A multi-cat individual-oriented long-term identity recognition method and system

Abstract

The application discloses a long-term identity recognition method and system for multiple cat individuals. The face three-dimensional point cloud data, cat eye iris data and palm print data of the cat are collected and converted to the same three-dimensional joint coordinate system for alignment of the physical space. Then, three parallel feature extraction networks are used to extract features from the aligned data, and after splicing and dimension reduction, preliminary fusion features are obtained. The preliminary fusion features are subjected to attention enhancement and residual connection to obtain a fusion feature vector. The fusion feature vectors of an individual at different periods are saved in a historical feature sequence in chronological order to construct a multi-individual feature library. The most similar individual to the individual to be identified is found in the feature library as a candidate cat through cosine similarity comparison. The time sequence evolution features of the candidate cat are predicted, the similarity between the time sequence evolution features, the historical features and the individual to be identified is judged to determine the identity of the individual to be identified, and the feature library is updated.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and relates to biometric recognition technology, specifically to a method and system for long-term identification of multiple cat individuals. Background Technology

[0002] In smart devices such as pet medical management, pet shelters, and stray cat houses, pet identification is an important task. Its core purpose is to uniquely identify different animal individuals through biometrics in order to achieve personalized services and management.

[0003] Existing solutions for cat identification, based on computer vision and AI classification algorithms, primarily focus on single or simple multimodal data processing. Compared to common non-biological targets, biological targets exhibit significant individual changes, especially in kittens. After six months, their facial features change considerably, causing a noticeable decline in the recognition rate of single-modal methods after this age. Furthermore, cross-modal features exhibit inconsistencies in physical space, leading to substantial deviations and introducing noise during fusion, severely impacting accuracy and robustness. In addition, traditional solutions rely on static feature libraries or passive, simple feature replacements only when recognition fails. This mechanism cannot learn and predict the intrinsic evolution of biological features over time, thus failing to proactively adapt to long-term, gradual feature drift.

[0004] In summary, existing technologies lack a robust solution for long-term identification of animal targets. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a long-term identification method and system for multiple individual cats. By deeply fusing multimodal features such as computer vision, near-infrared imaging, and bioimpedance detection, learning a highly discriminative feature space, and actively predicting temporal evolution, this invention solves the problem of feature drift failure caused by growth in existing cat identification methods.

[0006] A long-term identification method for multiple cats, comprising the following steps:

[0007] Step 1: Collect multimodal data of the cat, including facial 3D point cloud data, cat eye iris data, and palm print data.

[0008] Step 2: Using the cat's nose tip as the origin of the three-dimensional joint coordinate system, transform the iris data and palm print data into the three-dimensional joint coordinate system through matrix transformation to ensure that the multimodal data are aligned in three-dimensional space.

[0009] Step 3: Use three parallel feature extraction networks to extract features from the aligned data in Step 2. Then, concatenate the feature vectors output by the three feature extraction networks in terms of dimensions and project them back to the original input dimension to obtain the preliminary fused feature F'.

[0010] Step 4: Enhance the initial fused feature F' using a self-attention mechanism, and then perform a residual connection with the initial fused feature F' to obtain the fused feature vector. .

[0011] Step 5: Combine the feature vectors of each individual at different times. The features are stored in chronological order in a historical feature sequence to build a feature library.

[0012] Search the feature library for fused feature vectors that match the individual to be identified. The closest feature vector is selected as the candidate cat. The historical feature sequences of the candidate cats are input into a recurrent neural network to generate predicted feature vectors. The fused feature vector is then calculated separately. The similarity between the current cat's most recent historical feature vector and its predicted feature vector is weighted and summed to obtain the identity confirmation confidence score. If the identity confirmation confidence score is lower than a set threshold, the current cat is determined to be an unknown individual, and its features are added to the feature library. Otherwise, the current cat is determined to be a candidate cat, and its feature vector is fused. The feature library is intelligently updated by adding the historical feature sequence of the candidate cats and removing the oldest historical feature.

[0013] A long-term identification system for multiple cats includes a data acquisition module, a spatial alignment module, a feature extraction module, a nonlinear fusion module, and a feature library update module.

[0014] The data acquisition module includes a ToF depth sensor for acquiring three-dimensional point cloud data of the cat's face, a multispectral camera operating in the near-infrared band for capturing the texture of the cat's iris, and a 4x4 bioimpedance sensor array for acquiring impedance distribution grid data of the paw pads when the cat makes contact.

[0015] The spatial alignment module uses the cat's nose tip as the origin of the coordinate system to transform the multimodal data collected by the data acquisition module into the same three-dimensional joint coordinate system.

[0016] The feature extraction module uses three parallel feature extraction networks to extract features from the aligned data.

[0017] The nonlinear fusion module performs dimensional concatenation and projection on the features output by the feature extraction network to obtain preliminary fused features. Then, a self-attention mechanism is applied to enhance the preliminary fused features, and residual connections are made with the preliminary fused features to output a fused feature vector.

[0018] The identification and feature update module uses a recurrent neural network to predict the temporal evolution of historical feature sequences. By comparing the similarity between predicted features, historical features, and fused features, the identity of the individual cat is confirmed, and the historical feature sequence is updated using the fused features.

[0019] The present invention has the following beneficial effects:

[0020] 1. It introduces temporal evolution prediction, transforming from passively compensating for changes in biological characteristics to actively predicting the trend of characteristic evolution, solving the problem of characteristic drift, and accurately identifying individuals in the long-term growth and change process.

[0021] 2. By adopting a nonlinear fusion network based on the attention mechanism to replace the traditional linear weighting method, it can deeply explore and utilize the complex relationships between different biological features, generate fusion features with higher information density and stronger expressive power, and give full play to the advantages of multimodal collaboration.

[0022] 3. Through a rigorous multimodal spatial alignment process, cross-modal data bias caused by different sensor physical locations is eliminated, providing high-quality, unbiased input for all subsequent processing steps and ensuring the physical meaning and accuracy of the fusion results. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of multimodal data acquisition and spatial alignment.

[0024] Figure 2 This is a schematic diagram of feature extraction and preliminary fusion.

[0025] Figure 3 This is a schematic diagram of nonlinear fusion. Detailed Implementation

[0026] The present invention will be further explained below with reference to the accompanying drawings;

[0027] A long-term identification system for multiple cats includes a data acquisition module, a spatial alignment module, a feature extraction module, a nonlinear fusion module, and a feature library update module.

[0028] The data acquisition module includes a ToF depth sensor for acquiring three-dimensional point cloud data of the cat's face, a multispectral camera operating in the near-infrared band for capturing the texture of the cat's iris, and a 4x4 bioimpedance sensor array for acquiring impedance distribution grid data of the paw pads when the cat makes contact.

[0029] The spatial alignment module uses the cat's nose tip as the origin of the coordinate system to transform the multimodal data collected by the data acquisition module into the same three-dimensional joint coordinate system.

[0030] The feature extraction module uses three parallel feature extraction networks to extract features from the aligned data.

[0031] The nonlinear fusion module performs dimensional concatenation and projection on the features output by the feature extraction network to obtain preliminary fused features. Then, a self-attention mechanism is applied to enhance the preliminary fused features, and residual connections are made with the preliminary fused features to output a fused feature vector.

[0032] The identification and feature update module uses a recurrent neural network to predict the temporal evolution of historical feature sequences. By comparing the similarity between predicted features, historical features, and fused features, the identity of the individual cat is confirmed, and the historical feature sequence is updated using the fused features.

[0033] A long-term identification method for multiple cats, comprising the following steps:

[0034] Step 1, as follows Figure 1 As shown, three-dimensional point cloud data of the cat's face, iris data of the cat's eyes, and palm print data were collected.

[0035] Step 2: Spatial alignment of the multimodal features collected in Step 1 is performed to solve the problem of cross-modal feature fragmentation.

[0036] First, a lightweight pose estimation model, MediaPipe Animal Pose, is used to locate the cat's nose tip. The cat's nose tip is then used as the origin of a 3D joint coordinate system. Next, using pre-calibrated rotation and translation matrices, the coordinates of the iris and palmprint data are transformed into the 3D joint coordinate system, resulting in palmprint data in the same 3D joint coordinate system. Iris data and facial data This ensures that multimodal data is aligned in three-dimensional space.

[0037] Step 3, as follows Figure 2 As shown, three parallel feature extraction networks are used to process the palmprint data aligned in step 2. Iris data and facial data Feature extraction is performed, and then the feature vectors output by the three feature extraction networks are concatenated dimensionally. A fully connected layer and the ReLU activation function are used for non-linear transformation, followed by convolution. The result is projected back to the original input dimension, yielding a preliminary 256-dimensional fused feature F'.

[0038]

[0039] in, and These are the learnable weight matrix and bias vector. This represents the concatenated features. Fully connected layers can extract abstract features, while the ReLU activation function introduces non-linearity, enabling the network to learn complex relationships.

[0040] Specifically, 3D sparse convolutional networks are used to process unstructured facial data. Feature extraction was performed to obtain 256-dimensional three-dimensional geometric features. Using ResNet-18 residual network to process iris data from single-channel near-infrared images. Feature extraction was performed to obtain 256-dimensional iris features. A lightweight, depthwise separable CNN network was used to process palmprint data. Feature extraction was performed to obtain 256-dimensional bioimpedance features. .

[0041] Step 4, as follows Figure 3 As shown, a self-attention mechanism is applied to the initial fused feature F' to calculate the correlation between different elements within it, adaptively identifying the most critical feature dimension in the fusion task and increasing its weight, while suppressing relatively minor or redundant features. Then, residual connections are used to enhance the attention-enhanced features. The final fused feature vector is obtained by performing a residual connection with the initial fused feature F'. :

[0042]

[0043] Here, α is a learnable scalar parameter, initialized to 0, used to control the initial influence of the self-attention mechanism and ensure training stability.

[0044] By using residual connections, the features before and after attention enhancement are added together, ensuring that the network increases the weights for more useful features while also preserving the original features.

[0045] Step 5: Collect facial 3D point cloud data, cat eye iris data, and palm print data for the same cat at different times, and obtain the fused feature vectors for different times according to the methods in steps 2-4. As historical feature vectors, they are stored in the historical feature sequence in chronological order, and the historical feature sequences of different cats are stored in the feature database.

[0046] For the individual to be identified, the fused feature vector is obtained through steps 1-4. Traverse the historical feature sequences in the feature database and fuse the feature vectors of the individual to be identified. The cosine similarity is calculated with the last historical feature vector in the historical feature sequence, and the individual corresponding to the historical feature vector with the highest similarity is selected as the candidate cat i.

[0047] Select the most recent n historical feature sequences of candidate cat i. The historical feature sequence The input is fed into a Long Short-Term Memory (LSTM) network to obtain the predicted feature vectors for temporal evolution. .

[0048] Calculate the fusion feature vector of each individual to be identified. The most recent historical feature vector of candidate cat i Predicting feature vectors similarity between and Then calculate the final confidence level for identity verification. :

[0049]

[0050]

[0051]

[0052] in, Indicating historical similarity, Indicates evolutionary similarity. , The weights for historical similarity and evolutionary similarity are respectively, and these are hyperparameters that are set manually. In this embodiment... =0.4、 =0.6.

[0053] Set the confidence threshold τ = 0.85. When Confidence > τ, the identity of the individual to be identified is confirmed as candidate cat i. Added as the latest historical feature At the same time remove The earliest historical features , making the historical feature sequence The length remains n. If Confidence ≤ τ, then it is determined to be an unknown individual, a new identity is registered in the feature database, and... Add it to the historical feature sequence of the new individual.

Claims

1. A long-term identification method for multiple cat individuals, characterized in that: The specific steps are as follows: Step 1: Collect multimodal data of the cat, including facial 3D point cloud data, cat eye iris data, and palm print data; Step 2: Convert the facial 3D point cloud data, cat eye iris data, and palm print data to a 3D joint coordinate system using matrix transformation. Step 3: Use three parallel feature extraction networks to extract features from the aligned data in Step 2, then concatenate them dimensionally and project them back to the original input dimension to obtain the preliminary fused features F'. Step 4: Enhance the initial fused feature F' using a self-attention mechanism, and then perform a residual connection with the initial fused feature F' to obtain the fused feature vector. ; Step 5: Fuse the feature vectors of each individual cat at different times. The features are stored in chronological order in the historical feature sequence, and the historical feature sequences of different cats constitute the feature library; Search the feature library for fused feature vectors that match the individual to be identified. The closest individual feature vector is selected as the candidate cat; the historical feature sequence of the candidate cat is input into a recurrent neural network to generate a predicted feature vector; the fused feature vector is calculated separately. The similarity between the current cat and the candidate cat's most recent historical feature vector and predicted feature vector is weighted and summed to obtain the identity confirmation confidence score. If the identity confirmation confidence score is lower than a set threshold, the current cat is determined to be an unknown individual, its features are added to the feature library, and the historical feature sequence of the new individual is registered; otherwise, the current cat is determined to be a candidate cat, and the fused feature vector is... The feature library is intelligently updated by adding the historical feature sequence of the candidate cats and removing the oldest historical feature.

2. The long-term identification method for multiple cat individuals as described in claim 1, characterized in that: The position of the cat's nose tip in the facial 3D point cloud data is used as the origin of the 3D joint coordinate system. Then, the coordinates of the iris data and palm print data are transformed into the 3D joint coordinate system through a pre-calibrated rotation and translation matrix.

3. The long-term identification method for multiple cat individuals as described in claim 2, characterized in that: The cat's nose tip was located using the pose estimation model MediaPipe Animal Pose.

4. The long-term identification method for multiple cat individuals as described in claim 1, characterized in that: Feature extraction was performed on aligned facial 3D point cloud data, cat eye iris data, and palm print data using 3D sparse convolutional networks, ResNet-18 residual networks, and lightweight deep separable CNN networks, respectively.

5. The long-term identification method for multiple cat individuals as described in claim 4, characterized in that: The feature vectors output by the three feature extraction networks are concatenated in dimension, and a non-linear transformation is performed using a fully connected layer and the ReLU activation function. Then, a convolution operation is performed to project the vectors back to the dimension of the original input, resulting in the preliminary fused feature F'.

6. The long-term identification method for multiple cat individuals as described in claim 1, characterized in that: Select historical feature sequences of candidate cats The input is fed into a Long Short-Term Memory (LSTM) network to obtain the predicted feature vectors for temporal evolution. .

7. The long-term identification method for multiple cat individuals as described in claim 6, characterized in that: The confidence level of identity verification for: in, Indicating historical similarity, Indicates evolutionary similarity; , These are the weights for historical similarity and evolutionary similarity, respectively. This represents the most recent historical feature vector of the candidate cat.

8. The long-term identification method for multiple cat individuals as described in claim 7, characterized in that: Set weights =0.4、 =0.6, and the confidence threshold for identity verification τ=0.

85.

9. A long-term identification system for multiple individual cats, characterized in that: The method for implementing the identity recognition method as described in any one of claims 1 to 8 includes a data acquisition module, a spatial alignment module, a feature extraction module, a nonlinear fusion module, and a feature library update module. The data acquisition module is used to acquire three-dimensional point cloud data of the cat's face, iris data of the cat's eyes, and palm print data; The spatial alignment module is used to transform the multimodal data acquired by the data acquisition module into the same three-dimensional joint coordinate system; The feature extraction module uses three parallel feature extraction networks to extract features from the aligned data. The nonlinear fusion module performs dimensional concatenation and projection on the features output by the feature extraction network to obtain preliminary fused features. Then, a self-attention mechanism is applied to enhance the preliminary fused features, and a residual connection is made with the preliminary fused features to output a fused feature vector. The feature library update module uses a recurrent neural network to perform temporal evolution prediction from the historical feature sequence. By comparing the similarity between the predicted features, historical features, and fused features, the identity of the individual cat is confirmed, and the historical feature sequence is updated using the fused features.

10. The long-term identification system for multiple cats as described in claim 9, characterized in that: Three-dimensional point cloud data of the face is obtained through a ToF depth sensor, cat eye iris data is obtained through a multispectral camera operating in the near-infrared band, and palm print data is obtained through a bioimpedance sensor array.

Images