Young deer behavior recognition and health monitoring system and method based on edge calculation

Abstract

The invention relates to the technical field of computer vision, deep learning and intelligent animal husbandry, and discloses a young deer behavior recognition and health monitoring system and method based on edge computing, and the system comprises a data collection module which is used for collecting video data of a target young deer in a breeding house of a breeding farm, preprocessing the video data to obtain a preprocessed video; the target detection and tracking module is used for performing target detection and multi-target tracking on each frame of the preprocessed video, and comprises a detection sub-module for identifying a bounding box position and a bounding box confidence score of a target young deer in each frame by adopting an improved YOLO series detector, according to the method, the behavior recognition and health monitoring system oriented to the young deer breeding scene and based on edge calculation is constructed, accurate recognition of young deer behaviors and real-time evaluation and early warning of the health state are achieved, and compared with an existing method depending on manual observation or a traditional single sensor, the method has remarkable advantages.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning and smart animal husbandry, and in particular to a system and method for recognizing the behavior and monitoring the health of young deer based on edge computing. Background Technology

[0002] Currently, the management of enclosed fawns mainly relies on manual patrols and observations. Staff regularly check the fawns' behavior and health status to identify any abnormalities. However, this method has significant limitations: manual inspections are inefficient, lack real-time accuracy, are highly susceptible to subjective judgment, and cannot achieve continuous, uninterrupted, and accurate monitoring of fawns 24 hours a day. When fawns exhibit abnormal behavior or physiological abnormalities, they often fail to be detected immediately, delaying intervention and impacting survival rates and health. As the scale of fawn farming expands, manual methods are no longer sufficient to support the real-time management needs of every fawn in large-scale enclosures, leading to prominent problems such as delayed disease warnings and a lack of behavioral monitoring.

[0003] Meanwhile, the enclosure environment is typically complex and variable, with significant diurnal variations in lighting conditions, dim indoor lighting, and susceptibility to shading. Young fawns are small and live in dense groups, often experiencing occlusion and confusion between individuals with similar appearances. This makes it difficult for traditional vision-based detection and tracking methods to achieve stable and reliable detection accuracy. In actual production environments, factors such as daytime and nighttime lighting variations, differences in camera angles, and confined spaces can all lead to a decline in the recognition performance of young fawns. Directly applying general target detection algorithms to young fawn monitoring can easily result in missed detections, false detections, and tracking drift, making it difficult to guarantee consistency between detection results and actual individual behavior.

[0004] In behavior recognition, traditional classification methods relying on single-frame images struggle to reliably discriminate fine-grained behaviors due to the subtle categories of fawn behavior and the rapid and frequent changes in individual movements. Furthermore, behavior recognition often depends on a large number of labeled samples, but fawn behavior data collection is limited by actual farming conditions, making it difficult for model training samples to fully cover all scene features, leading to insufficient cross-scene adaptability. Existing research has largely focused on behavior recognition datasets and model building for livestock such as sheep and cattle. However, fawns, being smaller, with weaker features, and more frequently subject to occlusion, present greater challenges in behavior recognition, making direct transfer of existing technologies difficult.

[0005] In terms of health monitoring, current methods often rely on human experience or inferring the health status of fawns from single behavioral signals, lacking comprehensive analysis of multidimensional physiological and environmental indicators. Furthermore, traditional health monitoring methods typically require the pre-collection and uploading of large amounts of data to cloud servers for unified processing, increasing network transmission burden and making the monitoring process dependent on network quality stability. This makes it difficult to meet the needs of remote farming scenarios for low-latency, localized intelligent monitoring. The long-term storage and transmission of large-scale data also increases system energy consumption and operating costs, limiting the widespread deployment of intelligent monitoring systems.

[0006] With the development of smart animal husbandry and deep learning technology, video surveillance combined with computer vision provides a feasible technical approach for animal behavior recognition. However, most existing work adopts general object detection and tracking methods, which do not fully consider the special characteristics of fawns in actual feeding scenarios. These methods still have shortcomings in terms of recognition accuracy, tracking stability, low-light adaptability, and multi-source data fusion capabilities. Especially under conditions of poor lighting at night, frequent occlusion, and subtle movements, fawn behavior recognition models are prone to recognition errors, making it difficult to detect health abnormalities in a timely manner. Furthermore, limited computing power at the system level and insufficient inference speed of complex models on edge devices make it difficult to support large-scale real-time monitoring needs.

[0007] In summary, there is still a lack of effective overall solutions for the behavioral recognition, health monitoring, and real-time data processing of enclosed fawns. There is an urgent need for an intelligent fawn monitoring method that combines multi-source data fusion and edge computing to achieve accurate detection, stable tracking, and fine-grained behavioral recognition of fawns, and to conduct health risk assessments by combining physiological and environmental data. This would overcome the shortcomings of traditional methods in terms of real-time performance, accuracy, stability, and scalability, thereby improving the level of intelligence in fawn monitoring and management. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing a system and method for recognizing the behavior and monitoring the health of fawns based on edge computing.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: An edge computing-based system for recognizing and monitoring the behavior and health of fawns, the system comprising: The data acquisition module is used to collect video data of the target fawns in the breeding farm enclosure, and to obtain preprocessed video after preprocessing the video data; The object detection and tracking module performs object detection and multi-object tracking on each frame of the preprocessed video, including: The detection submodule uses an improved YOLO series detector to identify the bounding box location and bounding box confidence score of the target fawn in each frame; The tracking submodule uses the ByteTrack algorithm to establish the identity association of the target fawn through two-stage matching of high-confidence detection boxes and low-confidence detection boxes and Kalman filter prediction, thereby obtaining the image sequence of the target fawn. The behavior recognition module is used to extract the spatial features of the target fawn based on the image sequence using a convolutional neural network, and then feed the spatial features into a time series modeling network for time series analysis in chronological order to identify the behavior category of the target fawn. The multimodal data fusion and health monitoring module is used to acquire the physiological parameters of the target fawn and the environmental data of the enclosure. It integrates the environmental data with the physiological parameters and behavioral categories of the target fawn to construct a health feature vector of the target fawn. The module calculates a health risk score based on the health feature vector through a neural network model. When the risk score exceeds a preset threshold, it outputs a health warning message. The edge computing module is used to run data acquisition modules, target detection and tracking modules, behavior recognition modules, multimodal data fusion and health monitoring modules on embedded computing devices deployed in the farm.

[0010] Preferably, the data acquisition module specifically includes: The video data acquisition unit is used to collaboratively collect multi-channel video data of the target fawn through multiple high-definition visible light cameras and infrared cameras deployed in the breeding farm pens; The preprocessing unit is used to perform image enhancement, noise reduction, illumination equalization, and background modeling on the video data from each stream to obtain the preprocessed video.

[0011] Preferably, the improved YOLO series detector enhances the model's ability to perceive the features of fawns by introducing a coordinate attention mechanism, and adopts the YOLOv8 model suitable for low-light environments to improve the detection accuracy of fawns under low illumination conditions.

[0012] Preferably, the time series modeling network includes one of a bidirectional long short-term memory network and a Transformer model.

[0013] Preferably, the target fawn's behavior categories include standing, walking, eating, drinking, lying down, socializing, and being sick.

[0014] Preferably, the health risk score is calculated using the Sigmoid function for normalization, and the health risk score ranges from 0 to 1.

[0015] Preferably, the edge computing module specifically includes: Embedded computing devices are used to deploy data acquisition modules, target detection and tracking modules, behavior recognition modules, multimodal data fusion and health monitoring modules, and to run the inference computing of the above modules locally on the farm. The model lightweighting unit is used to perform model pruning, model quantization, and inference acceleration optimization on the models in the target detection and tracking module, behavior recognition module, and multimodal data fusion and health monitoring module, so as to reduce the computational load of the model and improve the inference speed at the edge. The real-time processing unit is used to perform local real-time processing on video data, physiological parameters of the target fawns, and environmental data of the enclosure, and generate monitoring results. The communication unit is used to transmit the monitoring results and health warning information to the user terminal for display via a local area network structure.

[0016] Preferably, the user terminal includes a computer monitoring terminal and a mobile terminal.

[0017] A method for identifying and monitoring the behavior and health of fawns based on edge computing, the method comprising: Video data of the target fawns in the farm enclosure was collected, and the video data was preprocessed to obtain preprocessed video. Perform object detection and multi-object tracking on each frame of the preprocessed video, including: An improved YOLO series detector was used to identify the bounding box location and bounding box confidence score of the target fawn in each frame; The ByteTrack algorithm is used to establish the identity association of the target fawn through two-stage matching of high-confidence detection boxes and low-confidence detection boxes and Kalman filter prediction, thereby obtaining the image sequence of the target fawn; Based on the image sequence, the spatial features of the target fawn are extracted using a convolutional neural network. The spatial features are then fed into a time series modeling network in chronological order for time series analysis to identify the behavioral categories of the target fawn. The physiological parameters of the target fawn and the environmental data of the enclosure are obtained. The environmental data and the physiological parameters and behavioral categories of the target fawn are fused to construct the health feature vector of the target fawn. A health risk score is calculated based on the health feature vector through a neural network model. When the risk score exceeds a preset threshold, a health warning message is output. The embedded computing devices deployed in the farm run data acquisition modules, target detection and tracking modules, behavior recognition modules, and multimodal data fusion and health monitoring modules.

[0018] The beneficial effects of this invention are as follows: This invention constructs an edge computing-based behavior recognition and health monitoring system for fawn farming, which enables accurate identification of fawn behavior, real-time assessment and early warning of fawn health status. Compared with existing methods that rely on manual observation or traditional single sensor methods, it has significant advantages.

[0019] 1. Regarding real-time monitoring, this invention employs an edge-deployed inference architecture, enabling the model to complete behavior recognition and anomaly detection within hundreds of milliseconds, achieving true real-time monitoring. When fawns exhibit abnormal behavior, the system can issue a health risk alert immediately, shortening the time to disease detection by several hours to several days compared to manual patrols, significantly improving the early detection rate. According to the description, this invention can increase the average early detection rate of diseases in fawns by approximately 40%.

[0020] 2. Regarding recognition accuracy, improvements to the YOLO detector, feature pyramid enhancement, nighttime YOLOv8 model optimization, and the fusion of the ByteTrack multi-target tracking strategy significantly enhanced the accuracy and stability of fawn target detection and tracking. Individual fawns can still be stably identified even in complex lighting environments and occluded scenes, achieving continuous trajectory acquisition and eliminating identity switching issues, providing high-quality data input for subsequent behavior recognition and health monitoring.

[0021] 3. In terms of behavior recognition, combining the spatial feature extraction capability of convolutional neural networks with the time series modeling capability of Bi-LSTM / Transformer enables the capture of fine-grained features and temporal dependencies in fawn movements, significantly improving the accuracy and robustness of behavior classification. Experiments show that its behavior recognition accuracy can reach 90%, and its health risk assessment accuracy exceeds 85%, making it more objective and reliable than traditional manual observation methods.

[0022] 4. In terms of health monitoring, this invention is the first to propose a multimodal fusion of behavioral characteristics such as the frequency and duration of fawn behavior with physiological parameters such as heart rate and body temperature, as well as environmental indicators such as temperature, humidity, and ammonia concentration, to construct a health feature vector and use a neural network to assess health risks. This method overcomes the problem of judgment bias caused by relying on a single behavioral signal and can achieve more accurate detection of health anomalies.

[0023] 5. Regarding system operating efficiency, by utilizing model pruning, quantization, and inference acceleration strategies, complex deep learning models can run smoothly on embedded edge devices, achieving a real-time processing capability of over 20 FPS for 1080p video processing. Compared to cloud-based inference, this invention significantly reduces system latency and data transmission costs, completely solving problems such as poor network conditions, high costs, and high risks of data leakage in livestock farms, and improving the overall maintainability of the system.

[0024] 6. In terms of reducing labor costs and improving management efficiency, this invention replaces a large number of manual pen patrol tasks through an automatic identification and early warning mechanism. This frees up breeders from repetitive labor and allows them to promptly grasp the health trends of fawns, improving the scientific and precise level of feeding management. The system also reduces missed and false detections, maintains consistency in monitoring records, and reduces subjective judgment differences.

[0025] 7. In terms of adaptability and promotion, the method of this invention has good versatility and can be extended to other deer species or animals with similar body size and behavioral characteristics. Meanwhile, cloud storage and multi-terminal display methods give the system good scalability, allowing for application in large-scale farming scenarios and providing a basic platform for digital animal health management.

[0026] In summary, this invention has achieved significant results in terms of real-time monitoring capabilities, identification accuracy, health assessment accuracy, system operating efficiency, reduced labor costs, scientific management, and adaptability for widespread application, and can provide a reliable intelligent health monitoring solution for fawn farming. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the architecture of the edge computing-based fawn behavior recognition and health monitoring system proposed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of the edge computing module proposed in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0029] An edge computing-based system for recognizing and monitoring the behavior and health of fawns, the system comprising: The data acquisition module is used to collect video data of the target fawns in the breeding farm enclosure, and to obtain preprocessed video after preprocessing the video data.

[0030] In this embodiment of the invention, the data acquisition module is mainly used to acquire the original video data of the target fawn in the breeding farm enclosure, and to perform various preprocessing operations on the acquired video data to improve image quality and enhance the robustness of the subsequent detection and recognition modules.

[0031] Firstly, regarding data acquisition, this invention preferably employs a combination of high-definition cameras and infrared cameras to address the impact of day / night cycles and lighting variations on monitoring. During the day, a high-definition visible light camera is used to acquire clear images of the fawns; at night, an infrared camera is activated to capture the fawns' outlines and activity information. The collaborative operation of multiple cameras provides multi-angle views, reduces blind spots, and improves overall monitoring coverage.

[0032] While acquiring data at the front end, the system performs a series of preprocessing operations on the video stream to improve the reliability of subsequent algorithms. On one hand, to address insufficient lighting and image quality fluctuations, the system scales the image to adapt its size to the input requirements of the detection model and improves the uniformity of image resolution. This invention employs a bilinear interpolation method, scaling the original image according to the image scale variation relationship. The basic principle of the bilinear interpolation method is to interpolate and sample pixels in the new image by calculating their corresponding positions in the original image, as shown in the formula:

[0033] in, For the original image, The image is scaled up. , The height and width of the original image. , The height and width of the scaled image. , These are the coordinates of the scaled image.

[0034] The above interpolation operation can scale the image proportionally to make its size conform to the model input requirements.

[0035] On the other hand, to enhance image quality under low-light conditions, this invention further employs image brightness enhancement technology to improve overall image brightness and detail resolution. The basic principle of image brightness enhancement technology is to redistribute the grayscale values ​​of the image to make them more uniform, thereby improving the visibility of the foreground area of ​​the fawn. The specific processing for grayscale enhancement is as follows:

[0036] in, The image grayscale levels (usually 256). This represents the total number of pixels in the image. The grayscale value is The number of pixels, grayscale value The new grayscale value after transformation This indicates rounding down to the nearest integer.

[0037] This formula linearly transforms the accumulated pixel grayscale histogram, resulting in a more balanced image brightness, making it suitable for image enhancement in low-light environments.

[0038] Furthermore, to further improve image clarity, this invention also performs noise suppression and contrast enhancement processing on the image to mitigate the adverse effects of sensor noise or low-light imaging. After the above enhancement processing, image details and contour information are preserved, providing more reliable input data for subsequent fawn target detection and behavior recognition algorithms.

[0039] Through the above data acquisition and preprocessing operations, the present invention can continuously output high-quality video data in complex environments such as changes in lighting, noise interference, and multi-camera collaboration, providing a solid data foundation for the accuracy and robustness of subsequent monitoring modules.

[0040] Furthermore, the edge computing-based fawn behavior recognition and health monitoring system also includes: The object detection and tracking module performs object detection and multi-object tracking on each frame of the preprocessed video, including: The detection submodule uses an improved YOLO series detector to identify the bounding box location and bounding box confidence score of the target fawn in each frame; The tracking submodule uses the ByteTrack algorithm to establish the identity association of the target fawn through two-stage matching of high-confidence and low-confidence detection boxes and Kalman filter prediction, thereby obtaining the image sequence of the target fawn.

[0041] In this embodiment of the invention, the fawn target detection and multi-target tracking module is used to perform real-time fawn individual detection on the input video stream to obtain the location region of the fawn individual in each frame of the image. This module adopts the current advanced deep learning object detection framework YOLO series algorithms, and improves the detection network according to actual needs.

[0042] Specifically, this invention uses the YOLOv12 detector as the basic model, which performs well in both detection accuracy and speed. Considering the small size and frequent movement of fawns in a farm environment, making the detection task quite challenging, this invention makes two targeted improvements to the YOLO model: First, a Coordinate Attention (CA) mechanism is introduced to enhance the model's ability to perceive salient feature regions of fawns, reducing the risk of missed detections of individual fawns in occluded environments. Second, addressing the characteristics of low image brightness and increased noise in nighttime or low-light environments, a YOLOv8 model specifically optimized for nighttime scene training is employed. By adding low-light nighttime sample data during training, the model gains stronger adaptability to low-light environments. Through these improvements, the YOLO model can accurately locate the bounding box of each fawn in complex scenes and output its corresponding confidence score.

[0043] Building upon target detection, this invention further incorporates a multi-target tracking (MOT) algorithm to transform the detected fawn locations into continuous individual trajectories. The system employs the leading-edge ByteTrack tracker to associate the targets detected by YOLO. The ByteTrack algorithm uses a "detection-detection" matching approach, leveraging high-confidence detection boxes for high-quality target matching and then reassigning low-confidence boxes to improve the overall detection rate. Compared to traditional tracking methods that only use high-confidence targets, ByteTrack fully utilizes more detection information, significantly improving the stability of fawn tracking.

[0044] In its implementation, this invention uses the YOLO model to infer the location bounding box, confidence level, and feature vector of each frame of the image. These detected features are then input into the ByteTrack tracker for a two-stage matching process: the first stage uses high-confidence bounding boxes for global matching; the second stage matches unmatched targets with low-confidence bounding boxes again to improve overall backtracking capability. Through this matching strategy, each fawn can obtain a unique and continuous identity ID and maintain stable tracking throughout the video frame sequence.

[0045] To adapt to the computing power limitations of edge devices, this invention also performs lightweight compression on the YOLO model: reducing model size by decreasing redundant parameters and accelerating inference using TensorRT. The optimized model can run efficiently on embedded hardware, making full use of limited GPU resources and ensuring the real-time responsiveness of the overall system.

[0046] Through the aforementioned detection and tracking modules, this invention can acquire the movement trajectory and spatial location of each individual in a herd of fawns over time, providing fundamental data for further behavioral identification. The continuous and stable tracking process avoids interruptions in behavioral segments due to individual loss, contributing to improved integrity and reliability of the overall monitoring.

[0047] Furthermore, the edge computing-based fawn behavior recognition and health monitoring system also includes: The behavior recognition module is used to extract the spatial features of the target fawn from the image sequence using a convolutional neural network. The spatial features are then fed into a time series modeling network in chronological order for time series analysis to identify the behavior category of the target fawn.

[0048] In this embodiment of the invention, the fawn behavior recognition and temporal analysis module is used to discriminate and sequence the behavioral states of each fawn after obtaining its location and movement trajectory. This module combines image spatial feature extraction and time series modeling techniques, improving the classification accuracy of multi-category behaviors of fawns through the combination of spatial feature extraction and time series modeling.

[0049] First, for each sequence of tracked fawn images, this invention employs a convolutional neural network (CNN) to extract static spatial observation features. A pre-trained ResNet-50, Inception-V3, or a custom lightweight CNN structure can be used to convert the fawn's image at a given moment into a spatial feature vector containing information such as its morphology and posture. Subsequently, the CNN features from one or more consecutive frames are fed sequentially into a recurrent neural network for time-series feature modeling. Preferably, a bidirectional long short-term memory network (Bi-LSTM) or a Transformer-based temporal encoder is used to capture the dynamic patterns of fawn behavior over time.

[0050] For Bi-LSTM-based methods, the model can simultaneously utilize behavioral state features from consecutive frames, improving the stability and generalization ability of time series modeling. Transformer-based methods, on the other hand, can utilize self-attention mechanisms to automatically model long-term dependencies between different frames, making them suitable for capturing categories with strong behavioral persistence. Using these models, this invention can efficiently identify various fawn behaviors, including eating, drinking, resting, and walking.

[0051] Building upon this foundation, to achieve the aggregation of behavior recognition results in the time domain, this invention further constructs a behavior scoring mechanism to comprehensively reflect the behavioral state of fawns within a certain time window. The behavior results identified from the video sequence can record the frequency and duration of each behavior within the window. This invention defines a behavior scoring mechanism. The calculation formula is as follows:

[0052] in, The behavioral scoring results are used to determine whether the behavior is "healthy". The total number of types of behavior. For the first The total duration of the behavior For window monitoring duration, For the first The importance weight of certain behaviors (e.g., "eating" is more important than "resting").

[0053] For example, a high frequency of "eating" behavior indicates that the fawn is in good health; while an excessive increase in "lying down" behavior may mean that the fawn is in poor health. Based on the statistical analysis of the identification results, the model can generate behavioral scores that reflect the stability and health trends of the fawn's behavior.

[0054] The aforementioned scoring mechanism transforms complex behavioral time series into comparable numerical indicators, providing a quantitative basis for subsequent health status assessment. To improve the accuracy of behavior recognition and temporal modeling, this invention further optimizes the model structure. For example, a cross-frame attention structure is introduced to enhance the model's ability to focus on key behavioral segments. Furthermore, this invention can also employ a PBC-Transformer model architecture to improve the model's ability to jointly model spatial and temporal patterns of behavioral categories.

[0055] By using the above-mentioned spatial feature extraction and time series feature fusion method, the present invention can distinguish various behaviors of fawns in detail and capture their changing trends, providing reliable behavioral data for subsequent health monitoring.

[0056] Furthermore, the edge computing-based fawn behavior recognition and health monitoring system also includes: The multimodal data fusion and health monitoring module is used to acquire the physiological parameters of the target fawn and the environmental data of the enclosure. It integrates the environmental data with the physiological parameters and behavioral categories of the target fawn to construct a health feature vector of the target fawn. The module calculates a health risk score based on the health feature vector through a neural network model. When the risk score exceeds a preset threshold, it outputs a health warning message.

[0057] In this embodiment of the invention, a multi-source data fusion strategy is further introduced to perform real-time scoring and risk monitoring of the health status of fawns. This module comprehensively utilizes feature information from different sources, including: time-series features such as behavior category, frequency, and duration provided by the behavior recognition module; key physiological parameters such as fawn body temperature; and environmental factors such as temperature and humidity in the enclosure. By fusing these multiple features, a fawn health status feature vector is constructed to reflect the overall health trend of the fawn in the current time period.

[0058] During system operation, the health monitoring module acquires the current physiological parameters of the fawns from various sensor devices. This data, combined with historical environmental change data from the farm, activity level variation curves, and behavioral frequency characteristics, forms a multi-dimensional health representation vector. Furthermore, this invention incorporates the individual fawn's historical health records and behavioral patterns, using long-term behavioral deviations to aid in judgment. For certain key physiological indicators, such as body temperature changes, the system employs a sliding time window method to smooth the indicators and reduce noise interference.

[0059] Furthermore, to quantify the health status of fawns, this invention constructs a health risk scoring mechanism. This mechanism, based on multimodal feature vectors, uses a neural network model or linear combination method to weighted aggregate multiple types of features to obtain a fawn health risk score. The anomaly detection approach involves teaching the model the typical ranges of various indicators under normal, healthy conditions in fawns, thereby enabling it to provide a higher risk assessment for deviations from the normal pattern. Scoring Results Normalized to a value between 0 and 1, a higher value indicates a higher health risk. The calculation method is shown in the following formula:

[0060] in, The "overall score" mechanism for calculating the health of fawns ranges from 0 to 1, with scores closer to 1 indicating a higher health risk. For the first Behavioral characteristics (e.g., eating time, exercise frequency); For the first Physiological or environmental characteristics (such as body temperature, humidity, oxygen concentration); , The weight of each feature (indicating how much this indicator affects health risk); This is used as a bias (to adjust the overall result); The sigmoid activation function compresses the results to the 0-1 range.

[0061] By introducing this scoring mechanism, the system can automatically adjust the weights based on the joint changes of multiple features, so that... It is more closely aligned with the actual health conditions of fawns. When a fawn exhibits abnormal behavior or physiological indicators, such as a continuous decrease in activity level or a significant increase in body temperature, its health risk score will increase accordingly, and the system can issue a health warning, enabling real-time monitoring of the fawn's health status.

[0062] In summary, the multimodal data fusion and health monitoring module of this invention can comprehensively utilize the multidimensional characteristics of fawn behavior, physiology, and environment to provide a reliable model foundation and decision-making basis for health monitoring.

[0063] Furthermore, the edge computing-based fawn behavior recognition and health monitoring system also includes: The edge computing module is used to run data acquisition modules, target detection and tracking modules, behavior recognition modules, multimodal data fusion and health monitoring modules on embedded computing devices deployed in the farm.

[0064] In this embodiment of the invention, the advantages of edge computing architecture are fully utilized, and the aforementioned functional modules are integrated and deployed on edge devices at the farm site to achieve real-time, efficient, and secure operation. Compared to the traditional solution of uploading video streams to the cloud for processing, edge computing has significant advantages. First, in the field environment, the amount of video data is large while the transmission bandwidth is limited. Uploading all the original video to the cloud is not only costly and has high latency, but it is also prone to data loss due to network instability. However, the analysis and processing are completed directly on-site at the edge, and the overall latency of all behavior recognition and health assessment can be controlled within 200 milliseconds. Therefore, when fawns exhibit dangerous behavior, the system can issue an alarm almost instantly, giving farm managers valuable time to respond.

[0065] Secondly, edge deployment avoids large-scale transmission of video data, thus strengthening data privacy protection. Since the videos contain sensitive information such as the growth status and activity patterns of fawns, uploading them all to the cloud would not only require significant bandwidth and storage resources but also pose a risk of privacy breaches. Local processing at the edge avoids these problems, effectively reducing security risks associated with data transmission.

[0066] Third, edge deployment can also reduce the impact of network failures. Because the original video is stored locally, the system can continue to process and analyze data even if the external network connection is unstable or interrupted, ensuring stable system operation.

[0067] In this specific implementation, the edge computing module performs lightweight processing on the model. Specifically, through optimization methods such as model pruning and parameter quantization, the object detection model, behavior recognition model, and data fusion model achieve high efficiency when running on edge devices. This invention converts the inference framework into a TensorRT engine suitable for edge devices, enabling the model to fully utilize hardware acceleration resources during inference and further reducing computational latency.

[0068] In terms of hardware configuration, this invention employs edge computing devices with high computing power. For example, an embedded high-performance computing platform such as NVIDIA Xavier NX can be used, achieving a frame rate exceeding 30 FPS when processing 1080p video, meeting the real-time requirements of farms. It is important to emphasize that this invention has specifically refined and optimized the model structure and inference process to adapt to edge hardware, enabling large-scale models to run stably on embedded devices.

[0069] Furthermore, this invention incorporates contingency plans for system anomalies. For instance, when excessively high device temperatures may lead to decreased computing performance, this invention uses a system monitoring module to record the operational status of edge devices in real time. When the hardware temperature exceeds a preset threshold, the system can promptly reduce the model's workload or activate a backup mode, thereby ensuring overall system stability. In cases of data surges or temporary increases in data acquisition frequency, the edge device can also automatically cache short-term data to prevent data loss.

[0070] Furthermore, to achieve collaboration between the edge and cloud systems, this invention designs an edge-cloud collaborative architecture. When users require long-term data storage or model update capabilities, the edge device can periodically upload health assessment results to the cloud platform for historical analysis, model updates, and remote management. Through edge computing and cloud collaboration, this invention constructs an efficient, reliable, and sustainably scalable intelligent monitoring system, enabling the system to operate independently and stably on the farm and providing continuous monitoring services for fawn behavior and health over the long term.

[0071] In another preferred embodiment of the present invention, the method for identifying and monitoring the behavior and health of fawns based on edge computing includes: Video data of the target fawns in the farm enclosure was collected, and the video data was preprocessed to obtain preprocessed video. Perform object detection and multi-object tracking on each frame of the preprocessed video, including: An improved YOLO series detector was used to identify the bounding box location and bounding box confidence score of the target fawn in each frame; The ByteTrack algorithm is used to establish the identity association of the target fawn through two-stage matching of high-confidence detection boxes and low-confidence detection boxes and Kalman filter prediction, thereby obtaining the image sequence of the target fawn; Based on the image sequence, the spatial features of the target fawn are extracted using a convolutional neural network. The spatial features are then fed into a time series modeling network in chronological order for time series analysis to identify the behavioral categories of the target fawn. The physiological parameters of the target fawn and the environmental data of the enclosure are obtained. The environmental data and the physiological parameters and behavioral categories of the target fawn are fused to construct the health feature vector of the target fawn. A health risk score is calculated based on the health feature vector through a neural network model. When the risk score exceeds a preset threshold, a health warning message is output. The embedded computing devices deployed in the farm run data acquisition modules, target detection and tracking modules, behavior recognition modules, and multimodal data fusion and health monitoring modules.

[0072] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fawn behavior recognition and health monitoring system based on edge computing, characterized in that, The system includes: The data acquisition module is used to collect video data of the target fawns in the breeding farm enclosure, and to obtain preprocessed video after preprocessing the video data; The object detection and tracking module performs object detection and multi-object tracking on each frame of the preprocessed video, including: The detection submodule uses an improved YOLO series detector to identify the bounding box location and bounding box confidence score of the target fawn in each frame; The tracking submodule uses the ByteTrack algorithm to establish the identity association of the target fawn through two-stage matching of high-confidence detection boxes and low-confidence detection boxes and Kalman filter prediction, thereby obtaining the image sequence of the target fawn. The behavior recognition module is used to extract the spatial features of the target fawn based on the image sequence using a convolutional neural network, and then feed the spatial features into a time series modeling network for time series analysis in chronological order to identify the behavior category of the target fawn. The multimodal data fusion and health monitoring module is used to acquire the physiological parameters of the target fawn and the environmental data of the enclosure. It integrates the environmental data with the physiological parameters and behavioral categories of the target fawn to construct a health feature vector of the target fawn. The module calculates a health risk score based on the health feature vector through a neural network model. When the risk score exceeds a preset threshold, it outputs a health warning message. The edge computing module is used to run data acquisition modules, target detection and tracking modules, behavior recognition modules, multimodal data fusion and health monitoring modules on embedded computing devices deployed in the farm.

2. The edge computing based fawn behavior recognition and health monitoring system according to claim 1, wherein, The data acquisition module specifically includes: The video data acquisition unit is used to collaboratively collect multi-channel video data of the target fawn through multiple high-definition visible light cameras and infrared cameras deployed in the breeding farm pens; The preprocessing unit is used to perform image enhancement, noise reduction, illumination equalization, and background modeling on the video data from each stream to obtain the preprocessed video. 3.The edge computing based fawn behavior recognition and health monitoring system of claim 1, wherein, The improved YOLO series detectors introduce a coordinate attention mechanism and employ the YOLOv8 model suitable for low-light environments.

4. The edge computing based fawn behavior recognition and health monitoring system of claim 1, wherein, The time series modeling network includes a bidirectional long short-term memory network and a Transformer model.

5. The edge computing-based fawn behavior recognition and health monitoring system according to claim 4, characterized in that, The target fawn's behavioral categories include standing, walking, eating, drinking, lying down, socializing, and being sick.

6. The edge computing-based fawn behavior recognition and health monitoring system according to claim 5, characterized in that, The health risk score is calculated using the Sigmoid function for normalization, and the health risk score ranges from 0 to 1.

7. The edge computing-based fawn behavior recognition and health monitoring system according to claim 1, characterized in that, The edge computing module specifically includes: Embedded computing devices are used to deploy data acquisition modules, target detection and tracking modules, behavior recognition modules, multimodal data fusion and health monitoring modules, and to run the inference computing of the above modules locally on the farm. The model lightweighting unit is used to perform model pruning, model quantization, and inference acceleration optimization on the models in the target detection and tracking module, behavior recognition module, and multimodal data fusion and health monitoring module. The real-time processing unit is used to perform local real-time processing on video data, physiological parameters of the target fawns, and environmental data of the enclosure, and generate monitoring results. The communication unit is used to transmit the monitoring results and health warning information to the user terminal for display via a local area network structure.

8. The edge computing-based fawn behavior recognition and health monitoring system according to claim 7, characterized in that, The user terminal includes a computer monitoring terminal and a mobile terminal.

9. A method for recognizing the behavior and monitoring the health of fawns based on edge computing, characterized in that, The method includes: Video data of the target fawns in the farm enclosure was collected, and the video data was preprocessed to obtain preprocessed video. Perform object detection and multi-object tracking on each frame of the preprocessed video, including: An improved YOLO series detector was used to identify the bounding box location and bounding box confidence score of the target fawn in each frame; The ByteTrack algorithm is used to establish the identity association of the target fawn through two-stage matching of high-confidence detection boxes and low-confidence detection boxes and Kalman filter prediction, thereby obtaining the image sequence of the target fawn; Based on the image sequence, the spatial features of the target fawn are extracted using a convolutional neural network. The spatial features are then fed into a time series modeling network in chronological order for time series analysis to identify the behavioral categories of the target fawn. The physiological parameters of the target fawn and the environmental data of the enclosure are obtained. The environmental data and the physiological parameters and behavioral categories of the target fawn are fused to construct the health feature vector of the target fawn. A health risk score is calculated based on the health feature vector through a neural network model. When the risk score exceeds a preset threshold, a health warning message is output. The embedded computing devices deployed in the farm run data acquisition modules, target detection and tracking modules, behavior recognition modules, and multimodal data fusion and health monitoring modules.

Images