and monitoring system based on artificial intelligence according to claim 1, characterized in that: The behavior trajectory analysis and pattern recognition module includes a trajectory data input unit, a time series analysis unit, a graph model construction unit, a GAN anomaly detection unit and a pattern recognition and output unit. The trajectory data input unit is responsible for receiving and processing animal trajectory data, the time series analysis unit is responsible for performing time series analysis on the input trajectory data, the graph model construction unit is responsible for constructing a graph model to describe the behavior patterns of animals and the associations between them based on the results of the time series analysis, the GAN anomaly detection unit is responsible for performing anomaly detection on the trajectory data using the discriminator network of GAN, the pattern recognition and output unit is responsible for identifying the behavior patterns of animals based on the results of time series analysis, graph model construction and GAN anomaly detection, and outputting the results, the trajectory data input unit passes the collected animal trajectory data to the time series analysis unit, the time series analysis unit passes the extracted time series features to the graph model construction unit, and the trajectory data input unit simultaneously passes the data to the GAN anomaly detection unit.
8. The tracking and monitoring system based on artificial intelligence according to claim 7, characterized in that: The time series analysis unit data uses an autoregressive moving average model to perform time series analysis on the input trajectory data. The specific steps are: Data preparation: Determine whether the input trajectory data is a stationary non-pure random sequence. If the data is not stationary, differential processing is required; Model identification and order determination: Use the autocorrelation function and partial autocorrelation function to analyze the processed stationary series. By observing the graphs of the autocorrelation function and partial autocorrelation function, the order of the autoregressive moving average model can be preliminarily determined. By minimizing AIC or BIC, the optimal model order (p and q) can be determined. Autoregressive moving average model establishment: Autoregressive moving average model expression: X t =c+φ1X t-1 +φ2X t-2 +...φ p X t-p +ε1+θ1ε t-1 +θ2ε t-2 +...+θ q ε t-q , where X t are the observed values at the current time step, which describe the relationship between the current value of the time series and its past values, φ1, φ2, ..., φ p are the parameters of the autoregressive model, θ1,θ2,...,θ q is the parameter of the moving average model, c is a constant term, ε t is the white noise error term, p is the order of the autoregressive part, q is the order of the moving average part; Model diagnosis and testing: Residual analysis: Check whether the residuals of the fitted model meet the white noise assumption, that is, whether the residuals are independent, have a mean of zero, and a constant variance; Adaptability test: If the residual test passes, it means that the model has good adaptability and can proceed to the next step. If the test fails, it is necessary to re-identify the model and determine the order until a model with good adaptability is found. Prediction and result interpretation: Prediction: Predict the input trajectory data and predict the trajectory value at a future time point. Result interpretation: Based on the prediction results, analyze the future change trend of the trajectory data and explain the possible reasons behind it.
9. The tracking and monitoring system based on artificial intelligence according to claim 1, characterized in that: The real-time data processing and monitoring module includes a data receiving unit, an image recognition unit, a video analysis unit, a sensor data processing unit, a data synchronization and storage unit and a user interface unit. The data receiving unit is responsible for receiving images, videos and sensor data from the farm in real time. The image recognition unit is responsible for processing the received image data and identifying the animal species and individuals in the image through image recognition technology. The video analysis unit is responsible for real-time analysis of the video stream. The sensor data processing unit is responsible for parsing and processing the received sensor data and extracting information related to animal health and environmental status. The data synchronization and storage unit is responsible for synchronizing the results of image recognition, video analysis and sensor data processing. The user interface unit provides users with a real-time monitoring interface and data analysis tools. The data receiving unit transmits the received images, videos and sensor data to the image recognition unit, the video analysis unit and the sensor data processing unit respectively. The image recognition unit, the video analysis unit and the sensor data processing unit transmit their respective processing results to the data synchronization and storage unit, and the data synchronization and storage unit transmits the synchronized data to the user interface unit.
Citation Information
Cited By
Field panda behavior monitoring method and system based on artificial intelligence
CN119863824A
Artificial Intelligence-Based Method and System for Monitoring the Behavior of Wild Giant Pandas
CN119863824B
Video identification method and system for zoo animals
CN120913242A
