Pet target detection method and device and camera

Abstract

The invention relates to the technical field of target detection, and discloses a pet target detection method and device and a camera, and the method comprises the steps: carrying out the motion triggering collection and image enhancement preprocessing of a front end region of a feeder, and obtaining an enhanced image frame sequence; performing feature extraction of dynamic receptive field adjustment on the enhanced image frame sequence to obtain pet feature descriptors and position information; behavior time sequence feature analysis is executed, and pet behavior sequence feature vectors are obtained; constructing a state transition diagram according to the pet behavior sequence feature vector, and performing time sequence consistency analysis to obtain a pet state judgment result; power management and decision execution are carried out on the feeder based on the pet state judgment result, feeding control under the low-power-consumption condition is achieved, behavior misjudgment caused by posture fluctuation is effectively avoided, the behavior recognition accuracy is improved, the accurate feeding control problem in a multi-pet family is solved, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a pet target detection method, device and camera. Background Art

[0002] Traditional pet feeders either rely on simple infrared or pressure sensor trigger mechanisms, which are unable to accurately identify the pet's identity and behavioral status; or use high-power visual processing systems. Although this improves recognition accuracy, the battery life is severely insufficient, making it difficult to meet the needs of long-term stable operation.

[0003] Pet feeders face particularly challenging challenges in complex indoor environments. Factors such as low light, varying distances, and interference from multiple pets cause traditional recognition algorithms to perform poorly in low-power environments. This is particularly true in low-light conditions at night, where detection success rates often fall below 50% and false trigger rates exceed 30%. Furthermore, existing detection and decision-making algorithms often struggle to balance power consumption and performance. They either sacrifice detection accuracy to reduce power consumption, or maintain high accuracy but significantly increase energy consumption, making them difficult to operate stably and long-term on battery power. Summary of the Invention

[0004] The present invention provides a pet target detection method, device and camera, which effectively avoids behavior misjudgment caused by posture fluctuations, improves behavior recognition accuracy, solves the problem of precise feeding control in multi-pet households, and enhances user experience.

[0005] In a first aspect, the present invention provides a pet target detection method, the pet target detection method comprising: Perform motion-triggered acquisition and image enhancement preprocessing on the front area of ​​the feeder to obtain an enhanced image frame sequence; Performing feature extraction of the enhanced image frame sequence with dynamic receptive field adjustment to obtain pet feature descriptors and position information; Performing behavior temporal feature analysis based on the pet feature descriptor and position information to obtain a pet behavior sequence feature vector; Constructing a state transition diagram based on the pet behavior sequence feature vector and performing time sequence consistency analysis to obtain a pet state judgment result; The feeder is powered and decision-making is performed based on the pet status judgment result to achieve feeding control under low power consumption conditions.

[0006] In combination with the first aspect, in a first implementation of the first aspect of the present invention, the motion-triggered acquisition and image enhancement preprocessing of the front end area of ​​the feeder to obtain an enhanced image frame sequence includes: The camera's working state is controlled by a motion trigger mechanism. When a moving object is detected in the front area of ​​the feeder, the image acquisition process is activated to obtain raw image data. A dual-resolution acquisition mechanism is used to perform resolution conversion on the original image data to obtain a stable imaging frame sequence; performing image noise suppression and color space conversion on the stabilized imaging frame to obtain a single-channel compressed image, and performing edge-preserving blur processing on the single-channel compressed image to obtain a preprocessed image frame sequence; The preprocessed image frame sequence is input into a lightweight residual context attention network for quality enhancement processing to obtain an enhanced image frame sequence.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present invention, inputting the preprocessed image frame sequence into a lightweight residual context attention network for quality enhancement processing to obtain an enhanced image frame sequence includes: Inputting the preprocessed image frame sequence into the backbone feature extraction layer of the lightweight residual context attention network for convolution processing to obtain a basic feature map; Inputting the basic feature map into the context attention enhancement layer of the lightweight residual context attention network for parallel processing to obtain a multi-scale feature map; Calculating the brightness variance of the image region based on the multi-scale feature map, performing quantization grading based on the brightness variance of the image region and selecting a corresponding enhancement parameter group to obtain a lighting parameterized feature map; Executing a cross-layer feature selection strategy on the illumination parameterized feature map, and screening feature layer nodes by setting an activation threshold to obtain a target feature map; The target feature map is input into the texture detail recovery layer of the lightweight residual context attention network for feature upsampling and fusion mapping, and feature upsampling is performed through the deconvolution layer and feature fusion is performed through the 1×1 convolution layer to obtain an enhanced image frame sequence.

[0008] In combination with the first aspect, in a third implementation of the first aspect of the present invention, the feature extraction of the enhanced image frame sequence undergoing dynamic receptive field adjustment to obtain a pet feature descriptor and location information includes: Inputting the enhanced image frame sequence into a dynamic receptive field adjustment module, capturing detail features through a fixed small receptive field branch in the dynamic receptive field adjustment module, capturing medium-scale features through an adaptive medium receptive field branch in the dynamic receptive field adjustment module, and capturing overall contour features through an expanded large receptive field branch in the dynamic receptive field adjustment module, thereby obtaining an initial detail feature map, a medium-scale feature map, and an overall contour feature map; Performing dynamic feature weighting on the initial detail feature map, the medium-scale feature map, and the overall contour feature map to obtain a weighted detail feature map, a weighted scale feature map, and a weighted contour feature map; Performing matching analysis based on the weighted detail feature map, the weighted scale feature map, and the weighted contour feature map with a priori knowledge base of pet features to obtain pet target part positioning information; Performing channel compression and multi-scale feature fusion on the weighted detail feature map, the weighted scale feature map, and the weighted contour feature map to obtain a fused feature map; Based on the fused feature map, the bounding box coordinates and the center point confidence are extracted simultaneously to obtain the pet feature descriptor and position information.

[0009] In combination with the first aspect, in a fourth implementation of the first aspect of the present invention, performing behavior temporal feature analysis based on the pet feature descriptor and position information to obtain a pet behavior sequence feature vector includes: Inputting the pet feature descriptor and position information into the temporal feature extraction unit in the pet behavior recognition module, and performing temporal integration of continuous frame features through the temporal feature extraction unit to obtain pet posture and position change trend data; Inputting the pet posture and position change trend data into the behavior classification unit in the pet behavior recognition module for multi-layer decision analysis to obtain initial behavior type data, wherein the multi-layer decision analysis includes a first-layer classification based on the pet's position change speed, a second-layer classification based on the relative position of the pet and the feeder, and a third-layer classification based on the pet's head posture; The minimum continuous frame number thresholds of the approach, contact and eating behaviors are set according to the initial behavior type data to perform behavior confirmation filtering to obtain a pet behavior sequence feature vector.

[0010] In combination with the first aspect, in a fifth implementation of the first aspect of the present invention, constructing a state transition diagram based on the pet behavior sequence feature vector and performing a temporal consistency analysis to obtain a pet state judgment result includes: Extracting behavior type identifiers from the pet behavior sequence feature vector as nodes of a state transition diagram, forming connecting edges by calculating the transition frequencies between adjacent behavior types, and determining node weights and edge weights using behavior duration and intensity parameters to obtain a state transition diagram; Calculate the node connection pattern, transition probability distribution, and state similarity distribution of the hunger state, curiosity state, and satisfaction state based on the state transition graph; Combining the historical state sequence and the state similarity distribution, comparing the matching degree between the state transition trend and the typical state transition law, and obtaining the verified state data; The verified status data is comprehensively analyzed with the pet's individual behavior pattern library and environmental parameters to obtain a pet status judgment result.

[0011] In combination with the first aspect, in a sixth implementation of the first aspect of the present invention, calculating the node connection pattern, the transition probability distribution, and the state similarity distribution of the hunger state, the curiosity state, and the satisfaction state based on the state transition graph includes: Obtaining a node connection topology in the state transition graph, and calculating the connection frequency and directionality between the behavior nodes of approach, contact, exploration, wandering, eating, and leaving based on the node connection topology to obtain a node connection pattern feature value; Calculating behavior transition probability distribution data between nodes in the state transition graph based on the node connection pattern characteristic value; Based on the behavior transition probability distribution data and the standard pattern of hunger state, the degree of matching of high-frequency approach-contact-probing behavior cycles, long-term contact duration, and high-intensity probing behavior characteristics is calculated to obtain a hunger state similarity value; Based on the behavior transition probability distribution data and the curiosity state standard model, the matching degree of low-frequency approach-contact-leave behavior cycle, short-term contact duration, and frequent behavior transition characteristics is calculated to obtain the curiosity state similarity value; Based on the behavior transition probability distribution data and the standard pattern of the satisfied state, the matching degree of the single approach-contact-eating-leaving complete sequence, the long-term feeding duration, and the no-approach feature after leaving are calculated to obtain the satisfied state similarity value; The hunger state similarity value, the curiosity state similarity value, and the satisfaction state similarity value are combined to form a state similarity distribution.

[0012] In combination with the first aspect, in a seventh implementation of the first aspect of the present invention, the power management and decision execution of the feeder based on the pet status judgment result to achieve feeding control under low power consumption conditions include: Inputting the pet status judgment result into the non-communication adaptive power control system, monitoring battery parameters through the battery management unit in the non-communication adaptive power control system and dynamically allocating system power based on the pet status to obtain a power allocation strategy; performing power rail switching on the feeder according to the power allocation strategy to obtain an optimal power supply mode; Based on the optimal power supply mode and the pet state judgment result, segmented pulse power supply control is implemented to control the energy distribution of the feeder during the feeding process to obtain a feeding execution signal; Adaptive power consumption compensation control is performed on the feeding execution signal to obtain an energy-balanced execution control instruction, and a power-function mapping table is established to automatically adjust the function priority at different battery power stages to achieve feeding control under low power consumption conditions.

[0013] In a second aspect, the present invention provides a pet target detection device, comprising: An acquisition module is used to perform motion-triggered acquisition and image enhancement preprocessing on the front area of ​​the feeder to obtain an enhanced image frame sequence; A feature extraction module, configured to extract features of the enhanced image frame sequence by dynamically adjusting the receptive field to obtain pet feature descriptors and position information; A feature analysis module, configured to perform behavior temporal feature analysis based on the pet feature descriptor and position information to obtain a pet behavior sequence feature vector; A consistency analysis module is used to construct a state transition diagram based on the pet behavior sequence feature vector and perform time sequence consistency analysis to obtain a pet state judgment result; The decision execution module is used to perform power management and decision execution on the feeder based on the pet status judgment result, so as to realize feeding control under low power consumption conditions.

[0014] A third aspect of the present invention provides a camera, which is used to execute the above-mentioned pet target detection method.

[0015] The technical solution provided by this invention significantly reduces system standby power consumption by employing motion-triggered and dual-resolution acquisition mechanisms. This allows the feeder to maintain ultra-low power consumption when no pet is approaching, while rapidly activating the full processing flow upon pet approaching, achieving a dynamic balance between power consumption and performance. A lightweight residual contextual attention network efficiently runs on low-power processors. Through a spatial perception enhancement mechanism and a cross-layer feature selection strategy, it reduces computational effort by over 90% while maintaining image enhancement quality, providing high-quality input for subsequent pet detection. The dynamic receptive field adjustment module uses three parallel feature extraction branches to capture comprehensive pet features, from details to overall outlines. Combined with a dynamic feature weighting mechanism, it enhances attention to areas where the pet is moving, improving detection accuracy. The pet behavior recognition module and temporal feature analysis mechanism accurately distinguish different pet behavior patterns. Through multi-layer decision analysis and temporal consistency constraints, they effectively avoid misjudgments caused by posture fluctuations, improving behavior recognition accuracy. The state transition diagram and temporal consistency analysis method enable precise determination of the pet's state. By matching typical behavior patterns and accounting for individual pet differences, this significantly improves state determination accuracy and reduces the possibility of false triggering of feeding. The non-communication adaptive power control system dynamically adjusts system power consumption based on the pet's status. Combined with a segmented pulse power supply control strategy and an adaptive power compensation controller, this significantly extends battery life while ensuring the normal operation of core functions. This invention supports intelligent identification and personalized feeding management in multi-pet environments. By establishing a pet ID-feature mapping table and a hierarchical feeding control strategy, it solves the problem of precise feeding control in multi-pet households and improves the user experience.

[0016] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of an embodiment of a pet target detection method according to an embodiment of the present invention; Figure 2 FIG. 1 is a schematic diagram of an embodiment of a pet target detection device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] The terms "including," "having," and any variations thereof, as used in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to the process, method, product, or device.

[0021] To facilitate understanding of this embodiment, a pet target detection method disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 As shown, this method includes the following steps: Step 101: Perform motion-triggered acquisition and image enhancement preprocessing on the front area of ​​the feeder to obtain an enhanced image frame sequence; It is understandable that the execution subject of the present invention may be a pet target detection device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0022] Specifically, based on a highly energy-efficient perception mechanism, the system continuously monitors weak signals from the front area through an embedded low-power microcontroller. Upon detecting the presence of motion signatures of a moving object in the area in front of the feeder, the camera acquisition module is activated, switching from a static or low-frequency standby state to a high-efficiency image acquisition mode to obtain the current raw image data. A dual-resolution acquisition mechanism is employed, with a low-resolution 160×120 image frame for rapid initial judgment. Once the outline of a moving target emerges and it is suspected to be a pet, acquisition continues at a working resolution of 384×288. This resolution switching effectively improves the stability of the imaging frame and reduces the data bandwidth pressure in the image processing chain, resulting in a stable imaging frame sequence. Image noise suppression and color space conversion are performed on the stabilized imaging frames. Preliminary image noise reduction is achieved through a hardware-level 4×4 regional mean filter, effectively suppressing high-frequency noise introduced by electrical interference in the photosensitive device while maintaining the integrity of the image's main contour information. The noise-suppressed image is then converted to a color space using a linear mapping from RGB to YCbCr space, retaining the Y luminance channel as the primary output channel. This approach reduces over 66% of the image's redundant color data, preserving the brightness and shading of the image information while minimizing computational overhead when transmitted via a single channel. This single-channel compressed image is then fed into an edge-preserving blur processing module, employing a guided filtering algorithm based on a 5×5 window. This algorithm suppresses error noise introduced by small-scale textures while preserving the object's contours, edges, and structure through a guided graph mechanism. This achieves a balanced optimization between compression, noise reduction, and structure preservation, resulting in a sequence of preprocessed image frames. The preprocessed image frame sequence is input into a lightweight residual contextual attention network for quality enhancement. The network adopts a three-level convolutional architecture. First, two layers of shallow 3×3 convolutions are used to extract basic texture features. Then, a contextual attention residual module containing multi-scale convolution kernels of 1×1, 3×3, and 5×5 performs cross-scale information fusion. The channel attention mechanism is used to achieve adaptive enhancement of key image regions. Subsequently, a layer of deconvolution and a 1×1 convolution are used to restore the structure of texture details and fuse features, forming an enhanced image frame sequence with high contrast and high contour, which is suitable for subsequent feature extraction. At the same time, to achieve low-energy deployment goals, the network integrates cross-layer activation screening and 8-bit quantization strategies. Only key nodes with activation strength greater than a set threshold are retained, greatly reducing computational and storage pressure. While ensuring image enhancement quality, the overall computational overhead of the model is compressed to less than 5.1mJ per frame. This achieves stable acquisition, efficient compression, and deep enhancement of image data on resource-constrained embedded devices, constructing a complete front-end image preprocessing pipeline that spans perception, compression, and enhancement.

[0023] Step 102: Perform feature extraction of the enhanced image frame sequence by dynamic receptive field adjustment to obtain pet feature descriptors and position information; Specifically, the enhanced image frame sequence is input into a three-branch dynamic receptive field adjustment module. This module concurrently executes a fixed small receptive field branch, an adaptive medium receptive field branch, and an extended large receptive field branch to jointly capture detail features, local structural features, and overall contour features. The small receptive field branch uses two consecutive 3×3 standard convolutional layers to extract fine texture features such as hair edges, ear tips, and claws. The medium receptive field branch uses a 5×5 depthwise separable convolutional structure to capture medium-scale regional features such as the head and back contours. The large receptive field branch combines two 3×3 convolutional layers with a 2×2 max pooling operation to capture overall contour morphology and posture information at a global scale. These three feature maps are then output to form the initial detail feature map, the medium-scale feature map, and the overall contour feature map, respectively, while preserving their corresponding scale information. A structural change attention mechanism is introduced in the temporal dimension, and dynamic feature weighting is performed on the three feature maps mentioned above. For each feature map, a structural similarity index is calculated between the current frame and the past five frames. When a region's structure remains stable across multiple frames, its motion signature is considered weak. When the similarity exceeds 0.85, the weight of that region is dynamically reduced to 0.4. Regions with similarities below 0.45 are considered highly dynamic, and their weights are increased to 1.6. This emphasizes the response to the pet's motion, generating weighted detail feature maps, weighted scale feature maps, and weighted contour feature maps. After obtaining multi-scale features with dynamic responsiveness, a priori knowledge base of pet features is introduced. This knowledge base contains three structural description vectors for 32 common pets: head outline, body proportions, and fur texture. Each descriptor has a dimension of 128. These descriptors are used to perform local similarity matching analysis with the weighted feature maps, rapidly locating areas in the image suspected to belong to key pet parts, such as the head, back, and limbs, to obtain location information for the pet's target. Channel compression and multi-scale feature fusion operations are performed on the weighted detail feature map, weighted scale feature map, and weighted contour feature map. 1×1 convolution is applied to each of the three feature maps to achieve uniform channel compression. These are then fused using a weighted summation strategy. The weight coefficients are dynamically adjusted via hardware registers and are set to [0.25, 0.5, 0.25] by default to ensure relatively higher attention is paid to mid-scale backbone features. This generates a fused feature map. Target localization and confidence estimation are performed simultaneously based on the fused feature map. A set of weighted convolutional structures is used to predict the coordinate parameters (x, y, w, h) of the bounding box. A Sigmoid function is used to output the confidence score of the center point. This combined result forms an object detection result that contains both location and confidence information. A 128-dimensional feature vector for the corresponding region is extracted from the fused feature map and output as the pet feature descriptor.

[0024] Step 103: Perform behavior temporal feature analysis based on the pet feature descriptor and the position information to obtain a pet behavior sequence feature vector; Specifically, the pet feature descriptor and its corresponding position information are input into the temporal feature extraction unit in the pet behavior recognition module. This unit uses a sliding window mechanism to integrate and process the time-continuous feature sequence. The window length is set to 12 frames and the step size is 3 frames. By aggregating the information of the current frame and its adjacent historical frames in the time domain, the pet's posture evolution trend and position change trajectory over a short period of time are effectively extracted, including dynamic attributes such as position change speed, head orientation angle, and bounding box movement path, and high-dimensional time series feature data representing the pet's behavior change trend is formed. The above time series features are input into the behavior classification unit in the behavior recognition module for multi-layer decision analysis. This unit uses a lightweight tree-structured hierarchical modeling method to complete the process from coarse classification to fine recognition in a three-layer structure. In the first level of classification, the pet's relative position change speed V is calculated based on the spatial displacement of the feature points in consecutive frames, and the current behavior is preliminarily divided into three categories: stationary (V less than 0.05), wandering (V between 0.05 and 0.2), and approaching (V greater than or equal to 0.2), providing a speed semantic basis for subsequent behavioral intention analysis; in the second level of classification, the relative distance D between the pet and the feeder is analyzed, and the Euclidean distance between the center of the bounding box and the feeder position is evaluated and normalized, thereby subdividing the behavior into away The system then classifies the pet's behavior into three micro-behaviors: approach (D greater than or equal to 0.8), alert (D between 0.4 and 0.8), and contact (D less than 0.4). The system then enters the third level of classification, focusing on the pet's head posture angle θ and mouth state S. θ represents the tilt of the head toward the camera or feeding port, and S indicates whether the pet is in an open or feeding state. This further categorizes the behavior into three micro-behaviors: observation (θ greater than 30 degrees and S equal to 0), exploration (θ between 0 and 30 degrees and S equal to 0), and feeding (θ less than or equal to 10 degrees and S equal to 1). After completing the three-level classification, initial behavior type data is obtained. However, given that pets often experience brief posture fluctuations or interruptions in real-world scenarios, a behavior confirmation filtering mechanism is introduced. This mechanism verifies the stability of the behavior by setting a minimum frame count threshold for the behavior duration. "Approach" behavior is considered valid only if consistent judgments appear in three consecutive frames. "Contact" behavior requires consistent judgments for at least five frames. "Feeding" behavior is considered valid only if the behavior remains stable for at least eight consecutive frames. Outputs a formatted pet behavior sequence feature vector, which contains the confirmed behavior type code, the duration of the behavior (in frames), and a level indicator representing the behavior intensity. The behavior intensity is comprehensively evaluated with reference to posture stability, displacement speed, and frequency and quantified on a scale of 1 to 5.

[0025] Step 104: construct a state transition diagram based on the pet behavior sequence feature vector and perform time sequence consistency analysis to obtain a pet state judgment result; Specifically, the method uses the pet behavior sequence feature vector as the core input and behavior type identifiers as the basis for graph construction, creating a structured representation of behavioral states that occur continuously over time. Each identified behavior type is represented as a node in the state transition graph. The frequency of transitions between adjacent behavior types is calculated based on the order of behaviors in the behavior sequence, and this frequency serves as the basis for constructing edges connecting the nodes. Each behavior node is assigned a weight, determined by both the duration and intensity of the behavior. Nodes with longer durations and higher intensity have greater weights in the graph, reflecting their dominance over the current behavior pattern. Edge weights are determined by the frequency of behavior transitions between two nodes. Edges with high transition frequencies indicate frequent switching within the behavior sequence and a higher probability of state transitions, forming a dynamic behavioral state graph. After the state transition graph is constructed, graph structure analysis and state cluster matching are performed. By calculating connectivity metrics such as the out-degree and in-degree distributions, path lengths, and loop structures for each node in the graph, and combining them with edge weights, the transition probability distribution function between states is derived, forming a set of quantized state transition parameter vectors. Based on these transition feature parameters, the system uses built-in templates for typical pet behavior states to compare the current transition graph for similarity. These include typical behavior chain patterns such as the approach-contact-exploration cycle in the hungry state, the single meal followed by departure in the satisfied state, and the frequent but brief approach and departure in the curious state. By calculating the similarity distribution between the current transition graph and these three typical state transition structures, the system determines the degree of fit of the current state graph to each typical state, thereby completing preliminary state category identification. This process not only considers the degree of structural match but also incorporates node duration and behavior intensity to improve the temporal sensitivity of state matching and the accuracy of judgment. After calculating the state similarity distribution, the system integrates historical state sequences to verify the stability of the judgment results. A sliding window mechanism is used to extract state judgment records from the previous period and compare them with the current predicted state. If the current state's transition trend is highly consistent with the evolution of historical states—for example, if there has been continuous exploration behavior accompanied by a decrease in feeding behavior—then the prediction is more likely to be a hungry state. If the behavior pattern is short-lived and changes frequently, it may be a curious or alert state. This process verifies the state through smoothness judgment of transition trends and a time series reconstruction algorithm, effectively eliminating misjudgments caused by short-term abnormal fluctuations. The verified state data is input into the pet individual behavior pattern library and the environmental parameter adjustment module for comprehensive analysis and processing. The pet behavior pattern library is constructed based on each pet's unique ID, recording the transition probability, duration threshold, and intensity distribution between each individual's behaviors under normal conditions, and automatically fine-tuning them over time. At the same time, the system integrates auxiliary parameters such as ambient light intensity, indoor temperature, and background noise, and uses an environment-behavior deviation mapping function to correct the state judgment results, such as allowing certain behaviors to be delayed or the frequency to decrease in high temperature or low light environments.Through the behavior-individual-environment fusion mechanism, the final output is the pet state judgment result including state type coding, judgment confidence and state duration.

[0026] Step 105: Power management and decision execution are performed on the feeder based on the pet status judgment result to achieve feeding control under low power consumption conditions.

[0027] Specifically, the pet's status judgment results are input into a non-communication adaptive power control system, which consists of a battery management unit, a power allocation scheduler, and an execution controller. The battery management unit continuously monitors the current battery voltage, current, and temperature parameters. When it detects that the voltage is lower than 3.5V or the temperature exceeds 45°C, it activates the protection mechanism to enter low-power or thermal avoidance mode. At the same time, combined with the pet's current status, such as "hungry," "satisfied," or "curious," the system dynamically allocates power resources to various functional modules. For example, in the hungry state, the operation of the feeding motor and sensor module is prioritized, and the power share of the decision-making and execution levels is increased to more than 65% of the total power consumption. In the satisfied or curious state, the execution level power is reduced, and the detection level remains the main active module, thus forming a refined power allocation strategy. The power allocation strategy controls the switching mechanism of the power rails and adopts a dual-power rail architecture. That is, when an external power supply is connected, the high-performance power rail is enabled to transfer computing tasks to the processing unit with higher energy efficiency. When running only on battery power, the low-power power rail is preferentially enabled, and the load balancing algorithm of the local microcontroller unit is activated. By postponing the execution of non-critical tasks, reducing the image acquisition frequency, and suspending some high-energy-consuming computing units, energy consumption is compressed to achieve optimal switching of power supply modes. The switching process does not exceed 5ms and has the ability to resume execution from breakpoints to ensure functional continuity. After the power supply mode is determined, the segmented pulse power supply control mechanism is triggered jointly based on the mode and the pet's status judgment results. This mechanism divides the feeding execution action of the feeder into multiple stages and assigns specific energy pulse parameters to each stage. For example, the first stage is the motor pre-charging stage, which lasts 50ms and the current is controlled at 0.1A; the second stage is the main drive stage, which lasts 200ms and the current is increased to 0.3A to complete the main mechanical pushing action; the third stage is the recovery stage, which lasts 150ms and recovers and redistributes the remaining kinetic energy through the energy feedback module, with an energy recovery rate of up to 42%. This segmented pulse design optimizes the efficiency of unit energy consumption and achieves fine adjustment of the feeding action through pulse width precision control, reducing the phenomenon of misfeeding and reducing system jitter caused by excessive load. Adaptive power consumption compensation control is implemented on the above-mentioned feeding execution signal, that is, the execution method of the final control instruction is dynamically adjusted according to the pet's status level, feeding action intensity and current power load. For example, in a hungry state with a confidence level higher than 85%, the system's instantaneous power consumption is allowed to be increased to 120% of the nominal power consumption to ensure rapid response capability; when the status confidence level is low or the behavior is unclear, the feeding instruction is not executed immediately. Instead, the image sampling density is increased and the judgment window is extended to ensure decision accuracy.In addition, a power-function mapping table is established to correspond different battery power levels to a set of function activation priorities. For example, when the power level is lower than 30%, the image enhancement and multi-scale analysis modules are automatically turned off, retaining only basic behavior recognition and scheduled feeding functions; when the power level is lower than 10%, it enters extreme power saving mode, retaining only Bluetooth or remote wake-up mechanism and simple scheduled execution, thereby achieving optimal allocation of power resources and gradual degradation of control paths, and realizing feeding control under low power conditions.

[0028] In this embodiment of the present invention, the use of motion-triggered and dual-resolution acquisition mechanisms significantly reduces system standby power consumption, enabling the feeder to maintain an ultra-low power state when no pet is approaching. Meanwhile, the full processing flow is quickly activated upon the approach of a pet, achieving a dynamic balance between power consumption and performance. A lightweight residual contextual attention network efficiently runs on low-power processors. Through a spatial perception enhancement mechanism and a cross-layer feature selection strategy, it reduces computational effort by over 90% while maintaining image enhancement quality, providing high-quality input for subsequent pet detection. The dynamic receptive field adjustment module uses three parallel feature extraction branches to capture comprehensive features of the pet, from details to overall outlines. Combined with a dynamic feature weighting mechanism, it enhances attention to areas where the pet is moving, improving detection accuracy. The pet behavior recognition module and temporal feature analysis mechanism accurately distinguish different pet behavior patterns. Through multi-layer decision analysis and temporal consistency constraints, they effectively avoid misjudgments caused by posture fluctuations, improving behavior recognition accuracy. The state transition diagram and temporal consistency analysis method enable precise determination of the pet's state. By matching typical behavior patterns and accounting for individual pet differences, the accuracy of state determination is significantly improved, reducing the possibility of false triggering of feeding. The non-communication adaptive power control system dynamically adjusts system power consumption based on the pet's status. Combined with a segmented pulse power supply control strategy and an adaptive power compensation controller, this significantly extends battery life while ensuring the normal operation of core functions. This invention supports intelligent identification and personalized feeding management in multi-pet environments. By establishing a pet ID-feature mapping table and a hierarchical feeding control strategy, it solves the problem of precise feeding control in multi-pet households and improves the user experience.

[0029] In a specific embodiment, the process of executing step 101 may specifically include the following steps: The camera's working state is controlled by a motion trigger mechanism. When a moving object is detected in the front area of ​​the feeder, the image acquisition process is activated to obtain raw image data. A dual-resolution acquisition mechanism is used to convert the resolution of the original image data to obtain a stable imaging frame sequence; Performing image noise suppression and color space conversion on the stabilized imaging frame to obtain a single-channel compressed image, and performing edge-preserving blur processing on the single-channel compressed image to obtain a preprocessed image frame sequence; The preprocessed image frame sequence is input into the lightweight residual context attention network for quality enhancement to obtain the enhanced image frame sequence.

[0030] Specifically, a perception activation mechanism is constructed with low power consumption and responsiveness as its core goals. This mechanism is based on a motion trigger device located in the front area of ​​the pet feeder, implemented by an ultra-low-power infrared pyroelectric sensor or CMOS frame differential detection logic. Its essence is to continuously monitor the front-end monitoring field of view at a low frequency or in a static state. When a significant change in brightness difference is detected or contour disturbances appear in consecutive frames exceeding a preset threshold, it is determined to be a moving object, triggering the camera to wake up from standby or hibernation and start the image acquisition process. At this point, the camera enters full-function acquisition mode, performing high-speed image capture of the target area to obtain raw image data, and caching it in the local frame buffer for subsequent processing. At the same time, since this mechanism only activates the image path when a target enters, it generates almost no image power consumption when no pet is approaching. After the camera is activated and successfully captures the raw image, it enters the adaptive resolution processing phase. To balance imaging efficiency and image quality, a dual-resolution acquisition mechanism downsamples and performs hierarchical response processing on the raw image data. Initially, image acquisition defaults to a low-resolution configuration of 160×120, enabling rapid background modeling, moving target location, and rough boundary contour extraction. If closed edge structures or concentrated grayscale areas are detected in the image, indicating the presence of a pet, the system switches to a medium- to high-resolution mode of 384×288 to resample the image, capturing clearer details such as outlines and fur textures. This mechanism rapidly determines target presence at low resolution and supplements target details at high resolution, achieving a balanced balance between power consumption and imaging quality without incurring the overhead of continuous high frame rates. Furthermore, the system's underlying driver integrates image frame timestamps and dynamic ROI control to prioritize sampling and cache updates of target areas, outputting an image sequence with high image stability and sharp object definition. Image denoising and compression are performed on this stable image sequence, minimizing data volume and redundant information for subsequent computations while preserving contour, edge, and structural information. The image noise reduction component utilizes a hardware-optimized 4×4 regional mean filter, performing local window smoothing directly at the sensor data stream output. This effectively suppresses image speckle and pseudo-contours caused by acquisition device thermal noise, electromagnetic interference, and edge pixel transitions, while also preventing over-smoothing of edge grayscale gradients. After completing the initial noise reduction, the image is converted from RGB to YCbCr color space, retaining only the Y luminance channel as the primary processing data, as the Y channel contains the vast majority of structural information in low-light environments and texture recognition tasks. The image data is also compressed to one-third its size, significantly reducing data path bandwidth and subsequent computational resource consumption.This single-channel image data is then subjected to edge-preserving blurring, implemented using a filtering scheme based on a 5×5 guided filter. This filter preserves the object's outline and primary boundary structure while suppressing small-scale noise and texture artifacts, achieving structured sparsification of image details. The resulting output image is visually cleaner and more coherent, with enhanced structural consistency and computational compression. This sequence of preprocessed image frames is then transferred via DMA to the input buffer of the deep processing module. This preprocessed image frame sequence is then fed into a lightweight residual contextual attention network for quality enhancement. This enhancement network, specifically designed for low-power embedded platforms, consists of two shallow 3×3 convolutional layers, which initially extract texture edge features and generate intermediate feature maps. This network then enters a contextual enhancement structure consisting of three parallel paths. These structures use 1×1, 3×3, and 5×5 convolutional kernels to extract contextual information from different receptive fields. A channel-wise attention mechanism is then used to adjust feature importance, focusing on the main area of ​​the pet. At the same time, the illumination level of the current frame is estimated based on the brightness variance of the image region, and an appropriate set of enhancement parameters is assigned. For example, a gamma enhancement function with γ = 0.6 is applied to darker areas of the brightness distribution, and adaptive histogram equalization is used for overexposed areas. This ensures that the image exhibits clear textures and recognizable boundaries in all-light environments. After completing contextual enhancement, the network enters the texture detail recovery stage. A layer of inverse convolution is used to upsample the feature map to the original size. A 1×1 convolution is then used to fuse multiple channels to output a highly structurally consistent grayscale enhanced image. The resulting enhanced image frame sequence features rich texture, sharp edges, and distinct light-dark contrast, while preserving the original structural information.

[0031] In a specific embodiment, the step of inputting the pre-processed image frame sequence into the lightweight residual context attention network for quality enhancement to obtain the enhanced image frame sequence may specifically include the following steps: The preprocessed image frame sequence is input into the backbone feature extraction layer of the lightweight residual context attention network for convolution processing to obtain the basic feature map; The basic feature map is input into the context attention enhancement layer of the lightweight residual context attention network for parallel processing to obtain a multi-scale feature map; Calculate the brightness variance of the image region based on the multi-scale feature map, perform quantization and grading based on the brightness variance of the image region and select the corresponding enhancement parameter group to obtain the illumination parameterized feature map; A cross-layer feature selection strategy is implemented on the illumination parameterized feature map, and the feature layer nodes are screened by setting the activation threshold to obtain the target feature map; The target feature map is input into the texture detail recovery layer of the lightweight residual context attention network for feature upsampling and fusion mapping, and is then subjected to feature upsampling by the deconvolution layer and feature fusion by the 1×1 convolution layer to obtain an enhanced image frame sequence.

[0032] Specifically, the preprocessed image frame sequence is fed into the backbone of a lightweight residual contextual attention network (LCAN). This backbone feature extraction layer consists of two consecutive standard 3×3 convolutional layers, each with 8 and 16 kernel channels, respectively. The convolution operation uses a stride of 1 and an edge padding of 1. A Reluctant Unit (ReLU) activation function is introduced after each convolution layer to enhance the nonlinear feature response. In this architecture, the first convolution layer primarily extracts low-level texture and edge variation features of the image, capturing primary information such as local contours, grayscale gradients, and detail transitions. The second convolution layer further integrates neighborhood information based on the previous layer's feature map and generates a more abstract intermediate feature representation, thereby constructing a basic feature map with sufficient semantic expression and context recognition capabilities. This basic feature map is then fed into the Contextual Attention Enhancement Layer, the core module of the LCAN. This module consists of three parallel residual substructures, each employing convolutional kernels of different sizes to extract feature responses within different receptive fields. The first path uses a 1×1 convolution kernel to extract high-precision local features, suitable for processing high-frequency components such as edges and texture details. The second path employs a standard 3×3 convolution kernel to capture medium-range spatial relationships, such as the shape distribution and boundary structure of local regions. The third path introduces a 5×5 dilated convolution kernel, integrating broader contextual semantic information through a larger receptive field. This allows for stable regional representation in images despite occlusion, low light conditions, and structural ambiguity. Within each path, a residual structure uses bypass connections to sum input and output features, maintaining information flow consistency and preventing gradient vanishing or feature over-compression caused by deeper network layers. The channel-wise attention mechanism adaptively reweights each branch output. Through global average pooling, FC layer compression-expansion, and sigmoid mapping, it generates a channel weight vector, which is applied to the feature map channel dimension to emphasize responses to salient regions. The feature maps output by the three paths are concatenated and fused in the channel dimension to form a multi-scale feature map that captures multi-scale context. To address the impact of varying lighting environments on the visibility of image structures, the brightness variance of image regions is calculated based on this multi-scale feature map. A sliding window is used to statistically analyze brightness differences within each image region within the spatial domain, forming a brightness fluctuation matrix. This matrix is ​​then discretely quantized into five brightness level intervals, corresponding to typical lighting conditions: extremely dark, dark, neutral, bright, and overexposed. Based on a set of preset image enhancement parameters for each level mapping, the image dynamically selects the enhancement method that best suits the brightness distribution of the current region. For low-brightness regions, a gamma nonlinear enhancement function with γ = 0.6 is used to boost local grayscale values ​​and enhance visual visibility. For high-brightness regions, contrast-limited histogram equalization is used to suppress local overexposure while enhancing the sense of boundary depth.Enhancement parameters are applied at the per-pixel or per-region level, resulting in an illumination parameterized feature map. This enhanced map exhibits balanced structural clarity and contour separability across different illumination regions. To reduce redundant computation and increase feature information density, a cross-layer feature selection strategy is applied to the illumination parameterized feature map. This strategy selects nodes based on feature activation strength. Specifically, the activation value (i.e., the post-convolution response) is calculated for each spatial location within each feature map channel, and a global activation threshold T = 0.15 is set as the decision criterion. If the activation value at a location exceeds the threshold, the feature response at that location is retained for the next processing stage; otherwise, it is zeroed and considered an invalid node. This significantly reduces the number of floating-point operations required for subsequent network computation and avoids false enhancement or feature drift caused by the participation of low-energy regions. A sparse target feature map is generated. This target feature map is then fed into the texture detail recovery layer of a lightweight residual contextual attention network to restore the image's structure and enhance the output. The feature map is upsampled via deconvolution, restoring its spatial size from its compressed state to the original size. Figure 1 The proposed method achieves consistent resolution while retaining detailed edges and texture transition positions through learnable convolution kernel weights. A 1×1 convolution layer is then introduced to fuse and map the information between different channels to unify the expression style and response frequency of each path, avoiding structural distortion or image distortion caused by uneven distribution of information between channels, and ultimately outputting an enhanced image frame sequence.

[0033] In a specific embodiment, the process of executing step 102 may specifically include the following steps: The enhanced image frame sequence is input into the dynamic receptive field adjustment module, and the fixed small receptive field branch in the dynamic receptive field adjustment module is used to capture detail features, the adaptive medium receptive field branch in the dynamic receptive field adjustment module is used to capture medium-scale features, and the extended large receptive field branch in the dynamic receptive field adjustment module is used to capture overall contour features, thereby obtaining an initial detail feature map, a medium-scale feature map, and an overall contour feature map; Performing dynamic feature weighting on the initial detail feature map, the medium-scale feature map, and the overall contour feature map to obtain a weighted detail feature map, a weighted scale feature map, and a weighted contour feature map; Based on the weighted detail feature map, weighted scale feature map and weighted contour feature map and the pet feature prior knowledge base, the positioning information of the pet target part is obtained; Perform channel compression and multi-scale feature fusion on the weighted detail feature map, weighted scale feature map and weighted contour feature map to obtain a fused feature map; Based on the fused feature map, the bounding box coordinates and center point confidence are extracted simultaneously to obtain the pet feature descriptor and location information.

[0034] Specifically, the enhanced image frame is input into the triple-branch structure of the dynamic receptive field adjustment module, where the fixed small receptive field branch is mainly composed of two consecutive 3×3 standard convolutional layers, which is used to focus on the detailed elements in the image, such as the pet's hair texture, ear tip contour, claw shape and other high-frequency information. These features are densely distributed in space but have a small structural span, so the accuracy is maintained through a smaller receptive field; at the same time, the adaptive medium receptive field branch adopts a 5×5 depthwise separable convolution structure, which has the advantages of large receptive field coverage and low parameter count. It is suitable for extracting medium-scale body information such as head contour, back shape, and tail curvature, and has better regional aggregation characteristics; the extended large receptive field branch integrates two serially connected 3×3 convolutional layers and a 2×2 maximum pooling operation to capture low-frequency structural features such as overall contour, body distribution and movement direction. Especially when the pet's entire body is in the center of the picture or partially occluded, this path can integrate contextual information and make up for the missing boundaries. The three branches output an initial detail feature map, a mid-scale feature map, and an overall contour feature map, respectively, which together constitute a set of feature representations across multiple perceptual dimensions. A dynamic feature weighting mechanism is introduced to enhance the ability to focus on the response to target regions. This mechanism uses a sliding window to calculate a structural similarity metric (e.g., similarity based on the structural similarity index (SSIM)) for each location between the current frame and the previous five frames. This similarity value serves as the basis for generating dynamic weights to control the response strength of different regions in the current frame. If a region exhibits high similarity in the historical frames (e.g., SSIM greater than 0.85), it is considered static and stable, representing background information. A weight decay operation is applied to its feature response, reducing the response factor to 0.4. If a region exhibits low temporal similarity (SSIM less than 0.45), it indicates a pet's movement, and the system increases its response weight to 1.6, thereby enhancing the ability to capture features in dynamic regions. Weighted detail feature maps, weighted scale feature maps, and weighted contour feature maps are generated, respectively. The weighted feature map is matched and analyzed using a priori knowledge base of pet features. This knowledge base contains pre-built key feature templates for 32 common pet types, including 20 cats and 12 dogs. Each template contains three sub-feature domains: head outline, body proportions, and typical hair texture, and each sub-feature is encoded as a 128-dimensional vector. The matching process combines local region feature extraction with template similarity measurement. Local sub-vectors are extracted from different regions of the weighted feature map and matched with corresponding templates in the knowledge base using cosine similarity or dot product. When the similarity exceeds a set threshold, the system marks the region as a target feature, thereby obtaining preliminary positioning information for key parts of the pet target, such as the head, body, and limbs, in the image.Channel compression and fusion are performed on the three weighted feature maps. Three parallel 1×1 convolution kernels are used to compress the channels of each branch feature map, allowing them to be fused in a unified dimension to avoid feature stacking distortion caused by differences in the number of channels. A weighted summation strategy is then used for multi-scale fusion, with the default fusion weights set to [0.25, 0.5, 0.25]. Mid-scale features are given a higher weight due to their balance between boundary preservation and structural representation. After fusion, a unified fused feature map is output. Bounding box prediction and center point confidence calculation are performed based on the response peak and spatial distribution in the fused feature map. The four parameters (x, y, w, and h) of the bounding box are inferred from the location of the maximum response heat distribution and its spatial extension in the fused feature map. The center point confidence score is normalized using a Sigmoid function to score the aggregation strength of the target region's features. This confidence value serves as a filtering threshold for subsequent behavior determination. Simultaneously, a 128-dimensional pet feature descriptor is extracted based on the structural vector of the focused region in the fused feature map. This fused information includes feature subdomains such as type category, posture angle, and motion trend, and serves as input to the target recognition and behavior prediction modules. The resulting pet feature descriptor and location information exhibit high discrimination and expressive integrity. Furthermore, thanks to multi-receptive field feature fusion and semantic prior support, it possesses high structural stability and scene adaptability, making it suitable for pet tracking, behavior recognition, and interactive decision-making in complex environments.

[0035] In a specific embodiment, the process of executing step 103 may specifically include the following steps: The pet feature descriptor and position information are input into the temporal feature extraction unit in the pet behavior recognition module. The temporal feature extraction unit performs temporal integration on the features of consecutive frames to obtain the pet posture and position change trend data. Inputting the pet's posture and position change trend data into the behavior classification unit in the pet behavior recognition module for multi-layer decision analysis to obtain initial behavior type data, wherein the multi-layer decision analysis includes a first-layer classification based on the pet's position change speed, a second-layer classification based on the relative position of the pet and the feeder, and a third-layer classification based on the pet's head posture; According to the initial behavior type data, the minimum continuous frame number threshold of approach, contact and eating behaviors is set to perform behavior confirmation filtering to obtain the pet behavior sequence feature vector.

[0036] Specifically, the pet's feature descriptors and position information are combined into a standard frame-level feature vector and input into the temporal feature extraction unit in the behavior recognition module. This unit uses a sliding window mechanism to construct a temporal stacking structure, stacking and integrating N consecutive frames of data on the time axis to form a multidimensional time series tensor. Each tensor unit contains a 128-dimensional feature descriptor and bounding box parameters (x, y, w, h), as well as a center point confidence value. First-order difference, weighted averaging, and smoothing filtering are performed on this time series data to extract dynamic features between frames, resulting in temporal information such as the pet's posture evolution, position movement trend, and contour deformation trajectory over continuous time. The pet's posture and position change trend data are input into the behavior classification unit in the pet behavior recognition module for multi-layer decision analysis. This multi-layer classification structure uses a staged and layered decision logic to gradually transition from coarse classification to fine-grained behavior recognition and judgment. The first level of classification is mainly based on the displacement amplitude of the pet's position in continuous frames to determine whether it is in a moving state. Specifically, it calculates the Euclidean distance change rate V of the center point coordinates in continuous frames, and preliminarily divides the behavior according to the set speed threshold range. If V is less than 0.05, it is judged to be stationary; if 0.05≤V<0.2, it is judged to be wandering; if V is greater than or equal to 0.2, it is judged to be approaching, thereby achieving primary recognition of the pet's movement trend; the second level of classification further introduces the relative position parameter D of the feeder, calculates the normalized distance between the center point of the pet's target frame and the center point of the feeder's physical position, and divides the behavior into three spatial relationship states according to the relative distance value: distance (D≥0.8), alert (0.4≤D<0.8) and contact (D<0.4). This level of classification emphasizes whether the pet is Relational changes such as approaching, nearing the feeder, or engaging in interactive contact represent mid-level logical inferences based on spatial structure. The third-level classification focuses on the refined semantic recognition of behavior. This involves analyzing the pet's head posture angle θ and opening / closing state S to determine its specific action tendency. This information is extracted through feature point matching analysis within the descriptor or using the head orientation prediction subnetwork. The specific behavior is then determined based on whether the feeding area is open. If θ > 30° and S = 0, observation behavior is defined; if 0 < θ ≤ 30° and S = 0, exploration behavior; and if θ ≤ 10° and S = 1, feeding behavior. This constructs a three-layer nested behavior analysis structure: "movement trend - spatial approach - posture action." This structure ensures consistency and traceability of recognition decisions across the spatial, temporal, and action semantics levels. Because pet behavior can be subject to interference from instantaneous jitter, irregular movements, and recognition errors in real applications, behavior confirmation filtering is performed to improve the stability and semantic integrity of behavior recognition.This operation is based on frame-level behavior determination results. A minimum continuous frame count threshold for key behavior types is set as a conditional filter for validity. Approach behaviors must be recognized as approaching in at least three consecutive frames to be considered valid; approaches detected in only one or two frames are ignored. The threshold for contact behaviors is set at at least five consecutive frames to account for accidental touches or brief proximity. As a highly deterministic action type, eating behaviors are judged at a threshold of eight consecutive frames or more to ensure the integrity of the eating process and high-confidence recognition. The behavior confirmation module uses a sliding window verification mechanism to analyze the temporal stability of each behavior label sequence. If the current behavior type does not meet the continuous frame count requirement, the corresponding behavior label is automatically removed or rolled back to the previous stable state, ensuring the temporal coherence and structural consistency of the final output behavior. Through the above-mentioned temporal integration, multi-layer decision classification and behavior confirmation mechanism, a structured, temporally continuous and clearly categorized pet behavior sequence feature vector is obtained. This feature vector contains the final behavior type identification of the current frame, as well as the number of frames of behavior duration, average action intensity, temporal stability level and other auxiliary parameters for subsequent state reasoning and decision control.

[0037] In a specific embodiment, the process of executing step 104 may specifically include the following steps: The behavior type identifiers in the pet behavior sequence feature vector are extracted as nodes of the state transition diagram. The connection edges are formed by calculating the transition frequencies between adjacent behavior types and the node weights and edge weights are determined by the behavior duration and intensity parameters to obtain the state transition diagram. Based on the state transition graph, we calculate the node connection pattern, transition probability distribution, and the state similarity distribution of hunger, curiosity, and satisfaction. Combining historical state sequences and state similarity distributions, we compare the matching degree between state transition trends and typical state transition patterns to obtain verified state data. The verified status data is comprehensively analyzed with the pet's individual behavior pattern library and environmental parameters to obtain the pet's status judgment result.

[0038] Specifically, the algorithm uses a continuous sequence of pet behavior feature vectors as input, where each feature vector contains at least key information such as the behavior type identifier, the number of frames the behavior lasts, and the behavior intensity level. The behavior type identifier is extracted from this sequence, and different types of behaviors, such as "approach," "contact," "explore," "feed," and "leave," are treated as independent nodes in a graph structure. Next, the adjacent transition relationships between behaviors are counted according to the order in which they occur. Directed edges are established between nodes based on the frequency of transitions, forming connecting edges in the graph structure. For example, if the sequence "approach → contact → explore" occurs frequently within a certain time period, directed edges are established multiple times between the nodes "approach" and "contact," and between "contact" and "explore." Each edge is assigned an initial edge weight based on its frequency. A weight is calculated for each node based on the number of frames the corresponding behavior lasts (i.e., the duration of the behavior) and the intensity of the behavior (typically scaled from 1 to 5). Nodes with longer durations and higher intensity have higher weights, indicating a higher dominance of the behavior in the behavioral graph during that period. In addition to frequency, edge weights also consider the smoothness of transitions between two behaviors. Specifically, if two behaviors frequently and stably transition between each other, their edge weights can be dynamically increased to reflect the determinism of state transitions. After the state transition graph is constructed, graph-theoretic feature extraction and analysis are performed on the graph structure, including node out-degree and in-degree distribution, loop structure, and path connectivity, to capture node connectivity patterns. The transition probability distribution matrix between states is then calculated based on the edge weights of the connecting edges. Based on this, a standard state template is introduced: a standard transition model for three typical states: "hunger," "curiosity," and "contentment," established based on empirical data and manual annotation. The hunger state is characterized by a high-frequency "approach → contact → exploration" cycle, accompanied by long contact durations and high-intensity exploration behaviors. The curiosity state is characterized by a low-frequency "approach → contact → departure" cycle, with short contact durations and frequent path changes. The contentment state is characterized by a sequential path of "approach → contact → feeding → departure," with a long feeding duration and a behavioral sequence that ends with departure, followed by no further approach behaviors. Using a graph structure comparison algorithm, the current state graph is structurally compared with the three standard graphs. The similarities in path structure, node weight distribution, and edge transition probabilities are calculated. The similarity between the graph structures is then compared using methods such as cosine similarity, KL divergence, or graph edit distance. This results in a similarity value between the current state graph and the three standard states of hunger, curiosity, and satisfaction, and a state similarity distribution vector is constructed. Based on this state similarity distribution, a historical state sequence is introduced as a reference for temporal consistency verification. This involves extracting the historical trajectory of state judgments from the previous time period and comparing it with the current state evolution trend to assess whether the current state transition path is structurally coherent with past behavioral patterns.For example, if the system has observed multiple closed-loop transitions of "explore → contact → approach" within the past 30 seconds, accompanied by incomplete feeding behavior, the system is more likely to maintain or confirm the "hungry" or "exploring" state in the current state. On the other hand, if the historical state has repeatedly reverted to "leave" and the current state graph does not show a new feeding trend, even if "approach" behavior occurs within a short period of time, it may not be immediately recognized as "hungry." This improves the stability and robustness of state recognition through temporal trend analysis. The system then analyzes the consistency between the state similarity distribution and historical trajectories, ultimately outputting verified intermediate state data. After obtaining verified intermediate state data, the state determination results are comprehensively corrected by incorporating individual behavioral pattern differences and environmental changes. The system includes a built-in library of individual pet behavior patterns. After the user's first use, the system creates an individual profile based on the pet's ID. This profile contains dynamic features such as a statistical model of the pet's behavioral transition probability under normal conditions, the mean duration of key behaviors, and a behavior intensity distribution curve. As the system's operating time extends beyond 72 hours, the system fine-tunes individual state boundary thresholds using a gradient descent method based on sample accumulation. For example, for a particular pet, feeding time may be shorter than the average model, or approaching behaviors may occur more frequently. The system adjusts the state judgment sensitivity boundary accordingly. Furthermore, environmental parameters can perturb pet behavior patterns. Environmental sensors collect real-time environmental data and establish an environment-behavior bias mapping model. For example, in strong noise conditions, pets may wander more frequently or observe but avoid approaching the feeding port. The system compensates for these perturbations in state judgment. Combining verified state data, individual pet behavior models, and environmental parameter characteristics, the state fusion inference module executes a multi-factor judgment algorithm. The resulting output includes a complete pet state judgment, including a state type code (e.g., 000: hungry, 001: curious, 010: satisfied), a state confidence value (0-100), and a state duration. This output serves as the core input signal for the feeder control system.

[0039] In a specific embodiment, the process of calculating the node connection pattern, the transition probability distribution, and the state similarity distribution of the hunger state, the curiosity state, and the satisfaction state based on the state transition graph may specifically include the following steps: Obtain the node connection topology in the state transition graph, and calculate the connection frequency and directionality between the behavior nodes of approach, contact, exploration, wandering, eating, and leaving based on the node connection topology to obtain the node connection pattern feature value; Calculate the behavior transition probability distribution data between nodes in the state transition graph based on the node connection pattern eigenvalues; Based on the behavioral transition probability distribution data and the standard model of hunger state, the matching degree of high-frequency approach-contact-exploration behavior cycle, long-term contact duration, and high-intensity exploration behavior characteristics was calculated to obtain the hunger state similarity value; Based on the behavior transition probability distribution data and the standard model of curiosity state, the matching degree of low-frequency approach-contact-leave behavior cycle, short-term contact duration and frequent behavior transition characteristics is calculated to obtain the curiosity state similarity value; Based on the behavioral transition probability distribution data and the standard pattern of the satisfaction state, the matching degree of the single approach-contact-eating-leaving complete sequence, the long-term feeding duration, and the no-approach feature after leaving are calculated to obtain the satisfaction state similarity value; The hunger state similarity value, curiosity state similarity value, and satisfaction state similarity value are combined to form a state similarity distribution.

[0040] Specifically, a state transition graph is constructed based on the behavior type information in the feature vector of the pet's behavior sequence, in which each behavior type such as "approach", "contact", "exploration", "wandering", "eating" and "leaving" is treated as an independent graph node. By analyzing the behavior changes between consecutive frames, directed edges are established between two adjacent different behaviors to represent the behavior transition path between states. While recording each edge, the system counts its frequency of occurrence, that is, the number of times each behavior transitions from one type to another in the entire behavior sequence, and records the directional information of the connection, such as "approach→contact" and "contact→approach" are two edges with different directions. At the same time, the out-degree and in-degree of each node are counted to describe the activity level and behavior control centrality of the node in the overall graph structure, and obtain the connection topology of the entire graph. After obtaining the node connection topology, the transition probability distribution between behavioral nodes is calculated based on edge frequency data. This is done by taking the sum of the frequencies of all outgoing edges connected to any starting node as the normalized benchmark and then probabilizing the frequency of each outgoing edge from that node. For example, if the "approach" node has three outgoing edges pointing to "contact," "wander," and "leave," with frequencies of 12, 5, and 3, respectively, then the transition probabilities of "approach → contact" are 12 / 20, "approach → wander" is 5 / 20, and "approach → leave" is 3 / 20. This method constructs a transition probability matrix between six-order behavioral states. This matrix not only reflects the density of connections between states but also captures the dynamic trends and biases of behavioral evolution within the behavioral sequence. After obtaining the transition probability data, a standard model of hunger states is introduced for structural matching. This standard model is characterized by a high frequency of "approach → contact → exploration" cyclic paths, along with long durations of "contact" nodes and high intensity "exploration" behaviors. To achieve this matching, the system examines the current transition probability matrix to determine whether the three edges (approach → contact, contact → probe, and probe → approach) form a high-frequency closed loop. The product of these three transition probabilities is used as the loop matching score. Two weighting coefficients are calculated based on the average duration of the "contact" state and the intensity distribution of the "probe" state in the actual behavior sequence. If this closed loop exists and the probabilities of all three edges are greater than 0.3, and the average contact duration is greater than a preset threshold (e.g., 5 frames) and the probe intensity is above level 3, the state is considered highly compatible with the hunger state. Finally, these three factors are combined to calculate a hunger state similarity score, typically expressed as a real number between 0 and 1. The system then transitions to matching the curiosity state, which typically follows a low-frequency three-step behavioral sequence (approach → contact → departure). The duration of contact behavior is significantly shorter than that of the hunger state, and the behavior transitions are frequent, resulting in unstable state switching.The system searches the transition graph for a linear, non-closed-loop path between the three nodes "approach → contact → leave." If such a path exists and the transition probability distribution is uniform (e.g., the transition probability for each edge is between 0.2 and 0.4), a basic structural match is established. The system then checks whether the average duration of the "contact" node is less than a preset value (e.g., 3 frames) and calculates the average period of all state transitions in the entire state graph. If the average period is less than 6 frames (i.e., frequent switching) and the path loops are minimal, it further confirms the curiosity behavior pattern. Based on this, the curiosity state similarity value is calculated by integrating the aforementioned structural, temporal, and fluctuation factors. For matches that satisfy the state, the system then searches for a complete "approach → contact → eat → leave" sequential structure. It also verifies whether the "eat" node exhibits a long-term dwell time and whether there is a prolonged period of no approach behavior after the "leave" node, that is, whether the behavior sequence terminates in the leave state. In the transition graph, the system searches for the three edge paths of "approach → contact", "contact → eating", and "eating → leaving" to see if they are continuous and have a high transition probability (e.g., each path is greater than 0.4). At the same time, it verifies in the time series that the average duration of the "eating" state is greater than 8 frames, and there is no record of any approaching behavior within 10 frames after "leaving". Such features meet the definition of a satisfied state. The system calculates its similarity with the satisfied state based on the fusion of path integrity, continuity, and subsequent stationary behavior. The three similarity values ​​mentioned above - the hunger state similarity value, the curiosity state similarity value, and the satisfaction state similarity value - are proportionally combined into a set of state similarity distribution vectors. The numerical distribution of this vector reflects the overlap and fit between the current behavior state graph and the three typical behavior patterns. The sum is normalized to facilitate subsequent state fusion and behavior decision reasoning.

[0041] In a specific embodiment, the process of executing step 105 may specifically include the following steps: The pet status judgment result is input into the non-communication adaptive power control system, and the battery management unit in the non-communication adaptive power control system monitors the battery parameters and dynamically allocates the system power based on the pet status to obtain a power allocation strategy; Perform power rail switching on the feeder according to the power allocation strategy to obtain the optimal power supply mode; Based on the optimal power supply mode and the pet status judgment result, segmented pulse power supply control is implemented to control the energy distribution of the feeder during the feeding process and obtain a feeding execution signal; Adaptive power consumption compensation control is performed on the feeding execution signal to obtain energy-balanced execution control instructions, and a power-function mapping table is established to automatically adjust function priorities at different battery power stages to achieve feeding control under low power consumption conditions.

[0042] Specifically, the pet's state determination results are input into a communication-free adaptive power control system. This system, based on local sensing, autonomous logic, and embedded decision-making mechanisms, achieves real-time power consumption scheduling and response. The core module of the power control system is the battery management unit (BMU), which continuously monitors operating parameters such as battery voltage, current, and temperature, and updates the battery state model in real time via low-power analog-to-digital sampling circuits. Simultaneously, the BMU receives the pet's behavioral state (e.g., "hungry," "curious," "contented," or "uncertain") from the state determination module. Based on the battery health status of each state priority, it calculates a dynamic power resource allocation strategy for the system. If the pet is "hungry" and the battery temperature is normal, the system increases the power allocated to the actuators (i.e., the drive motor and control circuits). If the pet is "content" or the battery enters the thermal runaway zone, the system reduces the actuator power consumption and reserves resources for the detection and state recognition chain, thus forming a dynamic power allocation strategy. After the power allocation strategy is output, automatic switching of the feeder's power rails is executed according to the strategy. The entire system features a built-in dual-rail power supply design, consisting of a conventional low-voltage battery rail and an external high-performance power rail. By default, the battery rail powers all submodules when no external power is available. Upon detecting the insertion of an external power source (such as an adapter), the power switching module switches to the primary power rail within 5ms and migrates the computational load and image enhancement modules to the high-performance rail, freeing up low-power core resources for backup mode maintenance. In the absence of external power, the system automatically activates the local voltage regulation and power limiting circuitry on the backup rail based on the power policy. This automatically adjusts the image acquisition frequency, judgment window size, and behavior confirmation threshold based on feeding priority, maintaining the system in the lowest required power mode and ensuring basic functional stability of the main control logic even under limited battery life. After the power rail switch is complete and the system power supply stabilizes, the system executes the specific feeding action based on the pet's current state and the selected power mode, using a segmented pulse power supply mechanism to finely control the energy output process. This mechanism divides the entire feeding process into three phases: the first is the motor pre-start phase, which lasts approximately 50ms and applies a 0.1A constant-current pulse to establish the motor's magnetic flux and initiate torque preparation. The second is the main drive output phase, which lasts approximately 200ms and applies a main drive current of up to 0.3A. This phase completes the core action output of the food ejection or conveying device. The third is the feedback energy reduction phase, which lasts approximately 150ms and enters the energy recovery channel and activates the reverse voltage capture mechanism. This three-stage structure allows the system to effectively suppress energy waste while ensuring effective driving. It also achieves the optimal current control target by controlling the pulse amplitude and duration, generating a structured feeding execution signal.Adaptive power consumption compensation control is applied to this execution signal. This module compares the current voltage load with the theoretical energy consumption model of the feeding action in real time. When abnormal signals such as increased battery internal resistance or delayed current response are detected, the system automatically fine-tunes the current pulse, such as shortening the main output phase or increasing the energy recovery time, to ensure that energy consumption remains within the set budget and avoid sudden system power loss due to excessive energy consumption. Simultaneously, a power-to-function mapping table is constructed, dividing the remaining battery power into multiple stages, such as ≥70%, 30%-70%, 10%-30%, and <10%. Function activation priority and operating frequency strategies are defined for each stage. When the battery level is above 70%, the system opens all functional modules, including image enhancement, contextual attention recognition, and multi-branch feature fusion; when the battery level is between 30% and 70%, the system compresses the image processing chain to a medium-resolution model and lowers the frame rate; and when the battery level drops to the 10% to 30% range, the system shuts down some inference chains, retaining only target detection and feeding judgment; if the battery level drops below 10%, it enters extreme energy-saving mode, retaining only the timed feeding mechanism and remote wake-up interface, thereby ensuring that under extreme power consumption conditions, the system can still maintain a minimum operating state for more than 72 hours.

[0043] The above describes the pet target detection method in the embodiment of the present invention. The following describes the pet target detection device in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a pet target detection device includes: The acquisition module 201 is used to perform motion-triggered acquisition and image enhancement preprocessing on the front area of ​​the feeder to obtain an enhanced image frame sequence; Feature extraction module 202, for extracting features by dynamically adjusting the receptive field of the enhanced image frame sequence to obtain pet feature descriptors and position information; Feature analysis module 203, for performing behavior temporal feature analysis based on pet feature descriptors and position information to obtain pet behavior sequence feature vectors; The consistency analysis module 204 is used to construct a state transition diagram based on the pet behavior sequence feature vector and perform time sequence consistency analysis to obtain the pet state judgment result; The decision execution module 205 is used to perform power management and decision execution on the feeder based on the pet status judgment result, so as to realize feeding control under low power consumption conditions.

[0044] Through the collaborative efforts of these components, the use of motion-triggered and dual-resolution acquisition mechanisms significantly reduces system standby power consumption, enabling the feeder to maintain ultra-low power consumption when no pet is approaching. Meanwhile, the full processing flow is rapidly activated upon pet approach, achieving a dynamic balance between power consumption and performance. A lightweight residual contextual attention network efficiently runs on low-power processors. Through a spatial perception enhancement mechanism and a cross-layer feature selection strategy, it reduces computational effort by over 90% while maintaining image enhancement quality, providing high-quality input for subsequent pet detection. The dynamic receptive field adjustment module uses three parallel feature extraction branches to capture comprehensive features of the pet, from details to overall outlines. Combined with a dynamic feature weighting mechanism, it enhances attention to areas of pet motion, improving detection accuracy. The pet behavior recognition module and temporal feature analysis mechanism accurately distinguish different pet behavior patterns. Through multi-layer decision analysis and temporal consistency constraints, they effectively avoid misjudgments caused by posture fluctuations, improving behavior recognition accuracy. State transition diagrams and temporal consistency analysis methods enable precise determination of the pet's state. By matching typical behavior patterns and accounting for individual differences, they significantly improve state determination accuracy and reduce the possibility of false triggering of feeding. The non-communication adaptive power control system dynamically adjusts system power consumption based on the pet's status. Combined with a segmented pulse power supply control strategy and an adaptive power compensation controller, this significantly extends battery life while ensuring the normal operation of core functions. This invention supports intelligent identification and personalized feeding management in multi-pet environments. By establishing a pet ID-feature mapping table and a hierarchical feeding control strategy, it solves the problem of precise feeding control in multi-pet households and improves the user experience.

[0045] An embodiment of the present invention also provides a camera, which is used to execute the steps of the above-mentioned pet target detection method.

[0046] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0047] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0048] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A pet target detection method, characterized in that: include: Perform motion-triggered acquisition and image enhancement preprocessing on the front area of ​​the feeder to obtain an enhanced image frame sequence; Performing feature extraction of the enhanced image frame sequence with dynamic receptive field adjustment to obtain pet feature descriptors and position information; Performing behavior temporal feature analysis based on the pet feature descriptor and position information to obtain a pet behavior sequence feature vector; Constructing a state transition diagram based on the pet behavior sequence feature vector and performing time sequence consistency analysis to obtain a pet state judgment result; The feeder is powered and decision-making is performed based on the pet status judgment result to achieve feeding control under low power consumption conditions.

2. The pet target detection method according to claim 1, characterized in that: The motion-triggered acquisition and image enhancement preprocessing of the front end area of ​​the feeder to obtain an enhanced image frame sequence includes: The camera's working state is controlled by a motion trigger mechanism. When a moving object is detected in the front area of ​​the feeder, the image acquisition process is activated to obtain raw image data. A dual-resolution acquisition mechanism is used to perform resolution conversion on the original image data to obtain a stable imaging frame sequence; performing image noise suppression and color space conversion on the stabilized imaging frame to obtain a single-channel compressed image, and performing edge-preserving blur processing on the single-channel compressed image to obtain a preprocessed image frame sequence; The preprocessed image frame sequence is input into a lightweight residual context attention network for quality enhancement processing to obtain an enhanced image frame sequence.

3. The pet target detection method according to claim 2, characterized in that: The step of inputting the pre-processed image frame sequence into a lightweight residual context attention network for quality enhancement processing to obtain an enhanced image frame sequence comprises: Inputting the preprocessed image frame sequence into the backbone feature extraction layer of the lightweight residual context attention network for convolution processing to obtain a basic feature map; Inputting the basic feature map into the context attention enhancement layer of the lightweight residual context attention network for parallel processing to obtain a multi-scale feature map; Calculating the brightness variance of the image region based on the multi-scale feature map, performing quantization grading based on the brightness variance of the image region and selecting a corresponding enhancement parameter group to obtain a lighting parameterized feature map; Executing a cross-layer feature selection strategy on the illumination parameterized feature map, and screening feature layer nodes by setting an activation threshold to obtain a target feature map; The target feature map is input into the texture detail recovery layer of the lightweight residual context attention network for feature upsampling and fusion mapping, and feature upsampling is performed through the deconvolution layer and feature fusion is performed through the 1×1 convolution layer to obtain an enhanced image frame sequence.

4. The pet target detection method according to claim 1, characterized in that: The feature extraction of the enhanced image frame sequence by dynamic receptive field adjustment to obtain pet feature descriptors and position information includes: Inputting the enhanced image frame sequence into a dynamic receptive field adjustment module, capturing detail features through a fixed small receptive field branch in the dynamic receptive field adjustment module, capturing medium-scale features through an adaptive medium receptive field branch in the dynamic receptive field adjustment module, and capturing overall contour features through an expanded large receptive field branch in the dynamic receptive field adjustment module, thereby obtaining an initial detail feature map, a medium-scale feature map, and an overall contour feature map; Performing dynamic feature weighting on the initial detail feature map, the medium-scale feature map, and the overall contour feature map to obtain a weighted detail feature map, a weighted scale feature map, and a weighted contour feature map; Performing matching analysis based on the weighted detail feature map, the weighted scale feature map, and the weighted contour feature map with a priori knowledge base of pet features to obtain pet target part positioning information; Performing channel compression and multi-scale feature fusion on the weighted detail feature map, the weighted scale feature map, and the weighted contour feature map to obtain a fused feature map; Based on the fused feature map, the bounding box coordinates and the center point confidence are extracted simultaneously to obtain the pet feature descriptor and position information.

5. The pet target detection method according to claim 1, characterized in that: The performing behavior temporal feature analysis based on the pet feature descriptor and the position information to obtain the pet behavior sequence feature vector includes: Inputting the pet feature descriptor and position information into the temporal feature extraction unit in the pet behavior recognition module, and performing temporal integration of continuous frame features through the temporal feature extraction unit to obtain pet posture and position change trend data; Inputting the pet posture and position change trend data into the behavior classification unit in the pet behavior recognition module for multi-layer decision analysis to obtain initial behavior type data, wherein the multi-layer decision analysis includes a first-layer classification based on the pet's position change speed, a second-layer classification based on the relative position of the pet and the feeder, and a third-layer classification based on the pet's head posture; The minimum continuous frame number thresholds of the approach, contact and eating behaviors are set according to the initial behavior type data to perform behavior confirmation filtering to obtain a pet behavior sequence feature vector.

6. The pet target detection method according to claim 1, characterized in that: The step of constructing a state transition diagram based on the pet behavior sequence feature vector and performing a time sequence consistency analysis to obtain a pet state judgment result includes: Extracting behavior type identifiers from the pet behavior sequence feature vector as nodes of a state transition diagram, forming connecting edges by calculating the transition frequencies between adjacent behavior types, and determining node weights and edge weights using behavior duration and intensity parameters to obtain a state transition diagram; Calculate the node connection pattern, transition probability distribution, and state similarity distribution of the hunger state, curiosity state, and satisfaction state based on the state transition graph; Combining the historical state sequence and the state similarity distribution, comparing the matching degree between the state transition trend and the typical state transition law, and obtaining the verified state data; The verified status data is comprehensively analyzed with the pet's individual behavior pattern library and environmental parameters to obtain a pet status judgment result.

7. The pet target detection method according to claim 6, characterized in that: The calculating of the node connection pattern, the transition probability distribution, and the state similarity distribution of the hunger state, the curiosity state, and the satisfaction state based on the state transition graph includes: Obtaining a node connection topology in the state transition graph, and calculating the connection frequency and directionality between the behavior nodes of approach, contact, exploration, wandering, eating, and leaving based on the node connection topology to obtain a node connection pattern feature value; Calculating behavior transition probability distribution data between nodes in the state transition graph based on the node connection pattern characteristic value; Based on the behavior transition probability distribution data and the standard pattern of hunger state, the degree of matching of high-frequency approach-contact-probing behavior cycles, long-term contact duration, and high-intensity probing behavior characteristics is calculated to obtain a hunger state similarity value; Based on the behavior transition probability distribution data and the curiosity state standard model, the matching degree of low-frequency approach-contact-leave behavior cycle, short-term contact duration, and frequent behavior transition characteristics is calculated to obtain the curiosity state similarity value; Based on the behavior transition probability distribution data and the standard pattern of the satisfied state, the matching degree of the single approach-contact-eating-leaving complete sequence, the long-term feeding duration, and the no-approach feature after leaving are calculated to obtain the satisfied state similarity value; The hunger state similarity value, the curiosity state similarity value, and the satisfaction state similarity value are combined to form a state similarity distribution.

8. The pet target detection method according to claim 1, characterized in that: The power management and decision execution of the feeder based on the pet status judgment result to achieve feeding control under low power consumption conditions include: Inputting the pet status judgment result into the non-communication adaptive power control system, monitoring battery parameters through the battery management unit in the non-communication adaptive power control system and dynamically allocating system power based on the pet status to obtain a power allocation strategy; performing power rail switching on the feeder according to the power allocation strategy to obtain an optimal power supply mode; Based on the optimal power supply mode and the pet state judgment result, segmented pulse power supply control is implemented to control the energy distribution of the feeder during the feeding process to obtain a feeding execution signal; Adaptive power consumption compensation control is performed on the feeding execution signal to obtain an energy-balanced execution control instruction, and a power-function mapping table is established to automatically adjust the function priority at different battery power stages to achieve feeding control under low power consumption conditions.

9. A pet target detection device, characterized in that: For executing the pet target detection method according to any one of claims 1 to 8, the pet target detection device comprises: An acquisition module is used to perform motion-triggered acquisition and image enhancement preprocessing on the front area of ​​the feeder to obtain an enhanced image frame sequence; A feature extraction module, configured to extract features of the enhanced image frame sequence by dynamically adjusting the receptive field to obtain pet feature descriptors and position information; A feature analysis module, configured to perform behavior temporal feature analysis based on the pet feature descriptor and position information to obtain a pet behavior sequence feature vector; A consistency analysis module is used to construct a state transition diagram based on the pet behavior sequence feature vector and perform time sequence consistency analysis to obtain a pet state judgment result; The decision execution module is used to perform power management and decision execution on the feeder based on the pet status judgment result, so as to realize feeding control under low power consumption conditions.

10. A camera, characterized in that: The camera is used to execute the pet target detection method according to any one of claims 1 to 8.

Images