Multi-sensor based pet smart monitoring method and an air box

Abstract

The application relates to the technical field of pet intelligent monitoring, in particular to a pet intelligent monitoring method based on multiple sensors and an aviation box, aiming to solve the pain points of single monitoring dimension, inaccurate state identification of existing monitoring technology, and no intelligent monitoring and active intervention capability of traditional aviation boxes. The method relies on a distributed multiple sensor network to synchronously collect three types of real-time data of a video stream, environmental sensing and sound in a pet service area; extracts a pet skeleton key point sequence and generates real-time posture features, synchronously obtains temperature and humidity, air quality and pet call frequency and decibel values, aligns the time axis of multidimensional data to generate time series data; inputs a behavior discriminator to obtain three types of state labels of normal, anxious and listless, and generates corresponding service scheduling instructions accordingly. The aviation box is integrated with corresponding sensors and intelligent modules to realize hardware and software integration, accurately complete pet state monitoring and abnormal intervention, and improve the intelligent and refined level of monitoring.

Description

Technical Field

[0001] This application relates to the field of intelligent pet monitoring technology, and in particular to a multi-sensor-based intelligent pet monitoring method and an aviation crate. Background Technology

[0002] The pet consumer market is continuously upgrading. Pet air transport, professional boarding, short-term unattended home care, and short-distance pet travel have become core needs in daily pet care. The pet monitoring field is gradually shifting from traditional manual care to intelligent and automated methods. Monitoring and ensuring the well-being of pets in enclosed and unfamiliar environments has become a key focus of industry research and development. Pet carriers, as the core carrier for monitoring pets in enclosed environments, are evolving from simple pet containment and portable transport to integrated and intelligent solutions. The industry as a whole is placing higher demands on the accuracy, responsiveness, and hardware compatibility of pet monitoring.

[0003] Existing intelligent pet monitoring technologies and traditional pet carriers suffer from two core technical deficiencies. First, existing monitoring systems rely on a single sensor to collect data, either using only a camera for video monitoring or monitoring only a single environmental parameter. They fail to integrate and analyze the three core data points of pet posture, sound, and environment, making it impossible to accurately identify abnormal pet conditions. This results in a high rate of false positives and false negatives, hindering refined and reliable monitoring. Second, traditional pet carriers only provide basic containment and transport functions, lacking integrated multi-sensor intelligent monitoring modules and corresponding algorithms. They cannot monitor the pet's condition and the carrier environment in real time, nor can they provide proactive early warnings and targeted interventions for abnormalities. Pets are prone to stress reactions in enclosed spaces, and their discomfort cannot be detected and addressed in a timely manner. Summary of the Invention

[0004] To address the industry pain points of existing pet monitoring technologies, such as limited monitoring dimensions and insufficient accuracy in status assessment, as well as the lack of intelligent monitoring and proactive intervention capabilities in traditional pet carriers, this application provides a multi-sensor-based intelligent pet monitoring method and a pet carrier.

[0005] Firstly, this application provides a multi-sensor-based intelligent pet monitoring method, which adopts the following technical solution: A multi-sensor-based intelligent pet monitoring method includes: Acquire real-time monitoring data collected by a distributed multi-sensor network deployed in the pet service area, including video stream data, environmental sensor data, and sound acquisition data; Based on the video stream data, extract the skeletal key point sequence of the pet; Based on the skeletal key point sequence, the real-time posture features of the pet are generated; Based on the environmental sensing data, acquire temperature and humidity data and air quality data of the service area; Based on the sound data collected, the frequency and decibel value of the pet's barks are identified; The real-time posture features, temperature and humidity data, air quality data, vocal frequency, and decibel value are aligned along the time axis to generate multi-dimensional monitoring time-series data. The multidimensional monitoring time series data is input into the behavior discriminator to obtain the pet's status tags, which include normal state, anxious state or lethargic state. Based on the status label, a corresponding service scheduling instruction is generated.

[0006] Optionally, extracting the skeletal keypoint sequence of the pet based on the video stream data includes: Based on the video stream data, the location of the detection box for the pet is obtained using an object detection algorithm; Based on the location of the detection box, the pet area image is cropped; Instance segmentation is performed on the pet region image to obtain the pet's outline mask; Based on the contour mask and the pet region image, the skeletal topology of the pet is extracted; Based on the skeletal topology, a predetermined number of key joints are located, including the tip of the nose, the eye point, the base of the ear, the shoulder joint, the hip joint, and the base of the tail. The key joints are tracked and matched in multiple consecutive frames of images to generate the skeletal key point sequence.

[0007] Optionally, generating the pet's real-time pose features based on the skeletal keypoint sequence includes: Based on the coordinates of adjacent key points in the skeletal key point sequence, calculate the pet's limb angle parameters; Based on the displacement of specific key points in the skeletal key point sequence, calculate the pet's motion amplitude parameters; Determine the pet's movement status, which includes static state, small-scale activity, and large-scale running and jumping; If the movement state is static, then the prone or standing posture is determined based on the limb angle parameters. If the movement state is a small-range activity, then obtain the head swing frequency and tail swing amplitude; If the movement state is a large-scale running and jumping, then obtain the running and jumping duration and the running and jumping trajectory; The limb angle parameters, the motion amplitude parameters, the motion state, and the corresponding posture determination results are fused to generate the real-time posture features.

[0008] Optionally, inputting the multidimensional monitoring time-series data into the behavior discriminator to obtain the pet's status tags includes: Based on the real-time posture features, the frequency of abnormal postures of the pet is obtained; Based on the frequency of the call and the decibel value, obtain the abnormal call segment; Determine whether the frequency of the abnormal vocalization overlaps with the frequency of the abnormal posture in time; If there is overlap, extract the temperature and humidity data and the air quality data corresponding to the overlapping time period; Based on the temperature and humidity data and air quality data corresponding to the overlapping time periods, determine whether there are environmental stress triggering factors. If there are environmental stress triggers, the state label will be marked as an environmentally unwell state. If no environmental stress triggers are present, the state label will be marked as a state of physiological discomfort.

[0009] Optionally, after marking the state label as an unwell state if environmental stress triggers exist, the method further includes: Obtain body part information associated with the frequency of the abnormal posture; Based on the body part information, posture data of the same part in historical monitoring data are retrieved and compared; Based on the comparison results, it is determined whether the body parts exhibit persistent postural deviation. If persistent posture deviation exists, a medical examination suggestion instruction is generated, and the body part information is marked as a medical attention area. The medical examination suggestion instruction and the medical attention area are sent to the service terminal.

[0010] Optional, also includes: Obtain the environmental parameter type corresponding to the environmental stress triggering factor; Based on the environmental parameter types, generate environmental control instructions; Determine whether the status label changes from an unsuitable environmental state to a normal state within a preset time after the environmental control command is executed; If the pet does not return to a normal state within a preset time, the real-time posture characteristics and vocal frequency of the pet will be reacquired. The pet's status tag is updated based on the reacquired real-time posture features and vocal frequency.

[0011] Optionally, the service scheduling instruction includes a first scheduling instruction and a second scheduling instruction; generating the corresponding service scheduling instruction based on the status label includes: Determine whether the status label is in an abnormal state; If the status label is an abnormal status, then the abnormal start time point is extracted from the multidimensional monitoring time series data; Based on the anomaly start time point, backtrack to obtain the first environmental dataset before the anomaly start time point and the second environmental dataset after the anomaly start time point; The first environmental dataset is compared with the second environmental dataset to identify inflection points in environmental change; If the environmental change inflection point exists, a first type of scheduling instruction is generated based on the environmental parameters corresponding to the environmental change inflection point. The first type of scheduling instruction is used to automatically adjust the environmental equipment in the service area. If the environmental change inflection point does not exist, a second type of scheduling instruction is generated based on the abnormality type corresponding to the abnormal state. The second type of scheduling instruction is used to trigger manual intervention by service personnel.

[0012] Optionally, the second type of scheduling instructions includes reassurance and guidance instructions and medical intervention instructions; the generation of the second type of scheduling instructions based on the abnormality type corresponding to the abnormal state includes: Obtain the real-time posture features and vocal frequency corresponding to the abnormal state; Based on the real-time posture characteristics and the vocalization frequency, an abnormal behavior combination pattern is determined; Based on the abnormal behavior combination pattern, a preset abnormal cause library is matched to obtain a list of candidate abnormal causes; Obtain the pet's service history, which includes historical abnormal events and corresponding historical solutions; The candidate anomaly cause list is correlated with the service history record to filter out the target anomaly cause with the highest matching degree with the current anomaly behavior combination pattern; Based on the cause of the target anomaly, a targeted reassurance and guidance instruction or medical intervention instruction is generated and pushed to the mobile terminal of the target service personnel.

[0013] Optionally, after generating the corresponding service scheduling instruction, the method further includes: Obtain the pet's identification and a complete record of this monitoring; Link the complete record to the pet's electronic file; Based on the complete records, analyze the trend of pet's status changes during multiple service processes; Based on the aforementioned status change trend, personalized pet service recommendations are generated, including preferred service time periods and recommended service types.

[0014] Secondly, this application also discloses a flight case, which adopts the following technical solution: A pet intelligent monitoring flight case based on multiple sensors includes a main body, which is formed by a top, bottom, left side wall, right side wall, front wall and rear wall to form a closed cavity for accommodating a pet. The main body also integrates a case door, a detachable lick-type water feeder, a rotating feeding pull-out box and a visual intelligent AI module. The cabinet door is located on the front wall of the main body of the cabinet and is connected to the cabinet by a hinge for opening and closing. The detachable lick-type water feeder is installed on the right side wall of the main body of the enclosure. It includes a water storage chamber, a lick-type water outlet, and a snap-on mounting base. The water storage chamber is connected to the water outlet, and the entire water feeder can be quickly disassembled through snap-on fasteners. It is used to provide drinking water for pets and facilitates cleaning and hydration. The rotating feeding pull-out box includes a pull-out box door and a feeding bowl. The feeding bowl is fixedly connected to the inner wall of the pull-out box door, and the pull-out box door is hinged to the rear wall of the main body of the box for opening and closing. The visualized intelligent AI module is installed on the top of the main body of the enclosure and embedded inside the cavity of the enclosure from the top. It includes a distributed multi-sensor network and a control unit. The distributed multi-sensor network includes at least a camera, a temperature and humidity sensor, an air quality sensor, and a sound collector. The control unit is electrically connected to the camera, the temperature and humidity sensor, the air quality sensor, and the sound collector, respectively.

[0015] In summary, this application includes the following beneficial technical effects: By employing a distributed multi-sensor network to simultaneously collect three types of data—video streams, environmental sensors, and sound—this technology effectively addresses the issue of limited monitoring dimensions in existing pet monitoring technologies. By aligning skeletal key points, posture features, temperature and humidity, air quality, and vocal characteristics along a timeline and inputting this data into a behavior discriminator, it achieves accurate identification of pets' normal, anxious, and lethargic states, overcoming the shortcomings of traditional methods such as inaccurate state recognition and susceptibility to misjudgments and omissions. Simultaneously, it automatically generates corresponding service scheduling instructions based on the identified state tags, enabling timely responses and interventions to abnormal pet states. This solves the problem of traditional monitoring methods lacking proactive intervention capabilities, thus achieving more precise, intelligent, and reliable comprehensive monitoring of pets. Attached Figure Description

[0016] Figure 1 This is a main flowchart of a multi-sensor-based intelligent pet monitoring method according to an embodiment of this application; Figure 2 This application describes an embodiment of a flight case. Figure 1 ; Figure 3 This application describes an embodiment of a flight case. Figure 2 .

[0017] Explanation of reference numerals in the attached figures: 1. Main body of the feeding box; 2. Box door; 3. Detachable lick-type water feeder; 4. Rotating pull-out feeding box; 5. Visual intelligent AI module. Detailed Implementation

[0018] In the first aspect, this application discloses a pet intelligent monitoring method based on multiple sensors.

[0019] Reference Figure 1 A multi-sensor-based intelligent pet monitoring method includes steps S101 to S108: Step S101: Obtain real-time monitoring data collected by the distributed multi-sensor network deployed in the pet service area. The real-time monitoring data includes video stream data, environmental sensor data, and sound acquisition data.

[0020] Specifically, the pet service area refers to the enclosed, independent cavity inside the smart aviation crate. This cavity provides pets with their own activity space, isolating them from external environmental interference and adapting to the monitoring needs of enclosed scenarios such as pet transport, boarding, and short-distance travel. The distributed multi-sensor network adopts a balanced deployment method with multiple points. The core sensors are embedded in the center of the top of the crate, the upper middle section of the left and right side walls, and the inner side of the rear wall, forming a 360-degree all-around data collection network without blind spots, completely eliminating monitoring blind spots and avoiding data loss caused by pets curling up or lying in corners. The video stream data uses a standard video capture rate of 25 frames per second. This rate balances real-time performance with the computational load of embedded hardware, ensuring complete capture of the pet's continuous limb movements without frame loss, while also preventing data processing overload due to excessively high frame rates. Environmental sensor data is set to be collected once per second, a high-frequency, short-cycle acquisition mode that can capture subtle fluctuations in environmental parameters in real time and promptly detect sudden changes in temperature, humidity, and air quality. Sound data is acquired in a real-time, all-day pickup mode, coupled with an adaptive sensitivity adjustment function. All three types of data are collected, transmitted, and processed simultaneously, strictly ensuring the time consistency and timeliness of multi-dimensional data, laying the foundation for subsequent timing alignment.

[0021] Step S102: Extract the skeletal key point sequence of the pet based on the video stream data.

[0022] Specifically, the target detection algorithm uses the YOLOv5-tiny lightweight model. This model is a lightweight deep learning detection algorithm optimized for the limited computing power and memory of embedded terminal hardware. The processing latency is strictly controlled within 100ms, meeting the core low-latency requirement of real-time monitoring. It can quickly and accurately locate the rectangular detection box coordinates of the pet in the video frame, completely selecting the pet's entire body and effectively eliminating background interference factors such as the box structure, feeding utensils, and the edges of the pet's fur. After completing the detection box localization, the video frame is precisely cropped according to the selected coordinate parameters, retaining only the local area image of the pet, significantly reducing the amount of data for subsequent image processing and reducing the computational pressure on the algorithm. Subsequently, a lightweight improved version of the MaskR-CNN instance segmentation model is used to perform fine segmentation on the cropped pet area image, accurately distinguishing the pet's body contour from the background environment, generating a high-resolution binary contour mask, and completely eliminating irrelevant pixel interference. Based on contour masking, the topological structure of the pet's linear skeleton is further extracted, clearly showing the core connection relationships of the head, torso, limbs, and tail. Finally, six types of core key joints are located: nose tip, eye point, ear root, shoulder joint, hip joint, and tail root. These points are the core joints most sensitive to changes in the pet's posture, which can not only completely restore the dynamic posture of the pet's whole body, but also avoid computational redundancy caused by too many key points. Then, the Kalman filter algorithm is used to stably track and match the same key point in consecutive video frames, eliminating the problem of key point jumps caused by image shaking and slight displacement of the pet, generating a continuous and accurate sequence of skeletal key points.

[0023] Step S103: Generate the real-time pose features of the pet based on the skeletal keypoint sequence.

[0024] Specifically, the real-time posture feature generation core relies on the coordinate data of skeletal key points. Through the vector angle calculation formula, it accurately calculates the limb angle parameters between adjacent key points, such as the angle between the shoulder joint and the forelimb, the hip joint and the hindlimb, and the trunk and neck. It pre-sets the angle threshold range of the pet's normal posture. If the angle exceeds this range, it is judged as an abnormal posture such as limb tension or curling. At the same time, it selects several movement-sensitive points such as the tip of the nose, the base of the tail, and the ends of the limbs as specific key points. It calculates the pixel displacement distance within a single frame or three consecutive frames, converts it into a quantitative motion amplitude parameter, and intuitively reflects the intensity of the pet's movement. Then, it divides the movement state into three levels according to the displacement threshold: key point displacement of less than 5 pixels is judged as static, 5-20 pixels is judged as small-scale activity, and more than 20 pixels is judged as large-scale running and jumping. Feature parameters are extracted in layers for different movement states. In the static state, the angle between the torso and the supporting surface is used to distinguish between lying down, standing, or semi-lying down postures. In the small-range activity state, the focus is on the frequency of head swaying and the amplitude of tail swaying, which are core indicators for judging the pet's anxiety, curiosity, and unease. In the large-range running and jumping state, the focus is on monitoring the duration of running and jumping and the regularity of movement trajectory to distinguish between normal play and excessive agitation. Finally, all posture-related parameters are normalized in the 0-1 range and integrated into a one-dimensional continuous feature vector to form a comprehensive, quantifiable, real-time posture feature that can be directly input into the algorithm model, fully reflecting the pet's current body posture, movement amplitude, and behavioral state.

[0025] Step S104: Based on environmental sensor data, acquire temperature and humidity data and air quality data of the service area.

[0026] Specifically, environmental sensing data is collected by high-precision integrated sensors. The temperature and humidity sensors are set to a range of 0-40℃ and relative humidity of 30%-90%RH, covering the entire environmental range in which pets live daily. The collected data undergoes hardware calibration and error correction, and is converted into standard numerical signals to ensure accurate and reliable values. The air quality sensor focuses on core pollution indicators in the pet's enclosed space, with a focus on collecting two key parameters: ammonia concentration and carbon dioxide concentration. Ammonia is a harmful gas produced by the volatilization of pet excrement, while carbon dioxide concentration reflects the air circulation quality in the enclosed space. These two are core environmental indicators for judging environmental comfort and inducing stress in pets. The sensors are periodically zero-point calibrated to avoid data drift caused by long-term use. The collected data is transmitted to the control unit in real time as the core basis for judging environmental stress triggering factors.

[0027] Step S105: Based on the sound acquisition data, identify the frequency and decibel value of the pet's barking.

[0028] Specifically, sound data is acquired in real time using a high-sensitivity microphone. The raw audio signal is first analyzed in the frequency domain using a Fast Fourier Transform (FFT) algorithm to accurately extract the frequency values ​​of pet vocalizations, distinguishing between different vocalization types such as whimpering, barking, growling, and purring. The decibel measurement range is set to 0-120dB, covering the full volume range of pet vocalizations. An environmental noise filtering algorithm is also incorporated to filter out all low-frequency environmental noise below 50dB (such as the sound of the enclosure rubbing and faint external noise), retaining only the valid vocal signals from the pet and avoiding interference from environmental noise in the identification of abnormal vocalizations. Normal pet vocalization frequencies and decibel thresholds are preset in advance; vocalizations exceeding the threshold range are identified as abnormal, providing reliable audio evidence for subsequent emotion and state assessment.

[0029] Step S106: Align the real-time attitude characteristics, temperature and humidity data, air quality data, call frequency and decibel value with the time axis to generate multi-dimensional monitoring time series data.

[0030] Specifically, timeline alignment uses a unified, millisecond-level high-precision timestamp as the benchmark standard to synchronously match and correlate video stream posture data, temperature and humidity data, air quality data, and sound data. This strictly ensures that the posture, environment, and sound data corresponding to the same millisecond moment correspond one-to-one, eliminating the problem of time sequence misalignment of data from different dimensions. After alignment, the data is organized into a continuous multi-dimensional monitoring time series array in chronological order. Each set of time series data contains complete posture features, environmental parameters, and audio features at that moment, forming a comprehensive, multi-dimensional, and time-coherent complete dataset. This completely solves the drawbacks of isolated analysis of single-dimensional data, providing comprehensive, accurate, and time-consistent data support for subsequent behavior discriminators and avoiding misjudgments of state due to time sequence deviations.

[0031] Step S107: Input the multidimensional monitoring time series data into the behavior discriminator to obtain the pet's status tags, which include normal state, anxious state, or lethargic state.

[0032] Specifically, the behavior discriminator employs a lightweight CNN-LSTM hybrid deep learning model. The CNN convolutional neural network extracts spatial features from multidimensional data, while the LSTM long short-term memory network mines temporal variation features. This combination perfectly meets the analytical needs of time-series data. The model was trained and optimized using 5000 sets of pet samples covering different breeds, ages, and states, including normal states, agitated states, lethargic states, environmental discomfort, and physiological discomfort. After repeated iterative training and validation, the state recognition accuracy reached over 95%. The model can achieve real-time classification output of pet states, and further subdivides the basic state labels into two major subcategories: environmental discomfort and physiological discomfort. This clarifies the core causes of abnormal states, significantly improving the targeting and accuracy of state discrimination, and providing a clear basis for subsequent differentiated service scheduling.

[0033] Step S108: Generate the corresponding service scheduling instruction based on the status label.

[0034] Specifically, the service dispatch instructions adopt a tiered handling mechanism to achieve precise and differentiated intervention: when the pet is in a normal state, the system continuously monitors but does not output any dispatch instructions, only synchronizing real-time data to the terminal to maintain the current environment and operating mode; when the system determines that the environment is uncomfortable, it automatically generates corresponding environmental parameter adjustment instructions, triggering the operation of the built-in ventilation, cooling, humidification, and dehumidification equipment in the pet carrier, without the need for manual intervention, to quickly improve the internal environment of the carrier; when the system determines that the pet is in a state of physiological discomfort or severe anxiety, it immediately generates manual intervention instructions or medical examination suggestions, which are pushed to service terminals such as the pet owner's mobile APP and the boarding center's management backend in real time, while marking the abnormality type, core manifestations, and areas of concern, reminding staff to handle the situation in a timely manner, realizing an intelligent transformation from passive monitoring to proactive intervention, and preventing the pet's discomfort symptoms from worsening.

[0035] In one embodiment of this example, step S102, which extracts the skeletal keypoint sequence of the pet based on video stream data, includes steps S201 to S206: Step S201: Based on the video stream data, obtain the detection box position of the pet using an object detection algorithm.

[0036] Specifically, the lightweight YOLOv5-tiny object detection algorithm is optimized for embedded terminal hardware, abandoning complex network structures and compressing model size. It balances detection accuracy and processing speed, with overall processing latency strictly controlled within 100ms, meeting the low-latency requirements of real-time monitoring. During algorithm execution, it quickly scans and analyzes the input video frames, automatically identifies the pet's body outline, and outputs the coordinate parameters of the rectangular detection box of the pet in the image. These coordinates accurately encompass the entire pet, effectively isolating background interference factors such as the cage structure, feeding utensils, and shadows on the inner wall. This allows for locking the core area for subsequent image cropping and key point extraction, improving the efficiency and accuracy of subsequent processing.

[0037] Step S202: Based on the location of the detection box, crop out the pet area image.

[0038] Specifically, image cropping is performed strictly according to the coordinates of the rectangular detection box output by the target detection. A precise pixel-level cropping method is used to remove all background areas from the complete video frame, retaining only the local core image where the pet is located. After cropping, the image resolution is significantly reduced, and the amount of data processing is reduced by more than 60%, effectively alleviating the computational pressure on the embedded hardware. At the same time, it reduces the interference of background pixels on subsequent instance segmentation and skeleton extraction, greatly improving the image processing speed and key point positioning accuracy, and achieving efficient and accurate local image processing.

[0039] Step S203: Perform instance segmentation on the pet region image to obtain the pet's outline mask.

[0040] Specifically, instance segmentation employs a lightweight Mask R-CNN model, a pixel-level image segmentation technique that accurately distinguishes the pet's body, fur edges, and background environment. Each pixel is classified and labeled, ultimately generating a binary contour mask. In the binary mask, the pet's body area is marked as a white valid pixel, while the background area is marked as a black invalid pixel. This completely eliminates interference factors such as blurred fur edges and overlapping background clutter, locking in a clean pet body contour. This provides a clear image foundation for subsequent skeleton topology extraction and avoids key point localization errors caused by contour blur.

[0041] Step S204: Extract the skeletal topology of the pet based on the contour mask and the pet region image.

[0042] Specifically, the skeleton topology extraction relies on morphological image processing algorithms to refine, denoise, and extract the skeleton from the contour mask, simplifying the three-dimensional pet body into a single-pixel-width linear skeleton. This clearly presents the core connection relationships and limb directions of the head, neck, torso, limbs, and tail, restoring the core skeletal structure of the pet's body. This skeleton discards redundant information such as fur and volume, retaining only the core features of limb connections and joint positions. It serves as the structural basis for subsequent key joint point localization, ensuring that the key point localization matches the actual skeletal joint positions of the pet.

[0043] Step S205: Based on the skeletal topology, locate a preset number of key joints, including the tip of the nose, the eye point, the base of the ear, the shoulder joint, the hip joint, and the base of the tail.

[0044] Specifically, the selection of key joints follows the principle of "core coverage and computing power adaptation," selecting six types of core points: nose tip, eye spot, ear base, shoulder joint, hip joint, and tail base. These points correspond to the core joints of the pet's head, torso, and limbs, respectively, and are the most sensitive parts to reflect changes in limb posture and movement, covering the core posture representation area of ​​the whole body. The number of six types of points is moderate, which can not only fully capture the dynamic changes of the pet's whole body posture and accurately reproduce various postures such as lying down, standing, curling up, and restlessness, but also avoid excessive computing power due to too many key points, perfectly adapting to the computing power limitations of the embedded hardware in the flight case, and achieving a dual balance between posture recognition completeness and computing efficiency.

[0045] Step S206: Track and match key joints in multiple consecutive frames of images to generate a sequence of skeletal key points.

[0046] Specifically, through a prediction-update mechanism, the coordinates of key points in each frame are accurately predicted and corrected to eliminate key point jumps and loss caused by video image jitter, slight pet displacement, and changes in lighting. This ensures that the coordinates of the same key joint are continuous and the trajectory is stable in a continuous time sequence. After tracking and matching are completed, the temporal coordinates of each key point are organized in chronological order to form a continuous, stable, and error-free sequence of skeletal key points, ensuring the temporal coherence and data reliability of posture feature extraction.

[0047] In one embodiment of this example, step S103, which generates the pet's real-time pose features based on the skeletal keypoint sequence, includes steps S301 to S307: Step S301: Calculate the pet's limb angle parameters based on the coordinates of adjacent key points in the skeletal key point sequence.

[0048] Specifically, the limb angle parameters are calculated using a vector angle formula. Two adjacent key joints and the center point of the torso are selected to form a spatial vector, which is then substituted into the formula to accurately calculate the limb angle value. Normal posture angle threshold ranges are preset for different breeds of pets in advance. For example, the limb angle is smaller when the pet is relaxed and lying down, and larger when it is standing and stretching. When the calculated angle value exceeds the preset normal range, it is immediately judged as an abnormal posture such as limb tension, stiffness, or curling up. This achieves a quantitative judgment of whether the posture is normal or not, abandoning the fuzzy visual judgment and relying on numerical indicators to complete the initial posture screening throughout the process.

[0049] Step S302: Calculate the pet's motion amplitude parameters based on the displacement of specific key points in the skeletal key point sequence.

[0050] Specifically, the motion amplitude parameter adopts the pixel displacement quantization method. It selects specific key points with large motion amplitude and high sensitivity, such as the tip of the nose, the base of the tail, and the ends of the limbs, and calculates the pixel movement distance within a single frame or within three consecutive frames. Then, it is converted into a motion amplitude score of 0-100 points through a preset conversion formula. The higher the score, the more intense the pet's movement, and the lower the score, the more gentle the movement. This quantization method transforms the abstract motion amplitude into an intuitive value, which makes it easier for the algorithm to quickly determine the intensity of the pet's movement, accurately distinguish between normal activity and abnormal agitation, and avoid subjective judgment bias.

[0051] Step S303: Determine the pet's movement status, which includes static, small-scale activity, and large-scale running and jumping.

[0052] Specifically, the motion state classification relies on preset pixel displacement thresholds to achieve three levels of accurate classification: continuous displacement of key points less than 5 pixels is judged as static, corresponding to states such as lying down, sitting still, or sleeping without obvious limb movements; displacement between 5 and 20 pixels is judged as small-scale activity, corresponding to small local movements such as head turning, tail swaying, and slight ear twitching; displacement greater than 20 pixels is judged as large-scale running and jumping, corresponding to large-scale full-body movements such as standing, walking, jumping, circling, and digging. This threshold has been verified by a large number of pet behavior samples, adapts to the movement characteristics of small and medium-sized pets, and quickly completes the stratification of motion states, laying the foundation for subsequent detailed posture analysis.

[0053] Step S304: If the motion state is static, then determine the prone or standing posture based on the limb angle parameters.

[0054] Specifically, in a static state, the posture is further distinguished by the angle between the pet's torso and the support surface of the enclosure. An angle of less than 30° is identified as a lying posture, corresponding to a relaxed state where the pet's body is close to the bottom of the enclosure and its limbs are tucked in. An angle of more than 60° is identified as a standing posture, corresponding to a state where the pet's body is upright and its limbs are extended. An angle between 30° and 60° is identified as a semi-lying posture, and all of these are included in the overall posture characteristic parameters. By using angle thresholds, the static posture can be accurately quantified and distinguished, which intuitively reflects the pet's degree of body extension and physical comfort when it is static, providing a basis for subsequent state judgment.

[0055] Step S305: If the movement state is a small-range activity, then obtain the head swaying frequency and tail swaying amplitude.

[0056] Specifically, in a small-scale activity state, the head swaying frequency is obtained by counting the number of times the nose tip key point moves back and forth per minute, and the tail swaying amplitude is calculated by measuring the maximum pixel displacement distance of the tail root key point. Normal fluctuation thresholds are set in advance by combining pet's normal behavior data. Too fast a swaying frequency or too large a swaying amplitude is judged as a tendency of anxiety. These parameters are the core indicators for judging the pet's emotional fluctuations, environmental discomfort, loneliness and anxiety. Compared with single action recognition, they can more accurately reflect the pet's inner emotional state and make up for the shortcomings of visual monitoring in emotion recognition.

[0057] Step S306: If the movement state is a large-scale running and jumping, then obtain the running and jumping duration and the running and jumping trajectory.

[0058] Specifically, during periods of extensive running and jumping, two key indicators are monitored: first, the duration of running and jumping, which involves tracking the duration of continuous, large-amplitude movements. A duration exceeding one minute is considered excessive agitation, while a duration less than one minute is generally considered normal play behavior; second, the regularity of the running and jumping trajectory, determined by the movement paths at key points. Normal play exhibits regular trajectories and a concentrated activity range, while abnormal agitation presents chaotic trajectories and no fixed activity range. By combining these two indicators, we can accurately distinguish between normal play behavior and stress-induced abnormal agitation, avoiding indiscriminate intervention that could trigger secondary stress in pets.

[0059] Step S307: Fuse the limb angle parameters, motion amplitude parameters, motion state and corresponding posture determination results to generate real-time posture features.

[0060] Specifically, the real-time posture feature fusion adopts a normalization integration method, which unifies various heterogeneous data such as limb angle parameters, motion amplitude parameters, motion state, and refined posture judgment results into standardized values ​​in the 0-1 range. This eliminates the analysis bias caused by differences in the units and ranges of different parameters, and finally integrates them into a one-dimensional continuous feature vector. This feature vector completely contains all the core information of the pet's current posture and can be directly input into the subsequent behavior discrimination model as core input data, realizing the standardized and quantitative output of posture features and ensuring the consistency and accuracy of subsequent state discrimination.

[0061] In one embodiment of this example, step S107, which inputs multidimensional monitoring time-series data into the behavior discriminator to obtain the pet's status tag, includes steps S401 to S405: Step S401: Based on real-time posture features, obtain the frequency of abnormal postures of the pet.

[0062] Specifically, the frequency of abnormal postures refers to the cumulative number of times a pet exhibits abnormal postures such as limb tension, repeated curling, frequent tail wagging, abnormal lying down, and limb stiffness within a unit of time (1 minute). After sample verification, a frequency threshold of 5 times was set. If the number of abnormal postures exceeds 5 times per minute, it is judged as a significantly abnormal posture. If it is less than 5 times, it is considered as occasional movement fluctuation. This excludes temporary and accidental movement interference and improves the reliability of abnormal posture judgment.

[0063] Step S402: Obtain abnormal vocal segments based on vocal frequency and decibel value.

[0064] Specifically, the identification of abnormal vocalizations relies on a dual threshold standard. In terms of frequency, sharp barks above 2000Hz or low whimpers below 500Hz both exceed the normal vocalization frequency range of pets. In terms of decibels, vocalizations above 70dB are considered abnormally excited vocalizations. At the same time, a noise filtering algorithm is used to remove environmental noise, equipment operation noise, and other non-pet vocalization interference, retaining only abnormal vocalizations that meet the above thresholds, accurately distinguishing normal communication vocalizations from abnormal vocalizations of discomfort, anxiety, or pain.

[0065] Step S403: Determine whether the abnormal vocalizations overlap with the frequency of abnormal postures in time.

[0066] Specifically, time overlap determination refers to the period when the occurrence of abnormal vocalizations completely or partially overlaps with the period when abnormal postures are concentrated, and the overlap lasts for more than 3 seconds. Setting a 3-second overlap duration threshold is to eliminate the interference of single-dimensional instantaneous anomalies. Only when posture and sound anomalies occur synchronously and overlap continuously can they be determined as true state anomalies. This reduces the probability of misjudgment caused by single-dimensional data fluctuations from the root cause and improves the accuracy of state label determination.

[0067] Step S404: If there is overlap, extract the temperature and humidity data and air quality data corresponding to the overlapping time periods.

[0068] Specifically, after identifying the overlapping periods of abnormal posture and sound synchronization, the real-time average and peak data of temperature, humidity, and air quality within those periods are extracted, and key environmental indicators are selected. This type of data is used to investigate whether the abnormal state is induced by environmental factors, to distinguish between environmental stress and physiological discomfort, to avoid blindly determining the cause of the abnormality, and to provide accurate data support for subsequent targeted interventions.

[0069] Step S405: Based on the temperature and humidity data and air quality data corresponding to the overlapping time periods, determine whether there are environmental stress triggering factors.

[0070] Specifically, the determination of environmental stress triggers relies on the pet's comfortable environment threshold standard. This standard has been verified by pet physiological characteristics research and a large number of samples: the comfortable temperature range is 20-26℃, the comfortable relative humidity range is 40%-60%RH, the safe value of ammonia concentration is below 10ppm, and the safe value of carbon dioxide concentration is below 1500ppm. If any of these indicators exceeds the comfortable range, it is determined that there is an environmental stress trigger. Such abnormal indicators will directly cause the pet to feel uncomfortable and are the most important external causes of stress response in pets in enclosed spaces.

[0071] Step S406: If there are environmental stress triggers, mark the status label as an unwell state.

[0072] Specifically, environmental discomfort is categorized into a sub-type of agitation, clearly indicating that the core cause of the abnormal state is excessive environmental parameters, rather than the pet's own physiological problems. This sub-label can directly guide subsequent environmental control operations, enabling targeted treatment, avoiding indiscriminate intervention, and improving the efficiency of alleviating abnormal states.

[0073] Step S407: If there are no environmental stress triggers, mark the status label as a state of physiological discomfort.

[0074] Specifically, physiological discomfort is categorized into lethargy or severe anxiety. After ruling out abnormal environmental parameters as a cause, the pet's abnormal condition is determined to be caused by internal factors such as physical discomfort, underlying diseases, or joint injuries. Such conditions cannot be alleviated by environmental control and must be addressed through manual verification or medical intervention to avoid delaying the treatment of the pet's physical discomfort.

[0075] In one embodiment of this example, after marking the state label as an unwell state in step S406 if an environmental stress trigger exists, steps S501 to S505 are further included: Step S501: Obtain body part information associated with the frequency of abnormal postures.

[0076] Specifically, body part information is precisely extracted through skeletal key point localization, locking down the body areas corresponding to the concentration of abnormal postures, such as stiff hind limb joints, persistent head tension, stiff and immobile tail, and abnormally curled torso. Identifying abnormal body parts allows for precise location of the pet's discomfort area, avoiding large-scale vague screening and providing a clear target for subsequent historical data comparison and medical examination.

[0077] Step S502: Based on the body part information, retrieve the posture data of the same part from the historical monitoring data for comparison.

[0078] Specifically, historical monitoring data is stored in the pet's exclusive electronic file. Normal posture data and historical abnormal records of the same part of the pet within the past 7 days are retrieved and compared with the current abnormal posture data. The comparison dimensions include core parameters such as limb angle, range of motion, and frequency of posture. By referring to historical data, it is determined whether the current abnormality is a temporary fluctuation caused by temporary environmental stress or a persistent physical abnormality, and to distinguish between accidental abnormalities and pathological abnormalities.

[0079] Step S503: Based on the comparison results, determine whether there is a persistent posture shift in the body parts.

[0080] Specifically, the determination criterion for persistent posture deviation is that the posture parameters of the same body part deviate from the normal threshold range within 3 or more consecutive monitoring periods, rather than temporary abnormalities in a single instance or single period. This criterion can effectively distinguish between temporary stress responses and persistent problems such as potential joint injuries and physical illnesses, avoiding misjudging temporary discomfort as pathological abnormalities and preventing the omission of potential health hazards.

[0081] Step S504: If there is persistent posture deviation, generate a medical examination recommendation instruction and mark the body part information as the medical attention area.

[0082] Specifically, the medical examination recommendation instruction is a warning instruction to trigger professional medical intervention, and at the same time accurately mark the abnormal body part as the medical attention area, such as "persistent deviation of the hind limb joint" and "neck stiffness and tightness". Clearly marking the attention area can help pet owners and veterinarians quickly locate the discomfort area, shorten the inspection and troubleshooting time, improve the efficiency of medical treatment, and avoid delaying the illness.

[0083] Step S505: Send the medical examination recommendation instruction and the medical attention area to the service terminal.

[0084] Specifically, the instruction push is realized through the WiFi and Bluetooth wireless communication modules, and the medical examination recommendation instruction, the abnormal body part, and the real-time monitoring data are synchronously pushed to the pet owner's mobile APP, the foster care center management background, and the intelligent terminal supporting the airline container in real time, realizing the second-level synchronization of abnormal information, ensuring that relevant personnel can obtain the warning information in the first time and carry out the disposal work in time.

[0085] In one implementation manner of this embodiment, it further includes steps S601 to S605: Step S601: Obtain the environmental parameter type corresponding to the environmental stress trigger factor.

[0086] Specifically, the environmental parameter type refers to the specific environmental indicators that exceed the standard or are abnormal, which are subdivided into types such as too high / low temperature, too large / small humidity, excessive ammonia concentration, excessive carbon dioxide concentration, and substandard air quality. Accurately lock in single or multiple abnormal parameters, clarify the core objectives of environmental regulation, avoid blind regulation, and ensure that subsequent regulation instructions accurately correspond to abnormal indicators.

[0087] Step S602: Generate an environmental regulation instruction based on the environmental parameter type.

[0088] Specifically, the environmental control commands are one-to-one targeted commands, triggering the operation of corresponding equipment based on different abnormal parameters: excessively high temperature triggers ventilation and cooling, and semiconductor refrigeration commands; excessively low temperature triggers heat preservation and low-speed heating commands; excessively high humidity triggers dehumidification commands; excessively low humidity triggers atomized humidification commands; excessive ammonia and carbon dioxide trigger forced ventilation and air circulation commands; the commands directly control the environmental control module built into the flight case, with a fast response speed, no need for manual operation, and automatic correction of environmental parameters.

[0089] Step S603: Determine whether the status label changes from an unsuitable environmental state to a normal state within a preset time after the environmental control command is executed.

[0090] Specifically, the preset observation time is set to 8 minutes. This duration is based on the physiological characteristics of pets adapting to the environment, allowing both buffer time for environmental parameters to return to normal and time for the pet's physical perception and emotional calm. During the observation period, the pet's condition is continuously monitored to determine whether the pet has gradually recovered from an unpleasant state to a normal state after environmental regulation, thus verifying the effectiveness of environmental regulation.

[0091] Step S604: If the pet does not return to a normal state within a preset time, then reacquire the pet's real-time posture characteristics and vocal frequency.

[0092] Specifically, if the pet's condition does not return to normal within the 8-minute observation period, it indicates that the abnormal condition is not solely caused by environmental factors. It is necessary to immediately re-collect the pet's real-time posture characteristics, vocalization frequency, and decibel levels, eliminate environmental interference, focus on the pet's own behavior and physiological signals, and further investigate the root cause of the abnormality to avoid delays in treatment caused by ineffective single environmental control.

[0093] Step S605: Update the pet's status tag based on the reacquired real-time posture features and vocalization frequency.

[0094] Specifically, based on the newly collected core behavioral data, the status label is updated to "physiological discomfort," correcting the previous judgment of environmental discomfort and simultaneously triggering human intervention and medical warning instructions. This dynamic update mechanism avoids misjudgments caused by relying solely on environmental data, further improving the accuracy of status identification and ensuring that abnormal pet conditions are handled accordingly.

[0095] In one embodiment of this example, step S108, based on the status label, generates the corresponding service scheduling instruction, including steps S701 to S706: Step S701: Determine whether the status label is in an abnormal state.

[0096] Specifically, abnormal states refer to various discomforts, stresses, or pathological states that differ from normal states. These include states of anxiety, lethargy, environmental discomfort, and physiological discomfort. These states indicate that the pet's current physical sensations, emotions, or physiological conditions are abnormal and require targeted intervention through the system, rather than simply maintaining routine monitoring. Under normal conditions, the system only continuously completes multi-dimensional data collection, time-series alignment, and state discrimination processes, and simultaneously pushes routine monitoring data to the service terminal without triggering any scheduling intervention commands. By pre-judging abnormal states, the system achieves an intelligent scheduling logic of "silent monitoring when there are no abnormalities and graded response when there are abnormalities." This avoids meaningless interventions that interfere with the pet's normal activities while ensuring a rapid response to abnormal situations, balancing monitoring accuracy and pet comfort.

[0097] Step S702: If the status label is an abnormal status, extract the abnormal start time point from the multidimensional monitoring time series data.

[0098] Specifically, the anomaly start time point refers to the precise millisecond-level timestamp of the first abnormal posture or sound in a pet. This timestamp is the core benchmark for environmental data backtracking, used to locate changes in environmental parameters before and after the anomaly occurs, accurately trace the time node of the anomaly, and investigate the correlation between environmental changes and pet anomalies.

[0099] Step S703: Based on the anomaly start time point, backtrack to obtain the first environment dataset before the anomaly start time point and the second environment dataset after the anomaly start time point.

[0100] Specifically, based on the time point of the anomaly's onset, the first environmental dataset is extracted from the 10 minutes before that time point as a reference for the normal environment before the anomaly occurred; the second environmental dataset is extracted from the 10 minutes after the time point of the anomaly's onset as environmental fluctuation data after the anomaly occurred; the data from the 20 minutes before and after the anomaly form a complete comparative sample, comprehensively reflecting the trend of environmental parameter changes before and after the anomaly occurred, which facilitates accurate identification of mutation factors.

[0101] Step S704: Compare the first environmental dataset with the second environmental dataset to identify the inflection point of environmental change.

[0102] Specifically, the inflection point of environmental change refers to the time point when environmental parameters undergo significant changes in a short period of time. The criteria for judgment are that the change in parameters exceeds 20%, such as a sudden increase in temperature of more than 5°C within 10 minutes, a sudden change in humidity of more than 20%RH, or a sudden increase in ammonia concentration. Such inflection points are direct evidence of environmental-induced abnormalities in pets. By comparing datasets before and after the changes, the causal relationship between environmental changes and pet abnormalities can be quickly identified and clarified.

[0103] Step S705: If there is an inflection point in environmental change, a first type of scheduling instruction is generated based on the environmental parameters corresponding to the inflection point. The first type of scheduling instruction is used to automatically adjust the environmental equipment in the service area.

[0104] Specifically, the service dispatch instructions include the first dispatch instruction and the second dispatch instruction. The first type of dispatch instruction is a fully automated control instruction that does not require manual intervention. The control unit directly drives the built-in environmental regulation modules such as ventilation, cooling, humidification, and dehumidification in the flight case to operate, quickly correcting abnormal environmental parameters to the pet's comfort range. The entire process is unmanned and automated, with a fast response speed, and is suitable for mild abnormal states induced by environmental factors.

[0105] Step S706: If there is no environmental change inflection point, a second type of scheduling instruction is generated based on the abnormality type corresponding to the abnormal state. The second type of scheduling instruction is used to trigger manual intervention by service personnel.

[0106] Specifically, the absence of an inflection point in environmental change indicates that the pet's abnormality is unrelated to environmental parameters and is determined to be caused by its own physiological, emotional, or psychological factors. This immediately triggers a second type of manual intervention dispatch instruction, pushing abnormal information to the corresponding service personnel to remind them to promptly carry out manual treatment such as comforting, checking, and caring for the pet, preventing the abnormal state from continuing to worsen and ensuring the pet's safety.

[0107] In one embodiment of this example, step S706, generating a second type of scheduling instruction based on the exception type corresponding to the exception state, includes steps S801 to S805: Step S801: Obtain the real-time posture features and vocal frequencies corresponding to the abnormal state.

[0108] Specifically, extract core data such as real-time posture features, vocal frequency, and decibel values ​​from three or more consecutive sets during the abnormal period, remove instantaneous fluctuation data, and select the most representative stable abnormal data as the core basis for abnormal cause analysis to ensure the accuracy of subsequent behavior pattern determination and cause matching.

[0109] Step S802: Determine the abnormal behavior combination pattern based on real-time posture characteristics and vocal frequency.

[0110] Specifically, abnormal behavior combination patterns are a combination of multiple abnormal postures and abnormal vocalizations, rather than a single abnormal behavior. They are divided into three typical patterns: continuous pawing at the ground with high-frequency short vocalizations, curling up and remaining still with low-frequency whimpering, and repeated circling with intermittent barking. Different combination patterns correspond to different causes of discomfort in pets. Combination-based judgment is more accurate than single-behavior judgment and significantly reduces the misjudgment rate.

[0111] Step S803: Based on the abnormal behavior combination pattern, match the preset abnormal cause library to obtain a list of candidate abnormal causes.

[0112] Specifically, the abnormal cause database is a pre-established standardized database that stores common triggers corresponding to various abnormal behavior patterns. For example, high-frequency short cries often correspond to loneliness, anxiety, and a desire for companionship; low-frequency whimpering often corresponds to physical pain, gastrointestinal discomfort, and joint injury; and repeated digging often corresponds to restlessness and a depressing environment. By matching patterns, a list of candidate abnormal causes can be quickly generated, narrowing down the scope of cause investigation.

[0113] Step S804: Obtain the pet's service history, which includes historical abnormal events and corresponding historical solutions.

[0114] Specifically, the service history is linked to the pet's unique identification (chip number, QR code), containing comprehensive information such as the pet's past abnormal events, treatment methods, recovery status, dietary preferences, and environmental adaptation habits. This data is personalized and exclusive to the pet. Combining the history can eliminate interference from common causes and analyze the reasons for abnormalities in accordance with the individual characteristics of the pet.

[0115] Step S805: Perform correlation analysis between the candidate list of abnormal causes and the service history, and filter out the target abnormal cause that has the highest matching degree with the current abnormal behavior combination pattern.

[0116] Specifically, the matching degree between candidate abnormal causes and pet history is calculated using the cosine similarity algorithm, and the cause with a matching degree higher than 80% is selected as the target abnormal cause. This screening method combines a general cause database with the pet's personalized habits to eliminate irrelevant interference items and lock in the real abnormal cause that best fits the current pet, so as to achieve personalized and accurate judgment.

[0117] Step S806: Based on the cause of the target abnormality, generate targeted reassurance and guidance instructions or medical intervention instructions, and push the reassurance and guidance instructions or medical intervention instructions to the mobile terminal of the target service personnel.

[0118] Specifically, the second type of dispatch instructions includes reassurance and guidance instructions and medical intervention instructions. Reassurance and guidance instructions target non-health-related abnormalities such as anxiety, loneliness, hunger, and thirst, and include specific operational suggestions such as gentle reassurance, companionship and interaction, appropriate feeding and watering, and environmental reassurance. Medical intervention instructions target health-related abnormalities such as suspected physical injury, disease, and persistent discomfort, and include key examination areas, medical advice, and precautions for temporary care. Both types of instructions are accurately pushed to the mobile terminals of the corresponding service personnel, guiding staff to quickly carry out targeted treatment and improve intervention efficiency.

[0119] In one embodiment of this example, after generating the corresponding service scheduling instruction in step S108, steps S901 to S904 are further included: Step S901: Obtain the pet's identification and a complete record of this monitoring.

[0120] Specifically, pet identification uses a unique electronic chip number or QR code to accurately distinguish different pets and avoid data confusion. The complete monitoring record covers the entire lifecycle of data, including raw sensor data, posture feature data, status judgment results, scheduling instructions, execution results, environmental parameter change curves, etc., to achieve full traceability of the monitoring process.

[0121] Step S902: Link the complete record to the pet's electronic file.

[0122] Specifically, the pet's electronic record is stored in the cloud with encryption, linked to the pet's unique identification, and retains all-dimensional information such as monitoring records, status data, treatment plans, and health changes for a long time, forming a complete digital monitoring history for the pet; the record can be retrieved, viewed, and exported at any time, making it easy for pet owners, boarding facilities, and veterinarians to fully understand the pet's status and health.

[0123] Step S903: Based on the complete records, analyze the trend of pet status changes during multiple service processes.

[0124] Specifically, the state change trend analysis relies on the longitudinal comparison of multi-period monitoring data to explore the state adaptation patterns of pets in different environments, time periods, and scenarios, including the adaptation time to unfamiliar environments, preferred temperature and humidity ranges, periods of easy anxiety, activity patterns, and high-incidence scenarios of abnormalities, forming a quantitative trend report, upgrading from single monitoring to full-cycle behavior analysis.

[0125] Step S904: Based on the trend of status changes, generate personalized pet service suggestions, which include preferred service time periods and recommended service types.

[0126] Specifically, personalized pet service recommendations are generated based on the pet's condition change trends, and the service time is selected when the pet's condition is most stable and its emotions are most calm, reducing the stress risk during boarding and transportation. Recommended service types include environmental parameter settings, care frequency, soothing methods, and boarding plans that are adapted to the pet's habits, realizing an upgrade from standardized monitoring to personalized and precise monitoring, and meeting the pet's refined care needs.

[0127] Secondly, this application also discloses a pet intelligent monitoring flight case based on multiple sensors.

[0128] Reference Figure 2 and Figure 3This application also discloses an aviation crate, which is an enclosed integrated cabin structure adapted to the breeding and transportation standards of small and medium-sized dogs and cats. The core includes the main body 1, which adopts an integrated sealed enclosure structure. It is composed of six independent panels spliced ​​and fixed together: a top cover, a bottom load-bearing base plate, a left side panel, a right side panel, a front side panel, and a rear side panel. The six panels are seamlessly connected by an integrated injection molding process. The joints are all rounded and blunted, with no sharp edges or exposed splicing gaps. Together, they form an independent and closed pet-accommodating cavity, providing the pet with a dedicated activity space while isolating it from external interference.

[0129] The main body 1 of the box is equipped with four core functional components in one integrated manner, corresponding to the preset installation position: box door 2, detachable lick-type water feeder 3, rotating feeding pull-out box 4, and visual intelligent AI module 5. Each component is fixed and matched with the main body 1 through a dedicated connection structure, which is firm and easy to assemble and disassemble. There are no additional exposed accessories. The overall structure is compact and neat, taking into account structural stability, functional integrity and ease of use.

[0130] Specifically, the main body of the carrier 1 is made of thickened environmentally friendly ABS composite pressure-resistant material, with an internal fiberglass reinforcement layer. It possesses excellent impact resistance, compression resistance, drop resistance, and wear resistance. The material itself is non-toxic, odorless, and free of harmful substances, fully complying with international air transport and pet boarding material safety and hygiene standards. The carrier is lightweight, balancing portability with load-bearing requirements. The inner surface of the bottom load-bearing plate has an anti-slip textured finish and is equipped with a removable silicone pad. The pad is detachably connected to the bottom plate via snap fasteners, improving pet comfort and facilitating future removal, cleaning, and replacement. A one-piece molded handle is centrally located on the top of the main body 1, covered with an anti-slip silicone sleeve for a comfortable grip and easy carrying. Wear-resistant and anti-slip feet are fixed at all four corners of the bottom, ensuring stability and preventing slippage when placed on a flat surface or in the transport compartment, effectively reducing stress to the pet caused by carrier movement.

[0131] The enclosure door 2 is fitted to the pre-set opening position on the front panel of the main body 1. It is connected to the front panel via two sets of symmetrically distributed stainless steel silent hinges. The two sets of hinges are fixed to the upper and lower center positions of the left and right edges of the enclosure door 2, respectively. The hinge bases are securely connected to the enclosure door 2 and the front panel using countersunk screws, ensuring a firm and stable connection. The opening and closing process is silent and smooth, without any jamming or abnormal noise. The opening angle can be fully extended, facilitating pet entry and exit and internal cleaning. The enclosure door 2 features a high-density perforated mesh structure with perforations controlled at 8-10mm. This moderate size ensures continuous airflow between the internal cavity and the outside, preventing a confined space. The enclosure is designed to be warm and oxygen-deficient, effectively preventing pets from sticking their limbs out of the enclosure and causing external injuries. It also allows outsiders to easily observe the pet's condition inside. The enclosure door 2 is equipped with a double safety locking mechanism, consisting of a top push-button latch and a bottom rotating safety lock. The two sets of latches are independent and must be operated simultaneously to unlock. When closed, the double latches lock simultaneously, effectively preventing the door from loosening due to the pet's struggle or collision, eliminating the risk of the pet escaping on its own, and balancing safety and ease of operation. When the enclosure door 2 is closed, the edge of the door panel and the opening of the front panel are tightly sealed with a sealing strip, reducing external noise, strong light, and airflow into the enclosure, and reducing the pet's stress response to the unfamiliar environment.

[0132] The detachable lick-type water feeder 3 is installed in the upper middle section of the right side panel of the main body 1. A matching rectangular mounting slot is pre-drilled on the right side panel. The feeder is detachably and securely connected to the mounting slot via a snap-on quick-release base. The quick-release base is firmly attached to the inner wall of the right side panel with screws. The feeder body and base are connected by elastic snap-fit ​​mechanisms. Disassembly and assembly can be completed with one hand without any tools, facilitating daily water replenishment, cleaning, and disinfection. The detachable lick-type water feeder 3 consists of a sealed water storage chamber, a lick-type spout, and a connecting tube. These three parts are sealed and interconnected without leakage. The water storage chamber has a capacity of 300ml, but can be replaced as needed to fully fill the container. Designed to meet the drinking needs of small and medium-sized pets for short trips and long-term boarding, the water storage chamber is made of transparent PC material with capacity markings on the outer wall for easy and intuitive viewing of the remaining water level. A sealed water inlet is located at the top of the chamber; after filling, the sealing plug closes to achieve a complete seal, preventing water leakage even when the enclosure is tilted or shaken during transport. The bottom of the water storage chamber is connected to a licking-style water outlet via a connecting tube. This licking-style water outlet penetrates the right side panel of the main body 1, with the outlet facing inwards. Made of food-grade flexible silicone, it conforms to the pet's licking drinking habits, allowing for controlled water flow. Water is only dispensed when the pet is licking and automatically closes to stop the flow when not licking, effectively preventing choking and swallowing, while also preventing water leakage and contamination of the enclosure's interior.

[0133] The rotating pull-out feeding box 4 is embedded in the lower middle section of the rear side panel of the main body 1. The rear side panel has a corresponding pull-out mounting groove, and the entire box is completely embedded inside the groove. When closed, the outer side is flush with the surface of the box, without protruding from the outside and occupying no additional external space. The rotating pull-out feeding box 4 consists of four parts: a pull-out door, a detachable feeding bowl, side rotating hinges, and limiting buckles. The pull-out door is fixedly connected to the rear side panel via two sets of side rotating hinges located at the upper and lower ends of the left edge of the door, enabling lateral rotation for opening and closing. Fixed at 90°, it opens and closes smoothly without taking up external space; the feeding bowl is fixedly installed in the center of the inside of the pull-out box door, and is detachably connected to the pull-out box door by a snap fastener. It can be removed and cleaned separately after feeding, completely avoiding food residue and bacterial growth; there is a limit buckle on the right edge of the pull-out box door. When closed, the buckle is locked with the rear panel to prevent the pet from rushing and opening it on its own. When feeding or adding food, there is no need to open the main box door 1. Only the small pull-out box door needs to be opened to complete the operation. The pet is kept away from the outside environment throughout the process, reducing stress and collisions, and eliminating the risk of the pet escaping during the feeding process.

[0134] The Visual Intelligent AI Module 5 is embedded in the center of the top cover of the main body 1 of the enclosure. The module is embedded into the cavity of the enclosure from the inside of the top cover downwards, with the top of the module flush with the outer surface of the top cover. This does not occupy the pet's activity space inside the enclosure and does not affect the pet's normal lying, walking, and activity. The edge of the module is sealed to the top cover with a waterproof sealing ring to prevent dust, moisture, and pet saliva from entering the module and damaging the electronic components, thus extending the module's lifespan. The Visual Intelligent AI Module 5 is integrated into five parts: a pressure-resistant and flame-retardant shell, a distributed multi-sensor network, an embedded control unit, a wireless communication module, and a built-in power module. All electronic components are integrated and connected via a circuit board. The module shell is made of pressure-resistant and flame-retardant material, meeting protection standards and suitable for various travel and boarding scenarios. The embedded control unit is a high-performance embedded microcontroller chip. The chip is pre-programmed with the aforementioned complete set of pet intelligent monitoring algorithms, which can independently complete the entire process of data collection, analysis and processing, status judgment, command output, and information push without the need for external processing equipment, achieving integrated hardware and software operation.

[0135] The distributed multi-sensor network is integrated at the bottom of the module, facing the interior of the enclosure. It includes at least four core sensor types: a wide-angle high-definition camera, a high-precision temperature and humidity sensor, an air quality sensor, and a high-sensitivity sound collector. All four types of sensors are electrically connected to the embedded control unit via built-in wires. The control unit synchronously receives real-time data collected by various sensors and performs intelligent judgment. The wireless communication module supports WiFi and Bluetooth dual-mode connectivity, enabling real-time wireless linkage with pet owners' mobile apps and boarding center management backends. It can simultaneously push pet status, environmental parameters, and abnormal warning information. The built-in power module has a battery life of 6-8 hours, meeting the needs of short-distance transport and long-term boarding. It is also equipped with a Type-C universal charging interface, supporting continuous power supply from an external power source, and is suitable for use in multiple scenarios. In addition, the sensor layout eliminates monitoring blind spots. The wide-angle high-definition camera achieves all-round video acquisition inside the cavity. The temperature and humidity sensor monitors the temperature and humidity values ​​inside the cavity in real time. The air quality sensor focuses on monitoring key pollution indicators in the pet's enclosed space, such as ammonia and carbon dioxide concentrations. The sound collector is equipped with a noise filtering module to accurately pick up pet sounds and eliminate environmental noise interference, achieving comprehensive intelligent monitoring of the pet's status and the enclosure environment.

[0136] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A multi-sensor-based intelligent pet monitoring method, characterized in that, include: Acquire real-time monitoring data collected by a distributed multi-sensor network deployed in the pet service area, including video stream data, environmental sensor data, and sound acquisition data; Based on the video stream data, extract the skeletal key point sequence of the pet; Based on the skeletal key point sequence, the real-time posture features of the pet are generated; Based on the environmental sensing data, acquire temperature and humidity data and air quality data of the service area; Based on the sound data collected, the frequency and decibel value of the pet's barks are identified; The real-time posture features, temperature and humidity data, air quality data, vocal frequency, and decibel value are aligned along the time axis to generate multi-dimensional monitoring time-series data. The multidimensional monitoring time series data is input into the behavior discriminator to obtain the pet's status tags, which include normal state, anxious state or lethargic state. Based on the status label, a corresponding service scheduling instruction is generated.

2. The pet intelligent monitoring method based on multiple sensors according to claim 1, characterized in that, The extraction of the pet's skeletal keypoint sequence based on the video stream data includes: Based on the video stream data, the location of the detection box for the pet is obtained using an object detection algorithm; Based on the location of the detection box, the pet area image is cropped; Instance segmentation is performed on the pet region image to obtain the pet's outline mask; Based on the contour mask and the pet region image, the skeletal topology of the pet is extracted; Based on the skeletal topology, a predetermined number of key joints are located, including the tip of the nose, the eye point, the base of the ear, the shoulder joint, the hip joint, and the base of the tail. The key joints are tracked and matched in multiple consecutive frames of images to generate the skeletal key point sequence.

3. The pet intelligent monitoring method based on multiple sensors according to claim 1, characterized in that, The generation of real-time pose features of the pet based on the skeletal keypoint sequence includes: Based on the coordinates of adjacent key points in the skeletal key point sequence, calculate the pet's limb angle parameters; Based on the displacement of specific key points in the skeletal key point sequence, calculate the pet's motion amplitude parameters; Determine the pet's movement status, which includes static state, small-scale activity, and large-scale running and jumping; If the movement state is static, then the prone or standing posture is determined based on the limb angle parameters. If the movement state is a small-range activity, then obtain the head swaying frequency and tail swaying amplitude; If the movement state is a large-scale running and jumping, then obtain the running and jumping duration and the running and jumping trajectory; The limb angle parameters, the motion amplitude parameters, the motion state, and the corresponding posture determination results are fused to generate the real-time posture features.

4. The pet intelligent monitoring method based on multiple sensors according to claim 1, characterized in that, The step of inputting the multidimensional monitoring time-series data into the behavior discriminator to obtain the pet's status tags includes: Based on the real-time posture features, the frequency of abnormal postures of the pet is obtained; Based on the frequency of the call and the decibel value, obtain the abnormal call segment; Determine whether the frequency of the abnormal vocalization overlaps with the frequency of the abnormal posture in time; If there is overlap, extract the temperature and humidity data and the air quality data corresponding to the overlapping time period; Based on the temperature and humidity data and air quality data corresponding to the overlapping time periods, determine whether there are environmental stress triggering factors. If there are environmental stress triggers, the state label will be marked as an environmentally unwell state. If no environmental stress triggers are present, the state label will be marked as a state of physiological discomfort.

5. The pet intelligent monitoring method based on multiple sensors according to claim 4, characterized in that, After marking the state label as an unwell state if environmental stress triggers exist, the method further includes: Obtain body part information associated with the frequency of the abnormal posture; Based on the body part information, posture data of the same part in historical monitoring data are retrieved and compared; Based on the comparison results, it is determined whether the body parts exhibit persistent postural deviation. If persistent posture deviation exists, a medical examination suggestion instruction is generated, and the body part information is marked as a medical attention area. The medical examination suggestion instruction and the medical attention area are sent to the service terminal.

6. The pet intelligent monitoring method based on multiple sensors according to claim 4, characterized in that, Also includes: Obtain the environmental parameter type corresponding to the environmental stress triggering factor; Based on the environmental parameter types, generate environmental control instructions; Determine whether the status label changes from an unsuitable environmental state to a normal state within a preset time after the environmental control command is executed; If the pet does not return to a normal state within a preset time, the real-time posture characteristics and vocal frequency of the pet will be reacquired. The pet's status tag is updated based on the reacquired real-time posture features and vocal frequency.

7. The pet intelligent monitoring method based on multiple sensors according to claim 1, characterized in that, The service scheduling instruction includes a first scheduling instruction and a second scheduling instruction; generating the corresponding service scheduling instruction based on the status label includes: Determine whether the status label is in an abnormal state; If the status label is an abnormal status, then the abnormal start time point is extracted from the multidimensional monitoring time series data; Based on the anomaly start time point, backtrack to obtain the first environmental dataset before the anomaly start time point and the second environmental dataset after the anomaly start time point; The first environmental dataset is compared with the second environmental dataset to identify inflection points in environmental change; If the environmental change inflection point exists, a first type of scheduling instruction is generated based on the environmental parameters corresponding to the environmental change inflection point. The first type of scheduling instruction is used to automatically adjust the environmental equipment in the service area. If the environmental change inflection point does not exist, a second type of scheduling instruction is generated based on the abnormality type corresponding to the abnormal state. The second type of scheduling instruction is used to trigger manual intervention by service personnel.

8. A multi-sensor-based intelligent pet monitoring method according to claim 7, characterized in that, The second type of scheduling instructions includes reassurance and guidance instructions and medical intervention instructions; the generation of the second type of scheduling instructions based on the abnormality type corresponding to the abnormal state includes: Obtain the real-time posture features and vocal frequency corresponding to the abnormal state; Based on the real-time posture characteristics and the vocalization frequency, an abnormal behavior combination pattern is determined; Based on the abnormal behavior combination pattern, a preset abnormal cause library is matched to obtain a list of candidate abnormal causes; Obtain the pet's service history, which includes historical abnormal events and corresponding historical solutions; The candidate anomaly cause list is correlated with the service history record to filter out the target anomaly cause with the highest matching degree with the current anomaly behavior combination pattern; Based on the cause of the target anomaly, a targeted reassurance and guidance instruction or medical intervention instruction is generated and pushed to the mobile terminal of the target service personnel.

9. A multi-sensor-based intelligent pet monitoring method according to claim 1, characterized in that, After generating the corresponding service scheduling instruction, the following is also included: Obtain the pet's identification and a complete record of this monitoring; Link the complete record to the pet's electronic file; Based on the complete records, analyze the trend of pet's status changes during multiple service processes; Based on the aforementioned status change trend, personalized pet service recommendations are generated, including preferred service time periods and recommended service types.

10. A flight case, characterized in that, The enclosure includes a main body, which is formed by the top, bottom, left side wall, right side wall, front wall and rear wall to create a closed cavity for accommodating a pet. The main body also integrates an enclosure door, a detachable lick-type water feeder, a rotating pull-out feeding box and a visual intelligent AI module. The cabinet door is located on the front wall of the main body of the cabinet and is connected to the cabinet by a hinge for opening and closing. The detachable lick-type water feeder is installed on the right side wall of the main body of the enclosure. It includes a water storage chamber, a lick-type water outlet, and a snap-on mounting base. The water storage chamber is connected to the water outlet, and the entire water feeder can be quickly disassembled through snap-on fasteners. It is used to provide drinking water for pets and facilitates cleaning and hydration. The rotating feeding pull-out box includes a pull-out box door and a feeding bowl. The feeding bowl is fixedly connected to the inner wall of the pull-out box door, and the pull-out box door is hinged to the rear wall of the main body of the box for opening and closing. The visualized intelligent AI module is installed on the top of the main body of the enclosure and embedded inside the cavity of the enclosure from the top. It includes a distributed multi-sensor network and a control unit. The distributed multi-sensor network includes at least a camera, a temperature and humidity sensor, an air quality sensor, and a sound collector. The control unit is electrically connected to the camera, the temperature and humidity sensor, the air quality sensor, and the sound collector, respectively.

Images