Pet behavior training system based on AI vision

Abstract

The invention provides a pet behavior training system based on AI vision, and belongs to the technical field of pet training. The system comprises a pet positioner, a visual acquisition module, a data processing module, a strategy matching module and an execution feedback module, and realizes pet individual distinguishing, behavior feature extraction and behavior type and confidence judgment by integrating positioning information, physiological signals and image data. Analyzing the physiological information to obtain a basic emotional state and quantized emotional intensity and stability parameters; and pet types are determined in combination with a species classification model, a species exclusive reward and punishment strategy library is called, the reward and punishment strength is dynamically adjusted according to behavior attributes, confidence coefficients and emotion parameters, and a complete closed-loop training mechanism is formed. The method solves the problems of subjective misjudgment, poor species adaptability, neglect of pet emotions and the like in traditional training, improves the behavior recognition accuracy, emotion perception objectivity and reward and punishment strategy precision, and is suitable for multi-species and multi-scene scientific and efficient pet behavior training.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pet training, in particular to a pet behavior training system based on AI vision. BACKGROUND

[0002] In recent years, as pets become important family members, the multi-pet cohabitation feeding mode gradually popularizes, and people's demand for pet behavior training is significantly improved. At the same time, the relevance between the emotional state of the pet and the training effect is increasingly concerned, and a scientific training method needs to take into account the behavior performance and the inner emotion of the pet to achieve more efficient and more personalized guidance.

[0003] However, the existing pet training related technology has obvious limitations in the multi-pet scene: it is difficult to accurately distinguish the behavior of different pet individuals, often leading to the misapplication of rewards and punishments to non-target pets, and the training pertinence is greatly discounted; and in the training process, only fixed strategies are mechanically executed according to the behavior type, and the real-time emotional state of the pet cannot be perceived, so it is impossible to adjust the training intensity according to the emotional fluctuations, and it is also difficult to avoid the resistance or stress reaction caused by the emotional mismatch, which restricts the scientificity and effectiveness of the training. SUMMARY

[0004] The present application provides a pet behavior training system based on AI vision to solve the problem that the existing technology cannot distinguish different pet individuals, resulting in confusion of reward and punishment objects, and cannot adjust the training intensity according to the emotional fluctuations.

[0005] In order to achieve the above object, the embodiment of the present application provides a pet behavior training system based on AI vision, characterized in that the pet behavior training system comprises: a pet locator comprising a positioning device and a physiological signal sensor, which is used for collecting positioning information and physiological information of a pet in real time, and the pet locator is integrated with a voice prompter; a vision acquisition module comprising a camera installed at a specified position and a smart feeder, which is used for collecting image information of the pet; a data processing module, which is used for distinguishing different pet individuals according to the positioning information and the image information, and is also used for extracting features of the positioning data and the image data to obtain a behavior feature vector, and determining a behavior type and a behavior confidence of the pet behavior based on the behavior feature vector to judge whether the behavior is reliable, and is also used for analyzing the physiological information to determine a basic emotional state of the pet, and calculating an emotional intensity parameter and an emotional stability parameter of the pet through a preset algorithm; a strategy matching module, which is used for determining a pet species according to the image information, and selecting and executing a corresponding reward and punishment strategy in combination with the behavior type, the behavior confidence of the pet and a preset reward and punishment rule library, and the reward and punishment strategy is used for training the pet behavior by controlling the voice prompter and the smart feeder; and an execution feedback module, which is used for adjusting a core intensity of the reward and punishment strategy according to the emotional intensity parameter of the pet in the execution process of the reward and punishment strategy, and secondarily adjusting the adjusted core intensity according to the emotional stability parameter and a preset adjustment rule.

[0006] Optionally, each pet locator is internally provided with a distinguishing label, and the process of distinguishing different pet individuals by the data processing module comprises: receiving the positioning information and the corresponding distinguishing label sent by each pet locator; performing target detection on the collected image information to extract image features of each pet; matching the space-time coordinates of the positioning information with the space-time coordinates of the image information, associating the image features with the positioning information in the same space-time range to form association information; matching each association information with the corresponding distinguishing label to distinguish different pet individuals; after distinguishing the pet individuals for the first time, a mapping relationship database of the distinguishing label and the corresponding image features and positioning information is established, and the distinguishing of different pet individuals is realized by matching the distinguishing label in the mapping relationship database with the positioning information and the image information collected in real time.

[0007] Optionally, the feature extraction comprises: for the positioning information and the image information in the same association information, performing smoothing processing on the positioning information to obtain a continuous pet position sequence, and filtering clear frames and identifying a pet target area from the image information; calculating a displacement distance, a turning angle and a stay time in a unit time based on the position sequence to obtain a motion feature; obtaining a space interaction feature by counting a distance change sequence of the pet and a key environmental area; extracting a joint node coordinate, a limb angle and a motion trajectory based on the pet target area in the image information to obtain a limb feature; and fusing the motion feature, the space interaction feature and the limb feature to obtain the behavior feature vector of the pet.

[0008] Optionally, the data processing module is configured with a preset behavior template library, the behavior template library includes typical feature vectors of each behavior type, and the process of determining the behavior type and the behavior confidence includes: comparing the behavior feature vector with the typical feature vectors in the behavior template library to calculate a matching degree; predicting the behavior feature vector through a preset classification model to calculate probability values of each behavior type; fusing the calculated matching degree and probability values according to a preset weight to obtain a comprehensive score, selecting the behavior type with the highest comprehensive score as the recognition result, and performing standardization processing on the comprehensive score as the behavior confidence of the behavior type.

[0009] Optionally, the physiological signal sensor includes a heart rate sensor and an electrodermal sensor, which are respectively used to collect a heart rate signal and an electrodermal signal of the pet to constitute physiological information of the pet, and the analysis of the physiological information includes: performing filtering and noise reduction processing on the collected heart rate signal and electrodermal signal, and extracting heart rate variability and electrodermal conductance level; comparing the heart rate variability and the electrodermal conductance level with corresponding preset normal intervals to obtain a comparison result; and determining a basic emotional state of the pet according to a preset judgment rule based on the comparison result.

[0010] Optionally, the preset algorithm includes: calculating emotional components of the pet based on motion features, physiological information and behavior recognition results of the pet, and combining the basic emotional state of the pet, and calculating reliabilities of each emotional component, the emotional components include a heart rate related emotional component, an electrodermal waveform feature emotional component and an image behavior emotional component; correcting the emotional components based on the reliabilities of the emotional components; weighting the three corrected emotional components according to their reliabilities to constitute an emotional vector, and performing normalization processing; calculating an emotional intensity parameter based on the normalized emotional vector and preset weights corresponding to each emotional component; calculating fluctuation standard deviations of the three emotional components in a preset time window, and performing weighted summation and normalization to obtain an emotional stability parameter.

[0011] Optionally, the calculation of each emotional component and the corresponding reliability comprises: constructing a motion interference index based on motion features, and filtering the heart rate signal to obtain an emotion-related heart rate signal; extracting heart rate variability features from the emotion-related heart rate signal, and converting them into heart rate-related emotional components through a correlation function according to the exclusive reference value and the limit value of the species corresponding to the basic emotional state; combining heart rate signal quality and motion interference stripping effect to calculate the reliability of the heart rate-related emotional component; extracting the waveform features of the skin conductance signal within a preset time window and performing normalization processing; selecting the corresponding preset feature weight according to the basic emotional state, weighting the normalized waveform features, and obtaining the skin conductance waveform feature emotional component through function processing; combining waveform integrity and signal quality to calculate the reliability of the skin conductance waveform feature emotional component; based on the behavior recognition result, determining whether the behavior type is an out-of-training behavior; if so, according to the basic emotional state, the emotional type and the basic emotional value corresponding to the out-of-training behavior are retrieved from the preset mapping table, the basic emotional value is corrected according to the behavior confidence, the behavior duration feature and the image quality, and the image behavior emotional component is obtained; combining the behavior confidence and the emotional mapping certainty to calculate the reliability of the image behavior emotional component.

[0012] Optionally, the strategy matching module comprises a species classification model, which is trained and constructed based on a plurality of labeled sample images of pet species, and the determination of the pet species comprises: extracting a key image frame containing the complete shape of the pet from the collected image information, and performing species feature extraction; inputting the extracted species features into the preset species classification model, and the species classification model compares the species features with a preset species feature library to output the species label with the highest matching degree as the pet species.

[0013] Optionally, the selection process of the reward and punishment strategy comprises: the strategy matching module calls the reward and punishment sub-library corresponding to the determined pet species from the preset reward and punishment rule library based on the determined pet species, and the sub-library pre-stores a mapping relationship table between behavior types and reward and punishment strategies; retrieving the reward and punishment strategy consistent with the currently determined behavior type from the mapping relationship table as a candidate reward and punishment strategy; comparing the current behavior confidence according to the preset confidence threshold to determine whether the current behavior is a reliable behavior; if it is a reliable behavior, the candidate reward and punishment strategy is determined as the reward and punishment strategy to be executed; if it is not a reliable behavior, no reward and punishment is performed.

[0014] Optionally, the attribute of the behavior type includes a positive attribute and a negative attribute, and the execution feedback module is further configured to determine a behavior attribute corresponding to the behavior type of the pet according to the behavior type of the pet, and the adjustment of the reward and punishment strategy includes: acquiring an emotion intensity parameter of the pet in a reward and punishment strategy execution process in real time; when the behavior of the pet is judged to be a reliable behavior, grading the emotion intensity of the pet according to the emotion intensity parameter, and selecting a corresponding initial adjustment strategy according to the emotion intensity grade to adjust the core strength of the reward and punishment strategy; dividing a fluctuation grade according to the emotion stability parameter, and selecting a corresponding secondary adjustment strategy in combination with the behavior attribute to perform secondary adjustment on the reward and punishment strategy.

[0015] The pet behavior training system based on AI vision provided by the application is a closed-loop intelligent training system constructed by integrating a pet locator, a vision acquisition module, a data processing module, a strategy matching module and an execution feedback module. The system can acquire and fuse multi-dimensional data of the pet in real time, extract behavior characteristics and identify behavior types through the data processing module, and analyze emotion intensity and stability parameters reflected by physiological information; the strategy matching module selects and executes a reward and punishment strategy accordingly, and the execution feedback module can dynamically adjust the execution strength of the strategy according to the emotion parameters, so as to realize precise, adaptive and humanized training of the behavior of the pet, and effectively improve the scientificity and effectiveness of the training. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor. In the drawings: Figure 1 is a framework diagram of the pet behavior training system provided by the embodiment of the application; Figure 2 is a pet behavior recognition flowchart provided by the embodiment of the application; Figure 3 is a calculation flowchart of emotion intensity parameters and emotion stability parameters provided by the embodiment of the application; Figure 4 is a reward and punishment strategy adjustment flowchart provided by the embodiment of the application. DETAILED DESCRIPTION

[0017] The specific embodiments of the embodiments of the application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the embodiments of the application, and are not used to limit the embodiments of the application.

[0018] It should be noted that the acquisition, transmission, storage, use, processing, etc. of data in the technical solutions of the present application comply with the relevant provisions of national laws and regulations. In the embodiments of the present application, some existing industry solutions may be mentioned, such as software, components, models, etc. They should be considered as exemplary, and their purpose is only to illustrate the feasibility of the implementation of the technical solutions of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0019] As mentioned above, with the increasing demand for fine pet feeding, the traditional pet behavior training mode gradually exposes limitations, and artificial training relies on subjective experience and cannot quantify the emotional state of pets; the existing intelligent system mostly uses single vision or positioning data, and the behavior recognition accuracy is low; at the same time, in the multi-pet coexistence scene, the individual differentiation mechanism is missing, and the reward and punishment objects are easily confused, making it difficult to achieve targeted training. Therefore, it is an urgent need to develop an intelligent training system that integrates multi-source data, accurately perceives emotions, and adapts to multi-pet scenes.

[0020] To solve this problem, the present application provides a pet behavior training system based on AI vision: the positioner collects positioning and physiological information, and the vision module obtains images; the data processing module integrates feature recognition behavior and confidence to analyze physiological emotional parameters; the strategy matching module generates an initial strategy in combination with the species and rule base, the feedback module dynamically adjusts the reward and punishment intensity according to the emotional parameters, and finally the voice prompter and intelligent feeder are executed to solve the problems of lack of emotional perception, difficulty in individual differentiation, and rigid strategy in traditional training, and improve the training accuracy and pertinence.

[0021] The following will be described in detail Figures 1-4 The present application is described in detail.

[0022] As Figure 1As shown, the embodiment of the present application provides an AI vision-based pet behavior training system, characterized in that the pet behavior training system comprises: a pet locator comprising a positioning device and a physiological signal sensor, for collecting positioning information and physiological information of the pet in real time, and the pet locator is integrated with a voice prompter; a vision acquisition module comprising a camera installed at a specified position and a smart feeder, for acquiring image information of the pet; a data processing module for distinguishing different pet individuals according to the positioning information and the image information; also for extracting features from the positioning data and the image data to obtain a behavior feature vector, and determining a behavior type and a behavior confidence of the pet behavior based on the behavior feature vector to judge whether the behavior is reliable; also for analyzing the physiological information to determine a basic emotional state of the pet, and calculating an emotional intensity parameter and an emotional stability parameter of the pet through a preset algorithm; a strategy matching module for determining a pet species according to the image information, and selecting and executing a corresponding reward and punishment strategy in combination with the behavior type, the behavior confidence of the pet, and a preset reward and punishment rule library, the reward and punishment strategy being used to train the pet behavior by controlling the voice prompter and the smart feeder; and an execution feedback module for adjusting a core intensity of the reward and punishment strategy according to the emotional intensity parameter of the pet during execution of the reward and punishment strategy, and secondarily adjusting the adjusted core intensity according to the emotional stability parameter and a preset adjustment rule.

[0023] Among them, the behavior feature vector refers to high-dimensional quantitative data formed by fusing motion features, spatial interaction features and limb features after the data processing module extracts features from the positioning information and the image information, and is used to accurately describe the comprehensive features of the pet behavior; the behavior confidence is a quantitative index for measuring the reliability of the behavior, and the higher the value is, the more reliable the recognition result is; the core intensity refers to a key adjustment parameter in the reward and punishment strategy that directly acts on the pet, including the amount of snacks of the smart feeder, the volume and duration of the encouragement sound of the voice prompter for positive behavior; and the volume and duration of the warning sound of the voice prompter for negative behavior; the basic emotional state refers to the emotional category obtained by analyzing physiological signals such as heart rate and skin electricity of the pet, such as calm, excitement, stress, etc.; the emotional intensity parameter is an index quantified from the basic emotional state (range 0-1), reflecting the strength of the emotion; and the emotional stability parameter is an index quantifying the fluctuation degree of the pet emotion.

[0024] The pet locator in the application adopts the patent technology of the prior application of the applicant, the disclosure number is CN118749950A, and the invention name is a pet locator and a positioning method thereof. The specific structure includes a base, a female positioning device, a male positioning device, and a binding belt. The female positioning device integrates a first positioning chip, a signal processor, and a physiological signal sensor, including a heart rate sensor and a skin electricity sensor. The male positioning device is provided with a second positioning chip. The binding belt is designed to be elastic and cooperates with a magnetic mounting table to realize the connection and disconnection of the male positioning device. The above structure and basic working principle are consistent with the prior patent, and will not be repeated here. The traditional pet locator cannot obtain the heart rate, skin electricity activity and other key physiological signals of the pet, which leads to the inability to analyze the emotional state of the pet through physiological indicators. With the emergence of the prior application patent, the emotional state of the pet can be realized.

[0025] Specifically, after the system is started, the pet locator collects real-time positioning information of the pet through the positioning device, and simultaneously acquires physiological information by using the physiological signal sensor. The integrated voice prompter is on standby. The visual acquisition module synchronously collects pet image information through the camera on the specified position and the intelligent feeder. The data processing module distinguishes different pet individuals according to the collected data, extracts features of the positioning data and image data, generates a pet behavior feature vector, determines the behavior type and behavior confidence, analyzes the physiological information to obtain the basic emotional state, and calculates the emotional intensity parameter and the emotional stability parameter. The strategy matching module determines the pet species according to the image information, combines the behavior type, the behavior confidence, and the preset reward and punishment rule library, selects the corresponding reward and punishment strategy, and executes the reward and punishment strategy through the control of the voice prompter and the intelligent feeder. The execution feedback module adjusts the strategy core strength according to the emotional intensity parameter in the reward and punishment process, and then adjusts the adjusted strength according to the emotional stability parameter and the preset rule, to form a complete behavior training closed loop.

[0026] The pet behavior training system provided by the embodiment of the application realizes accurate differentiation of multiple pet individuals by integrating positioning, physiological signals, and visual data, avoids confusion of reward and punishment objects, and significantly improves the pertinence of training. At the same time, by extracting behavior features and quantifying behavior confidence, the reliability of behavior recognition is ensured. In combination with the emotional intensity and stability parameters obtained by analyzing the physiological information, the reward and punishment strategy can dynamically adapt to the real-time state of the pet, which not only reduces the stress reaction caused by mechanical execution of the strategy, but also improves the scientificity and effectiveness of the training, and meets the fine and personalized pet behavior training requirements.

[0027] Preferably, each of the pet locators is internally provided with a distinguishing label, and the process of distinguishing different pet individuals by the data processing module comprises: receiving the positioning information and the corresponding distinguishing label sent by each pet locator; performing target detection on the collected image information to extract image features of each pet; matching the space-time coordinates of the positioning information with the space-time coordinates of the image information, associating the image features and the positioning information in the same space-time range to form association information; matching each association information with the corresponding distinguishing label to distinguish different pet individuals; after distinguishing the pet individuals for the first time, a mapping relationship database of the distinguishing label, the corresponding image features and the positioning information is established, and the distinguishing label in the mapping relationship database is matched through real-time collection of the positioning information and the image information to realize the distinction of different pet individuals.

[0028] wherein the distinguishing label refers to unique identification information such as a unique ID code internally provided in the pet locator, each label is bound to a single pet, and is used as a core identification of the individual identity in a multi-pet scene to ensure the uniqueness of data association; the space-time coordinate matching refers to that the data processing module compares the time stamp of the positioning information with the collection time stamp of the image information, simultaneously calculates the spatial distance between the positioning coordinates and the physical position of the image collection device, and determines that it is in the same space-time range when the time and space conditions are satisfied at the same time, thereby realizing the preliminary association of the positioning information and the image features; the association information refers to the combined data of the positioning information and the image features after the space-time coordinate matching; and the mapping relationship database refers to a database storing the corresponding relationship between the distinguishing label, the image feature template and the positioning information features, wherein the image feature template is the feature fusion result of multiple clear images of the same pet, and the positioning information features are statistical features such as the commonly used activity area and the motion speed range of the pet, which are used for subsequent rapid matching.

[0029] For example, a family raises a golden retriever wearing a distinguishing tag ID001 and a Persian cat wearing a distinguishing tag ID002. The distinguishing process of different pet individuals is as follows: first, the locators of the two pets respectively send real-time positioning information to the data processing module. The golden retriever is in the living room coordinate X1, Y1, and the Persian cat is in the balcony coordinate X2, Y2, and their respective distinguishing tags; at the same time, the indoor camera collects the images of the living room and the balcony, and extracts the image features of the golden retriever "yellow short hair, medium body type" and the Persian cat "white long hair, small body type". The data processing module determines the same space-time range by comparing the time stamp and spatial distance of the positioning information, associates the positioning information of the golden retriever with the yellow short hair feature to form associated information, binds ID001, associates the positioning information of the Persian cat with the white long hair feature to form associated information, binds ID002, and completes the first individual distinguishing. Subsequent systems will establish a mapping relationship library based on these associated information, store ID001 corresponding to "yellow short hair template, living room commonly used activity area", and ID002 corresponding to "white long hair template, balcony commonly used activity area"; when the "yellow short hair" image or the living room positioning signal is detected again, the characteristics of ID001 in the library are matched, so that the golden retriever can be quickly identified, and the Persian cat can be identified in the same way, realizing the accurate distinguishing of multiple pet individuals.

[0030] The individual distinguishing scheme provided in the preferred embodiment of the present application realizes one-to-one binding among positioning, image and identity in the multi-pet scene by configuring a unique distinguishing tag for each pet and combining the spatio-temporal coordinate accurate matching of positioning information and image features; subsequently, the corresponding relationship between the tags and the features is stored by establishing a mapping relationship library, so that the individual distinguishing is upgraded from full-process matching to feature rapid comparison, which not only solves the core problems of behavior attribution confusion and reward and punishment object error in traditional multi-pet raising, but also greatly improves the distinguishing efficiency through feature library reuse, lays a foundation for subsequent accurate behavior training for different individuals, and significantly enhances the applicability and reliability of the system in the multi-pet family scene.

[0031] Preferably, the feature extraction comprises: for the positioning information and the image information in the same associated information, performing smoothing processing on the positioning information to obtain a continuous pet position sequence, and screening clear frames and identifying a pet target area of the image information; based on the position sequence, calculating a displacement distance, a turning angle and a stay time in a unit time to obtain a motion feature; based on the distance change sequence of the pet and the key environment area, obtaining a space interaction feature; based on the pet target area in the image information, extracting node coordinates, limb angles and motion trajectories to obtain a limb feature; and fusing the motion feature, the space interaction feature and the limb feature to obtain a behavior feature vector of the pet.

[0032] The smoothing processing adopts a sliding window average method for the positioning information to eliminate instantaneous positioning errors and generate a continuous pet position coordinate sequence, such as one coordinate point per second. The clear frame screening refers to retaining pet image frames without blurring and without occlusion by image sharpness score, and adopting a target detection algorithm such as YOLOv5 to segment the pet target region from the clear frame. The motion feature calculation includes: the total displacement distance in a unit of time, the change amount of the turning angle, and the stay time in the same coordinate region. The key environmental region refers to a preset key region of the home scene, such as a feeding area, a sofa area, a forbidden area, etc. The space interaction feature refers to calculating quantitative indicators such as the approach frequency and the stay proportion by statistically analyzing the distance change sequence of the pet and each key region. In the limb feature extraction, the joint point coordinates can be identified by using an improved version of OpenPose, which can identify the pixel coordinates of 12 key points such as the head, neck, torso, and limb ends of the pet. The limb angle refers to the angle formed by the connecting lines of adjacent joint points. The action trajectory is generated by the change path of the continuous frame joint point coordinates. The feature fusion adopts a weighted splicing method to combine the motion feature, the space interaction feature, and the limb feature according to the weight of 1:1:2 to form a multi-dimensional behavior feature vector, which completely represents the dynamic and static attributes of the pet behavior.

[0033] Specifically, the feature extraction process is as follows: for the positioning information and image information in the same association information, the positioning information is first smoothed by using a sliding window average method to eliminate instantaneous errors and generate a continuous position sequence; at the same time, the edge gradient value of each frame of image information is calculated, the clear frame without blurring and without occlusion is screened out, and the pet target region containing the complete torso and limbs is segmented from the clear frame by using a target detection algorithm. Based on the continuous position sequence, the total displacement distance in 1 minute, the change amount of the turning angle of the adjacent coordinate vector angle, and the stay time in the same region are calculated to obtain 3-dimensional motion features; the key environmental regions such as the feeding area and the sofa area are determined by user labeling, the distance sampling sequence of the pet and each region is calculated once per second, and 5-dimensional space interaction features such as the approach frequency and the stay proportion are calculated; the pixel coordinates of 12 key joint points such as the head, neck, torso, and limb ends of the pet are identified by using a skeleton detection model, the elbow joint angle and the hip joint angle formed by the connecting lines of adjacent joint points are calculated, and 20-dimensional limb features are generated by the continuous frame joint point coordinate trajectory. Finally, the motion feature, the space interaction feature, and the limb feature are weighted and spliced according to the weight of 1:1:2 to fuse into a 28-dimensional behavior feature vector, which completely represents the dynamic and static attributes of the pet behavior.

[0034] The feature extraction scheme provided by the preferred embodiment of the present application effectively eliminates noise interference in the original data through smoothing of the positioning information and screening of clear frames of the image, thereby providing high-quality basic data for feature extraction; through multi-dimensional feature splitting, a comprehensive depiction of pet behavior from a macroscopic to a microscopic perspective is realized, thereby avoiding one-sidedness of a single feature dimension; and finally, the behavior feature vector is fused, which not only retains unique information of each feature, but also highlights the key role of limb details in behavior recognition through weight distribution. This multi-source fusion feature extraction method improves the matching degree of subsequent behavior recognition and significantly reduces the probability of misjudgment of similar behaviors.

[0035] As shown in Figure 2 Preferably, the data processing module is configured with a preset behavior template library, the behavior template library includes typical feature vectors of each behavior type, and the process of determining the behavior type and the behavior confidence includes: comparing the behavior feature vector with the typical feature vectors in the behavior template library to calculate the matching degree; predicting the behavior feature vector through a preset classification model to calculate the probability value of each behavior type; fusing the calculated matching degree and probability value according to a preset weight to obtain a comprehensive score, selecting the behavior type with the highest comprehensive score as the recognition result, and standardizing the comprehensive score as the behavior confidence of the behavior type.

[0036] The behavior template library is a database for storing typical feature vectors of each preset behavior type such as dismantling a house, eating, quiet rest, etc., each behavior type contains 500-1000 sample vectors, and the samples cover different pet species and behavior intensity differences; the matching degree calculation adopts a cosine similarity algorithm, specifically, the cosine value of the angle between the to-be-recognized behavior feature vector and the standard feature center vector of a behavior type in the behavior template library, and the value closer to 1 indicates a higher feature coincidence degree; the preset classification model adopts a lightweight CNN-LSTM hybrid model, the input is a 28-dimensional feature vector, and the output is a prediction probability of each behavior type, the model is trained through labeled samples, and the parameters are optimized through cross-validation to ensure the generalization of cross-species behavior recognition; in the preset weight fusion, the matching degree weight is set to 0.4, the classification model probability value weight is set to 0.6, and the fusion formula is: comprehensive score = matching degree x 0.4 + highest probability value output by the classification model x 0.6.

[0037] Specifically, taking the identification of pet housebreaking behavior as an example, the data processing module inputs the behavior feature vector containing the features of pet biting sofa, long stay in the sofa area, rapid displacement, etc. into the system; first, the behavior template library is called, the cosine similarity between the vector and the standard feature center vector of the housebreaking behavior in the library is calculated, and the matching degree is 0.85; at the same time, the lightweight CNN-LSTM hybrid model is started, the model predicts the input vector, and outputs the probability value of housebreaking behavior 0.92; according to the preset weight fusion, the comprehensive score = 0.85*0.4+0.92*0.6=0.892; finally, the comprehensive score is standardized (0.8 corresponds to 1.0, 0.3 corresponds to 0.0), and the behavior confidence is 0.98, and the final output behavior type is "housebreaking" and the behavior confidence is 0.98.

[0038] The preferred embodiment of the application verifies the behavior feature vector and the template library matching degree and the classification model prediction probability value in two dimensions, which not only utilizes the feature similarity to ensure the basic accuracy of behavior recognition, but also improves the generalization ability in cross-species and complex scenes by means of CNN-LSTM model, effectively solving the limitations of single identification method. The behavior confidence obtained by standardizing the comprehensive score quantifies the reliability of the identification result, avoids the misjudgment and misexecution of low confidence behaviors, improves the behavior type recognition accuracy, significantly reduces the error rate of similar behavior differentiation, provides a core judgment basis for the accurate triggering of subsequent reward and punishment strategies, and significantly enhances the recognition reliability and applicability of the system to complex behaviors of pets.

[0039] Preferably, the physiological signal sensor includes a heart rate sensor and a skin electricity sensor, respectively used for collecting the heart rate signal and the skin electricity signal of the pet to constitute the physiological information of the pet, and the analysis of the physiological information includes: filtering and denoising the collected heart rate signal and skin electricity signal, and extracting heart rate variability and skin electricity level; comparing the heart rate variability and the skin electricity level with the corresponding preset normal interval to obtain a comparison result; and determining the basic emotional state of the pet according to the preset judgment rule based on the comparison result.

[0040] The filtering denoising processing includes removing high-frequency noise by using a 5Hz low-pass filter for the heart rate signal, eliminating motion artifacts by wavelet transform, using a 0.01-1Hz band-pass filter for the skin conductance signal, retaining the effective skin conductance fluctuation component, and eliminating slow drift by baseline correction; the heart rate variability refers to the time variation of consecutive normal heart intervals, and the R-wave peak of the preprocessed heart rate signal is detected; the preset normal interval includes the normal interval corresponding to different species; the judgment rule of the basic emotional state is as follows: when the heart rate variability is greater than the upper limit of the normal interval and the skin conductance level is less than the lower limit of the normal interval, it is judged to be calm; when both are within the normal interval, it is judged to be neutral; when the heart rate variability is less than the lower limit of the normal interval and the skin conductance level is greater than the upper limit of the normal interval, it is judged to be excited / tense; when the skin conductance level is greater than 20% of the upper limit of the normal interval and the heart rate variability is less than 20% of the lower limit of the normal interval, it is judged to be highly stressed.

[0041] Specifically, the physiological information analysis process is as follows: the heart rate sensor and the skin conductance sensor integrated in the pet locator collect pet heart rate signals and skin conductance signals respectively; a 5Hz low-pass filter is used to remove high-frequency noise from the heart rate signal and wavelet transform is used to eliminate motion artifacts, a 0.01-1Hz band-pass filter is used to retain the effective fluctuation of the skin conductance signal and baseline correction is used to eliminate drift; the R-wave peak is detected from the preprocessed heart rate signal, and the standard deviation of the adjacent R-wave interval is calculated as the heart rate variability index; the smoothed direct current component is extracted from the skin conductance signal as the skin conductance level index; the two indexes are compared with the preset normal interval, such as heart rate variability of 50-150ms and skin conductance of 1-5μS, the basic emotional state is judged by the judgment rule, for example, the heart rate variability of a pet is 160ms and the skin conductance level is 0.8μS, at this time the emotional state of the pet is calm.

[0042] The physiological information analysis scheme provided by the preferred embodiment of the present application breaks through the limitation of traditional manual subjective observation of pet emotions by converting pet physiological signals such as heart rate and skin conductance into quantifiable indexes, combining species-specific normal physiological intervals, and judging the basic emotional state according to clear rules, thereby realizing objective and accurate determination of the basic emotion of the pet. This quantitative analysis based on physiological signals provides a reliable basis for the generation of subsequent emotional intensity and stability parameters, effectively avoids the deviation of training strategy caused by subjective misjudgment of emotional state, and enhances the scientificity and accuracy of the system in perceiving the internal state of the pet.

[0043] As Figure 3As shown, preferably, the preset algorithm comprises: based on the motion characteristics, physiological information and behavior recognition result of the pet, combining the basic emotional state of the pet, calculating the emotional components of the pet, and calculating the reliability of each emotional component, the emotional components include heart rate associated emotional component, skin electrical waveform feature emotional component and image behavior emotional component; based on the reliability of each emotional component, the emotional component is corrected; the three corrected emotional components are weighted according to their reliability to form an emotional vector, and normalized processing is performed; based on the normalized emotional vector and the corresponding preset weight of each emotional component, the emotional intensity parameter is calculated; the fluctuation standard deviation of the three emotional components in the preset time window is calculated, and weighted summation and normalization are performed to obtain the emotional stability parameter.

[0044] Further preferably, the calculation of each emotional component and the corresponding reliability comprises: constructing a motion interference index based on the motion characteristics, and filtering the heart rate signal to obtain an emotional associated heart rate signal; the heart rate variability feature is extracted from the emotional associated heart rate signal, and according to the exclusive reference value and the limit value of the species corresponding to the basic emotional state, the heart rate associated emotional component is obtained through the association function; the reliability of the heart rate associated emotional component is calculated by combining the heart rate signal quality and the motion interference stripping effect; the waveform feature of the skin electrical signal is extracted in the preset time window and normalized; the normalized waveform feature is weighted and calculated according to the corresponding preset feature weight selected according to the basic emotional state, and the skin electrical waveform feature emotional component is obtained through function processing; the reliability of the skin electrical waveform feature emotional component is calculated by combining the waveform integrity and the signal quality; based on the behavior recognition result, it is determined whether the behavior type is an out-of-training behavior; if yes, the emotional type and the basic emotional value corresponding to the out-of-training behavior are retrieved from the preset mapping table according to the basic emotional state, the basic emotional value is corrected by combining the behavior confidence, the behavior duration feature and the image quality, and the image behavior emotional component is obtained; the reliability of the image behavior emotional component is calculated by combining the behavior confidence and the emotional mapping certainty.

[0045] Specifically, the motion interference index is used to quantify the influence degree of motion on heart rate signal, which is constructed based on the motion characteristics collected by the pet locator, reflecting the rule that the more intense the motion, the greater the interference on heart rate; the species exclusive reference value is the heart rate variability feature under the calm state of the pet, such as the standard deviation of normal heartbeat interval, and the limit value is the maximum value of the heart rate variability under the stress state of the pet; the motion interference index Based on the motion characteristics: 、 , and the species exclusive training coefficient: 、 、 , the formula is:

[0046] The emotion-related heart rate signal is obtained by low-pass filtering the original heart rate signal and stripping the influence of the motion interference indicator M through a Kalman filtering algorithm. The heart rate-related emotion component is obtained by extracting the heart rate variability feature from the emotion-related heart rate signal, comparing the species-specific reference value and the limit value corresponding to the basic emotional state, and obtaining the result through a sigmoid correlation function. The specific calculation formula is:

[0047] is a sigmoid function that maps to [0, 1]; , is the heart rate variability reference value in the pet's calm state; , is the maximum heart rate variability in the stress state; , is the feature weight, which is 0.6, 0.4 respectively. The reliability of the heart rate-related emotion component is calculated as follows:

[0048] is the noise ratio, which is the amplitude ratio of noise in the heart rate signal; is the filter residual ratio, which is the fitting degree of the emotion-related heart rate signal and the original signal. The closer they are to 1, the higher the reliability.

[0049] Specifically, the skin electrical waveform feature refers to the key morphological parameters of the skin electrical signal in a preset time window, including the maximum peak , the rise time , and the fluctuation frequency , reflecting the change in skin conductance level caused by emotional changes. The three features form a feature vector F=[ ]; the feature weight is the influence weight of each waveform feature on emotion judgment based on the random forest model training W=[ ; extracting and normalizing the waveform feature specifically refers to band-pass filtering the skin electrical signal, retaining the 0.01-1Hz frequency band, and filtering out irrelevant noise. In a preset time window, extract the maximum peak, rise time, and fluctuation frequency, and convert the feature value to a value in the [0, 1] interval through linear normalization; weighted calculation specifically refers to weighting the normalized features according to the feature weight corresponding to the basic emotional state, such as peak weight 0.5, rise time weight 0.3, and frequency weight 0.2, and then summing the weighted features to obtain the skin electrical waveform feature emotion component through the tanh function. The specific calculation formula is: ​​

[0050]

[0051] wherein, is the hyperbolic tangent function, which enhances the suppression of extreme values; =1.2, =0.1 are fitting parameters; is the characteristic reference value in a calm state, is the characteristic maximum value in a stress state; , , The reliability of the skin electrical waveform characteristic emotional component The calculation formula of is as follows:

[0052] wherein, is the proportion of complete waveform, which is the proportion of effective waveform containing peak value, rising edge and falling edge; is the signal-to-noise ratio, which is the ratio of signal power to noise power, and the closer to 1, the higher the reliability.

[0053] Specifically, the out-of-training behavior refers to a behavior that is not a system preset training target, such as natural behaviors such as hitching and hiding in corners; the preset mapping table pre-stores the corresponding relationship between the out-of-training behavior type-emotion type-basic emotion value, such as hitching corresponding to happiness, basic emotion value 0.7, and hiding in corners corresponding to nervousness, basic emotion value 0.6; the emotion mapping certainty indicates the association reliability of the behavior type and the corresponding emotion type, such as the mapping accuracy of hitching and happiness is 90%, and the certainty coefficient is 0.9; the calculation formula of the image behavior emotion component is as follows:

[0054] wherein, is the confidence weight, taking a value of 0.6; is the behavior duration proportion; is the behavior image clarity score. The calculation formula of the corresponding reliability is as follows:

[0055] wherein, is the behavior confidence, is the image clarity, is the emotion label certainty coefficient, and the closer to 1, the higher the reliability.

[0056] Specifically, the correction logic for emotion components is as follows: Using high-reliability emotion components as a benchmark, low-reliability emotion components are adjusted to ensure that all components reflect consistent emotional trends. High-reliability emotion components refer to those with a reliability reaching a preset threshold, such as ≥0.8, while low-reliability emotion components are those with a reliability below the preset threshold. When at least one of the three emotion components is high-reliability, correction is initiated: based on the emotion type corresponding to the high-reliability component, such as stress, a preset correction coefficient matrix is ​​retrieved. This matrix contains adjustment factors for each component under that emotion type, and the values ​​of the low-reliability components are fine-tuned. For example, if the heart rate-related emotion component is low-reliability, while the high-reliability skin conductance waveform feature emotion component indicates stress, then the value of the heart rate-related emotion component is increased by a correction coefficient, such as 1.2, to make it more consistent with the characteristics of stress.

[0057] Specifically, emotion vector The expression is ,in, , , These are the correction values ​​for the three emotion components. , , These represent their respective reliability; normalization refers to transforming the sentiment vector into a normalized vector with values ​​in the range [0,1] using the L2 norm. The formula is: , For vectors The L2 norm is the square root of the sum of squares of the elements; the formula for calculating the emotional intensity parameter is:

[0058] in, The preset weights for each emotion component are as follows: =0.3、 =0.3、 =0.4, "" indicates vector dot product operation; the calculation process for the emotion stability parameter is to first calculate the standard deviation of the fluctuation of the three emotion components within a preset time window, such as 3 seconds. , , Then, calculate the weighted summation according to the component weights. Finally, the emotion stability parameter is obtained through normalization, and its specific expression is:

[0059] in The maximum baseline value for the species' emotional fluctuations was determined through sample statistics.

[0060] Taking a dog pet in a stress basic emotional state as an example, and a sampling time window of 30s, the preset algorithm calculates the process and specific values of the emotional intensity parameter and the emotional stability parameter as follows: first, based on the motion characteristics collected by the pet locator, such as the displacement per unit time 0.8m / s and the acceleration 1.2m / s 2 , physiological information collected by the physiological signal sensor, such as the heart rate signal after filtering, the heart rate variability SDNN is 45ms, the skin electrical signal peak value is 3.2μS, and the behavior recognition result output by the visual acquisition module, such as the behavior type is cowering and trembling, and the behavior confidence is 0.9, combined with the stress basic emotional state, three emotional components are calculated respectively: The heart rate related emotional component: the motion interference index M is constructed by the motion characteristics M = 5 × 0.8 + 3 × 1.2 + 2 × 0.8 × 1.2 = 8.32, the emotional related heart rate signal is extracted after stripping the interference, combined with the HRV limit value of 50ms and the calm reference value of 80ms when the dog is stressed, and the sigmoid function is E1 = 0.82; the synchronization reliability R1 = (1-0.1) × 0.95 = 0.855. The skin electrical waveform feature emotional component: the skin electrical waveform feature is extracted, and the weighted sum is 0.78 according to the stress feature weight peak value 0.5, the rise time 0.3, and the frequency 0.2, and the tanh function is E2 = 0.68; the reliability R2 = 0.9 × 0.85 = 0.765, the complete waveform proportion is 0.9, the signal-to-noise ratio is 0.85, and <0.8 is low reliability. The image behavior emotional component: cowering and trembling corresponds to the basic emotional value 0.8 in the preset mapping table, combined with the behavior confidence 0.9, the behavior duration proportion 0.8, and the image definition 0.95, the calculation is E3 = 0.8 × (0.6 × 0.9 + 0.4 × 0.8) × 0.95 = 0.6624; the reliability R3 = 0.9 × 0.95 × 0.9 = 0.7695, the emotional mapping certainty is low reliability. Then taking E1 as the reference, the stress state correction coefficient matrix is called to correct E2 and E3: E2 = 0.68 × 1.1 + 0.05 × (0.82-0.68) = 0.734, E3 = 0.7618. Then the corrected components are weighted according to the reliability to construct the emotional vector V = [0.82 × 0.855, 0.734 × 0.765, 0.7618 × 0.7695] ≈ [0.7011, 0.5615, 0.5862], which is normalized by L2 norm ≈[0.587,0.469,0.491]; combining the preset weights W1=0.3, W2=0.3, W3=0.4 of each component, the emotional intensity parameter I=0.587*0.3+0.469*0.3+0.491*0.4≈0.512 is calculated. Finally, the standard deviations σ1=0.12, σ2=0.15, σ3=0.13 of the three components in 30s are calculated, and the total fluctuation value =0.12*0.3+0.15*0.3+0.13*0.4=0.133 is obtained by weighted summation, combined with the maximum reference value of emotional fluctuation of dogs under stress =0.3, the normalized emotional stability parameter S≈0.443 is obtained.

[0061] The preferred embodiment of the present application realizes the precise quantification of pet emotions by constructing heart rate, skin electricity, image behavior multi-dimensional emotional components and combining with basic emotional state, while introducing reliability evaluation and correction mechanism; wherein, the component correction and weighted fusion based on reliability ensure the accuracy of the emotional vector, and the normalization processing and weight adaptation calculation logic make the emotional intensity parameter accurately reflect the emotional strength and the emotional stability parameter effectively represent the emotional fluctuation degree, finally providing a scientific and dynamic emotional basis for the reward and punishment strategy adjustment in pet behavior training, significantly improving the training pertinence and effectiveness.

[0062] Preferably, the strategy matching module includes a species classification model, which is trained and constructed based on a plurality of labeled sample images of pet species, and the determination of the pet species includes: extracting a key image frame containing the complete shape of the pet from the collected image information, and performing species feature extraction; inputting the extracted species features into the preset species classification model, and the species classification model compares the species features with a preset species feature library, and outputs the species label with the highest matching degree as the pet species.

[0063] The key image frame refers to a single frame image of a complete pet body without occlusion selected from continuous images of the visual acquisition module; the species feature extraction pointer extracts species distinguishing visual features, including body size parameters, hair features, head features, etc., from the key image frame through a feature extraction network such as an improved ResNet-18; the species classification model is trained in a supervised learning manner, takes the labeled sample image as input, associates the extracted species feature vector with the corresponding species label, optimizes the model parameters through a cross-entropy loss function, and finally outputs the matching probability value of each species; the preset species feature library stores standard feature templates of each species, each template is the mean value of the feature vectors of multiple samples of the species, for example, the golden retriever template includes the fusion features of “medium body size, yellow short hair, and drooping ears”, which is used to assist the classification model in feature comparison; the calculation of the matching degree includes the calculation of the cosine similarity of the input feature vector and each species template in the library by the species classification model, and the combination of the species probability value predicted by the model itself to obtain the matching degree of each species.

[0064] The classification scheme provided by the preferred embodiment of the present application realizes high-precision identification of the species of the pet by extracting species distinguishing features from the key image frame, and combining the species classification model with the double comparison of the preset species feature library, thereby effectively solving the problem of poor strategy adaptability caused by the undifferentiated treatment of species in the traditional training system.

[0065] Preferably, the selection process of the reward and punishment strategy includes: the strategy matching module calls the reward and punishment sub-library corresponding to the determined pet species from the preset reward and punishment rule library based on the determined pet species, and the mapping relationship table between the behavior type and the reward and punishment strategy is pre-stored in the sub-library; the reward and punishment strategy consistent with the currently determined behavior type is retrieved from the mapping relationship table as a candidate reward and punishment strategy; the current behavior confidence is compared according to the preset confidence threshold to determine whether the current behavior is a reliable behavior; if it is a reliable behavior, the candidate reward and punishment strategy is determined as the reward and punishment strategy to be executed; if it is not a reliable behavior, no reward and punishment is performed.

[0066] The preset reward and punishment rule library is a structured rule set constructed according to the classification of pet species, and at least includes exclusive reward and punishment sub-libraries of common pet species such as Canidae and Felidae, and each sub-library is customized based on the ethology characteristics and training needs of the corresponding species; each record in the mapping relationship table includes a behavior type, at least one reward and punishment mode, a basic execution parameter, and an associated behavior attribute label; the confidence threshold is a quantitative standard for judging the reliability of the behavior, and is preferably 0.7.

[0067] Specifically, the selection process of the reward and punishment strategy is as follows: the strategy matching module first determines the pet species according to the image information, such as identifying as a canine, and calls the corresponding species-specific sub-library from the preset reward and punishment rule library; then, the mapping relationship table in the sub-library is searched to retrieve records consistent with the current identified behavior type, such as the behavior type of the canine's fixed-point defecation, to obtain candidate reward and punishment strategies, such as dispensing 10g of snacks + playing 2 seconds of encouragement sound; then, the confidence of the current behavior, such as 0.85, is compared with the fixed threshold value 0.7, and since 0.85 is greater than 0.7, it is determined that the behavior is reliable, and the candidate strategy is confirmed to be executed.

[0068] The preferred embodiments of the present application realize the accurate matching of the reward and punishment strategy with the pet species and the reliability of the behavior by the calling logic of the species-specific reward and punishment sub-library, the reliable behavior screening mechanism based on the fixed confidence threshold value, and the species adaptation priority selection rule of multiple candidate strategies: on the one hand, the species-specific sub-library avoids the problem of insufficient adaptability of the general strategy to different pets, ensuring that the strategy conforms to the behavior characteristics of the species; on the other hand, the fixed threshold value screening effectively filters the false reward and punishment of low-confidence behaviors, reducing invalid operations.

[0069] As shown in Figure 4 Preferably, the attributes of the behavior type include positive attributes and negative attributes, and the execution feedback module is further configured to determine the behavior attributes corresponding to the behavior type of the pet according to the behavior type of the pet, and the adjustment of the reward and punishment strategy includes: acquiring the emotion intensity parameter of the pet in the execution process of the reward and punishment strategy in real time; when the behavior of the pet is determined to be a reliable behavior, classifying the emotion intensity of the pet according to the emotion intensity parameter, and selecting a corresponding initial adjustment strategy according to the emotion intensity level to adjust the core intensity of the reward and punishment strategy; dividing the fluctuation level according to the emotion stability parameter, and selecting a corresponding secondary adjustment strategy in combination with the behavior attributes to perform secondary adjustment on the reward and punishment strategy.

[0070] Specifically, the execution feedback module determines the initial attribute corresponding to the behavior type through a preset mapping relationship; when the behavior confidence corresponding to the behavior type is greater than a preset attribute determination threshold value, the initial attribute is taken as the final attribute of the behavior; if the behavior confidence is lower than the attribute determination threshold value, the attribute is not determined temporarily. The positive attribute corresponds to the behavior of the pet that conforms to the training target, such as fixed-point defecation and following instructions, and the negative attribute corresponds to the behavior that needs to be corrected, such as destroying the house and defecating anywhere. The behavior list of the two types of attributes is stored in a behavior-attribute mapping table, and the mapping relationship refers to the one-to-one correspondence rule of the behavior type and the attribute in the mapping table, for example, fixed-point defecation is mapped to positive, and destroying the house is mapped to negative. For example, the behavior type of a pet is destroying the house, and according to the mapping relationship, it is determined that the behavior is a negative attribute. When the system identifies the feature vector of the behavior, the behavior confidence is calculated to be 0.85, which is greater than the determination threshold value, for example, 0.7, so the negative attribute is directly taken as the final attribute, triggering the corresponding punishment strategy.

[0071] Specifically, the emotion intensity grading is divided based on the emotion intensity parameter: 0-0.3 is low intensity, 0.3-0.7 is medium intensity, and 0.7-1 is high intensity; each pet emotion intensity grade corresponds to an initial adjustment strategy, such as low intensity corresponding to 1.2 times the basic force, medium intensity corresponding to 1.0 times, and high intensity corresponding to 0.8 times, for example, for positive behavior reward, if the emotion intensity is low, the core force is increased by 20%, such as the feeding amount is increased to 1.2 times; the fluctuation level is divided according to the emotion stability parameter value: 0.7-1 is stable, 0.3-0.7 is moderate fluctuation, and 0-0.3 is severe fluctuation. The secondary adjustment strategy combines the behavior attribute, when the positive behavior is rewarded, the stable level corresponds to the core force which is increased by 10%, and the severe fluctuation corresponds to the force after the initial adjustment; when the negative behavior is punished, the stable level corresponds to the force after the initial adjustment, and the severe fluctuation corresponds to the core force which is decreased by 20%.

[0072] Taking the pet housebreaking behavior as an example, when the system identifies the behavior confidence as 0.85, the reward and punishment adjustment is started: first, it is detected that the pet emotion intensity parameter is 0.8, which belongs to high intensity excited state, and the basic punishment force such as prompt volume is reduced by 20% according to the initial adjustment strategy; at the same time, it is monitored that the emotion stability parameter is 0.2, which belongs to severe fluctuation level, and the secondary adjustment is triggered by combining the negative behavior attribute, and the prompt volume is reduced by 20% again, completing the dynamic adjustment from the core force to the emergency strategy.

[0073] The preferred embodiment of the present application provides a technical solution by real-time combination of the emotion intensity parameter and the emotion stability parameter of the pet in the reward and punishment process, dynamic adjustment of the reward and punishment strategy according to the positive / negative attribute of reliable behavior: first, the initial adjustment force is matched according to the emotion intensity grading, and then the secondary optimization is carried out by combining the emotion fluctuation level and the behavior attribute, which avoids the problems of excessive reward and punishment for strong emotion or insufficient reward and punishment for weak emotion, and can flexibly adapt to the adjustment rhythm according to the emotion stability difference, so that the reward and punishment strategy is more suitable for the real-time emotion state and behavior attribute of the pet, and the precision of behavior training, the acceptance of the pet and the stability of training effect are significantly improved.

[0074] Although a pet training method is proposed in the traditional scheme, the pet training process needs to be combined with pet feedback, and if only image recognition feedback is used to feed back pet state information, the frequency of the training method design will be lengthened, and the update of the training strategy often needs a very long time to complete, so the traditional method has no advantage compared with artificial training. Therefore, it can be determined that the premise of intelligent training landing is to update the training strategy under a training purpose according to the pet situation quickly and accurately, so that the training strategy can be customized and dynamically adjusted in time according to the pet. The pet behavior training system based on AI vision provided by the present application constructs an adaptive training closed loop capable of quickly responding to pet state changes by integrating multi-source perception data and real-time emotion quantification analysis. The system comprehensively utilizes the accurate positioning and physiological information obtained by the pet locator, the image sequence captured by the visual acquisition module, and the multi-dimensional behavior characteristics and emotion parameters extracted by the data processing module to realize accurate interpretation of the pet behavior intention and emotional state; the strategy matching module generates an initial reward and punishment strategy according to the pet species characteristics and real-time behavior recognition results; the execution feedback module dynamically adjusts the reward and punishment intensity and mode in the strategy execution process based on the emotion intensity and stability parameters. This technical path effectively solves the problem of slow strategy update caused by feedback delay in the traditional scheme, realizes millisecond-level synchronization and adaptive optimization of the training strategy and the pet state change, and truly achieves the effect of personalized, accurate and real-time intelligent training.

[0075] In summary, the above only describes preferred embodiments of the technical scheme of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An AI vision-based pet behavior training system, characterized in that, The pet behavior training system comprises: A pet locator comprising a positioning device and a physiological signal sensor, which is used to collect real-time positioning information and physiological information of the pet, and the pet locator is integrated with a voice prompter; A visual acquisition module comprising a camera installed at a specified position and a smart feeder, which is used to collect image information of the pet; A data processing module, which is used to distinguish different pet individuals according to the positioning information and the image information, and is also used to extract features from the positioning data and the image data to obtain a behavior feature vector, and determine a behavior type and a behavior confidence of the pet behavior based on the behavior feature vector to judge whether the behavior is reliable, and is also used to analyze the physiological information to determine a basic emotional state of the pet, calculate an emotional component of the pet through a preset algorithm, and calculate an emotional intensity parameter and an emotional stability parameter in combination with the basic emotional state; A strategy matching module, which is used to determine a pet species according to the image information, and select and execute a corresponding reward and punishment strategy in combination with the behavior type, the behavior confidence of the pet and a preset reward and punishment rule library, and the reward and punishment strategy is used to train the pet behavior by controlling the voice prompter and the smart feeder; An execution feedback module, which is used to adjust a core intensity of the reward and punishment strategy according to the emotional intensity parameter of the pet during execution of the reward and punishment strategy, and secondarily adjust the adjusted core intensity according to the emotional stability parameter and a preset adjustment rule.

2. The pet behavior training system of claim 1, wherein, Each pet locator is internally provided with a distinguishing label, and the process of distinguishing different pet individuals by the data processing module comprises: Receiving positioning information and corresponding distinguishing labels sent by each pet locator; Performing target detection on the collected image information to extract image features of each pet; Matching a space-time coordinate of the positioning information with a space-time coordinate of the image information, associating the image features with the positioning information in the same space-time range to form associated information; Matching each associated information with a corresponding distinguishing label to distinguish different pet individuals; After distinguishing pet individuals for the first time, a mapping relationship database of the distinguishing labels and corresponding image features and positioning information is established, and the distinguishing labels in the mapping relationship database are matched by real-time collected positioning information and image information to realize the distinction of different pet individuals.

3. The pet behavior training system of claim 2, wherein The feature extraction comprises: For the positioning information and the image information in the same associated information, the positioning information is smoothed to obtain a continuous pet position sequence, and the image information is filtered to obtain clear frames and identify pet target regions; Based on the position sequence, a displacement distance, a turning angle and a stay time in a unit time are calculated to obtain motion features; A distance change sequence of the pet and a key environmental region is counted to obtain space interaction features; Based on the pet target regions in the image information, node coordinates, limb angles and action trajectories are extracted to obtain limb features; The motion features, the space interaction features and the limb features are fused to obtain a behavior feature vector of the pet.

4. The pet behavior training system of claim 1, wherein, The data processing module is configured with a preset behavior template library, the behavior template library comprises typical feature vectors of each behavior type, and the process of determining the behavior type and the behavior confidence comprises: Comparing the behavior feature vector with the typical feature vectors in the behavior template library to calculate a matching degree; The behavior feature vector is predicted by a preset classification model to calculate probability values of each behavior type; The matching degree and the probability values are fused according to a preset weight to obtain a comprehensive score, a behavior type with the highest comprehensive score is selected as a recognition result, and the comprehensive score is standardized to be a behavior confidence of the behavior type.

5. The pet behavior training system of claim 1, wherein, The physiological signal sensor includes a heart rate sensor and an electrodermal sensor, which are respectively used to collect a heart rate signal and an electrodermal signal of the pet to constitute physiological information of the pet, and the analyzed physiological information includes: The collected heart rate signal and electrodermal signal are filtered and denoised, and heart rate variability and electrodermal conductance level are extracted; The heart rate variability and the electrodermal conductance level are compared with corresponding preset normal intervals to obtain a comparison result; Based on the comparison result, a basic emotional state of the pet is determined according to a preset judgment rule.

6. The pet behavior training system of claim 1, wherein, The preset algorithm includes: Based on the motion feature, the physiological information and the behavior recognition result of the pet, the basic emotional state of the pet, an emotional component of the pet is calculated, and the reliability of each emotional component is calculated, the emotional component includes a heart rate related emotional component, an electrodermal waveform feature emotional component and an image behavior emotional component; Based on the reliability of each emotional component, the emotional component is corrected; The three corrected emotional components are weighted according to their reliabilities to constitute an emotional vector, and the emotional vector is normalized; Based on the normalized emotional vector and the preset weight corresponding to each emotional component, an emotional intensity parameter is calculated; The standard deviation of the fluctuation of the three emotional components in a preset time window is calculated, and weighted summation and normalization are performed to obtain an emotional stability parameter.

7. The pet behavior training system of claim 6, wherein, The calculation of each emotional component and the corresponding reliability includes: Based on the motion feature, a motion interference index is constructed, and the heart rate signal is filtered to obtain an emotional related heart rate signal; the heart rate variability feature is extracted from the emotional related heart rate signal, and the heart rate related emotional component is obtained by converting the heart rate variability feature according to the exclusive reference value and the limit value of the species corresponding to the basic emotional state through an association function; the reliability of the heart rate related emotional component is calculated in combination with the heart rate signal quality and the motion interference stripping effect; The waveform feature of the electrodermal signal in a preset time window is extracted and normalized; the normalized waveform feature is weighted and calculated according to the corresponding preset feature weight selected according to the basic emotional state, and the electrodermal waveform feature emotional component is obtained through function processing; the reliability of the electrodermal waveform feature emotional component is calculated in combination with the waveform integrity and the signal quality; Based on the behavior recognition result, it is determined whether the behavior type is an out-of-training behavior; if so, the emotional type and the basic emotional value corresponding to the out-of-training behavior are retrieved from a preset mapping table according to the basic emotional state, the basic emotional value is corrected in combination with the behavior confidence, the behavior duration feature and the image quality to obtain the image behavior emotional component; the reliability of the image behavior emotional component is calculated in combination with the behavior confidence and the emotional mapping certainty.

8. The pet behavior training system of claim 1, wherein, The strategy matching module includes a species classification model, the species classification model is constructed based on labeled sample images of multiple pet species, and the determination of the pet species includes: Extracting a key image frame containing the complete shape of the pet from the collected image information, and performing species feature extraction; Inputting the extracted species features into a preset species classification model, which compares the species features with a preset species feature library and outputs the species label with the highest matching degree as the pet species.

9. The pet behavior training system of claim 1, wherein, The selection process of the reward and punishment strategy includes: The strategy matching module calls the reward and punishment sub-database corresponding to the species from the preset reward and punishment rule database based on the determined pet species, and a mapping relationship table between behavior types and reward and punishment strategies is pre-stored in the sub-database; Retrieving the reward and punishment strategy consistent with the currently determined behavior type from the mapping relationship table as a candidate reward and punishment strategy; According to a preset confidence threshold, comparing the current behavior confidence to determine whether the current behavior is a reliable behavior; If it is a reliable behavior, the candidate reward and punishment strategy is determined as the reward and punishment strategy to be executed; If it is not a reliable behavior, no reward and punishment is performed.

10. The pet behavior training system of claim 9, wherein, The attributes of the behavior type include positive attributes and negative attributes, and the execution feedback module is further configured to determine the behavior attributes corresponding to the behavior type of the pet according to the behavior type of the pet, and the adjustment of the reward and punishment strategy includes: Real-time acquisition of the emotional intensity parameter of the pet during the execution of the reward and punishment strategy; When the pet behavior is determined to be a reliable behavior, the emotional intensity of the pet is graded according to the emotional intensity parameter, and the corresponding initial adjustment strategy is selected according to the emotional intensity level to adjust the core intensity of the reward and punishment strategy; According to the emotional stability parameter, the fluctuation level is divided, and combined with the behavior attributes, the corresponding secondary adjustment strategy is selected to perform secondary adjustment on the reward and punishment strategy.

Images