A chicken multi-pose estimation method based on a multi-view imaging system and a TriFuseNet-OSDA

Abstract

This invention discloses a chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA, comprising the following steps: acquiring chicken video data using a multi-view depth camera in a multi-camera system to form a raw chicken video dataset; extracting RGB images from three perspectives from the raw chicken video dataset and constructing an RGB image dataset; constructing a TriFuseNet-OSDA model; the TriFuseNet-OSDA model includes: an input end; a shallow feature enhancement module; a feature extraction network; a dynamic multi-branch complementary fusion structure; a classification head; and an output end connected to the classification head; training the TriFuseNet-OSDA model using the RGB image dataset; after training, inputting the RGB images of the chickens to be detected from the three perspectives into the model to estimate the chicken's multi-pose and outputting the chicken pose classification result.

Description

Technical Field

[0001] This invention relates to the field of non-invasive poultry pose estimation technology, specifically to a chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA. Background Technology

[0002] Animal health and welfare monitoring is of great significance in animal behavior, ecology, and intensive animal husbandry. Postural information can reflect an individual's behavioral patterns, physiological state, and potential health risks. Therefore, accurate posture estimation not only provides key features for behavioral identification and health assessment but also provides a basis for management and intervention decisions. In production practice, posture estimation is applied to a wide range of targets, from livestock animals to laboratory animals. For example, posture changes can be used to identify abnormal behaviors, assess stress responses, or monitor disease progression, thereby improving production efficiency and animal welfare.

[0003] Traditional pose estimation methods primarily rely on manual observation and recording. This approach is not only inefficient and labor-intensive, but also susceptible to the subjective influence of the observer, making it difficult to achieve large-scale, continuous, and objective monitoring. With the development of sensing and computer vision technologies, the use of non-contact sensors for automatic pose detection is gradually becoming a trend. While contact sensors can provide high-precision data, they suffer from high equipment maintenance costs and are prone to triggering stress responses, limiting their application scope. In contrast, computer vision-based pose estimation offers advantages such as non-invasiveness, high automation, and flexible deployment, with minimal interference to the monitored object, making it an important technological direction for animal pose detection.

[0004] Early animal pose estimation methods based on traditional vision techniques mainly relied on marker tracking or morphological feature extraction based on image processing. For example, researchers obtained gait information by placing marked balls on the hind limb joints of rats, used the morphological features of pig backs for pose recognition, and combined background subtraction and clustering segmentation to analyze the movement posture of fruit flies. In studies on birds, researchers combined color and depth information, using image correction and depth threshold-based segmentation to obtain the morphology of chickens, achieving automatic estimation of standing and lying postures. However, these traditional vision-based pose estimation methods still have shortcomings in terms of pose category coverage, estimation accuracy, and adaptability to complex environments.

[0005] The development of deep learning has driven rapid progress in animal pose estimation techniques, including methods based on single-frame image classification, keypoint detection combined with machine learning, and object detection combined with pose estimation models. For example, researchers have used top-view cameras and densely connected convolutional networks (DenseNet) to identify four chicken poses; they have combined residual network 50 (ResNet50) and DeepLabcut to detect keypoints and used a Naive Bayes classifier to estimate six animal poses; researchers have used the cascade encoder-decoder network (CEDNet) model to locate chickens and combined it with CBLFormer to detect multiple chicken poses; and researchers have used You Only Look Once X (YOLOX) and pre-trained EfficientNet to achieve keypoint detection and multi-pose recognition. These methods have made progress in multi-class pose recognition, but they still generally face challenges such as limited information dimensionality, sensitivity to occlusion and viewpoint changes, and insufficient ability to finely distinguish complex poses.

[0006] Multi-view information fusion provides an effective approach to improve the accuracy and robustness of pose estimation. By fusing complementary information from different perspectives, the limitations of single-view systems in terms of occlusion and perspective distortion can be significantly alleviated, enhancing the expressive power of pose features. Existing studies have validated the advantages of multi-camera systems in localization and pose estimation. Researchers have achieved all-weather accurate localization using four-way top / side view information, and researchers have performed high-precision detection of four poses in a multi-camera system and a multi-Source composite backbone (MSCB) fish swarm. Researchers have also used four cameras and the Social Behavior Atlas (SBeA) computational framework to estimate various social poses of mice. These results demonstrate that multi-view fusion is a general pose estimation enhancement strategy, applicable not only to the animal domain but also transferable to state recognition and pose estimation tasks for other targets.

[0007] Existing multi-view pose estimation still faces three technical bottlenecks in feature extraction and fusion. First, insufficient cross-view feature interaction. Many methods use simple concatenation or weighted summation in the fusion stage, making it difficult to deeply model the complementarity and correlation between viewpoints, thus limiting the fusion effect. Second, insufficient utilization of the hierarchical nature of feature representations. Low-level features contain rich texture and edge information, but in multi-view tasks, they are often over-sampling or weakened, affecting the capture of fine-grained pose cues. Third, a balance between lightweight design and high accuracy is difficult to achieve. While deep fusion structures improve accuracy, they often introduce significant parameter increases and computational overhead, hindering edge deployment and real-time monitoring. Therefore, a multi-view pose estimation framework that combines efficient cross-view interaction, effective utilization of shallow information, and lightweight structure is needed to achieve a balance between multi-pose coverage, estimation accuracy, and computational cost. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA.

[0009] A chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA includes the following steps: 1) Use multi-view depth cameras in a multi-camera system to collect chicken video data and form a raw dataset of chicken videos; 2) Extract RGB images from three perspectives from the original chicken video dataset and construct an RGB image dataset; 3) Construct the TriFuseNet-OSDA model; The TriFuseNet-OSDA model includes: Input terminal; The shallow feature enhancement module is connected to the input terminal; The feature extraction network connected to the shallow feature enhancement module; A dynamic multi-branch complementary fusion structure connected to the feature extraction network; The classification head is connected to the dynamic multi-branch complementary fusion structure; The output terminal connected to the classification head; 4) The TriFuseNet-OSDA model is trained using an RGB image dataset. After training, the RGB images of chickens from three perspectives to be detected are input into the trained TriFuseNet-OSDA model to estimate the multiple poses of the chickens and output the chicken pose classification results.

[0010] Step 1) The multi-view imaging system includes three depth cameras fixed to the top, front and side of the measurement platform, and equipped with RFID reading and writing devices to identify the chickens. When the chickens walk into the area for image acquisition according to the set trajectory, the three depth cameras start to collect image data from three perspectives simultaneously. The data acquisition time for each chicken is 300 seconds, and the data collected by each camera is named in a way that combines identity and perspective to facilitate subsequent data processing.

[0011] Step 2) involves constructing the dataset by extracting RGB and depth images from the MKV video data for each viewpoint. A 300-second MKV video can ultimately yield 1800 RGB images. The original resolution of each image is 1920×1080, but for ease of model training, the resolution is uniformly adjusted to 640×360. Image data for eight postures (eating, flapping, head-down, preening, resting, standing, stooping, and walking) from three viewpoints were manually selected. This invention obtained 3734 sets of chicken posture image data from three viewpoints, totaling 11202 images.

[0012] In step 3), a TriFuseNet-OSDA model for chicken multi-pose estimation is constructed. In the feature extraction stage, this model employs an optimized lightweight OverLoCK-xt-light to reduce parameters and computational cost. In the fusion stage, a self-developed dynamic multi-branch complementary fusion structure is introduced. This structure uses a "sharing-separation-complementarity-fusion" strategy to share parameters at low levels to control costs, establish independent perspective branches at mid-to-high levels to enhance differential representation, and achieve adaptive fusion through cross-branch attention. Simultaneously, DynamicTanhECA channel attention is designed for adaptive weighting, and an optimized SRFD-enhanced shallow feature enhancement module is introduced to improve fine-grained cue retention. This is combined with the self-developed Attn-GeM-Lite lightweight classification head to achieve efficient inference.

[0013] In the shallow feature enhancement module, the feature extraction stage replaces the Conv convolution module with the DWConv module and adds a Coordinate Attention module. In the downsampling stage, the hierarchical Cut-Fusion structure replaces the single cascaded downsampling structure.

[0014] The feature extraction network adopts the OverLoCK-xt framework, which includes three parts: base, overview, and focus. The DWCony convolutional modules in the base and overview parts are replaced with Ghost-DWConv convolutional modules, and channel compression and upsampling alignment operations are added to the overview part.

[0015] The dynamic multi-branch complementary fusion structure includes: A feature extraction layer sharing basic features is used to obtain basic features, which include: basic texture information, edges, colors, and simple shapes. A feature extraction layer for separating complex features, connected to the feature extraction layer that shares basic features, is used to obtain complex features, which include complex spatial relationships, local pose details, and viewpoint-specific information. A complementary layer, wherein the basic features and complex features are exchanged and complemented to obtain three-branch feature information; The fusion layer combines the feature information from the three branches to form a fused feature output.

[0016] The complementary layer exchanges and complements the basic features and complex features to obtain three-branch feature information, specifically including: 3.1) Weighted optimization is performed on complex features to obtain features after removing redundancy; 3.2) The redundant features are then processed through the interaction of each branch to obtain complete branch feature information; 3.3) The complete branch feature information is interactively weighted with the basic features to obtain the three-branch feature information.

[0017] The classification head is used to classify and output the fused features, specifically including: First, the fused features are subjected to GhostConv convolution and channel recalibration. Then, channel weighting is performed through global statistics. Subsequently, a global representation is obtained by learningable generalized mean and average aggregation. Finally, the classification result is output by gating fusion of cosine metric classifier and linear classifier.

[0018] The DMCF architecture works as follows: The overall design employs a strategy of "shared low-level features + independent high-level branches," enabling the model to balance high accuracy with lightweight efficiency. In the low-level stage, the model primarily captures basic texture information, edges, colors, and simple shapes from the input image. These features exhibit strong commonality and consistency across different viewpoints, making them suitable for unified extraction using parameter sharing. The model achieves parameter sharing across three different viewpoint inputs during the low-level feature extraction stage, completing the low-level feature encoding through unified feature extraction. Compared to designing independent low-level branches for each viewpoint, parameter sharing significantly reduces the number of parameters in the model while promoting alignment and consistency of basic information across different viewpoints, avoiding feature distribution drift and redundancy. This design not only improves computational efficiency but also enhances the model's feature representation.

[0019] In the mid- and high-level stages, features gradually shift towards more abstract semantic representations, incorporating complex spatial relationships, local pose details, and viewpoint-specific information. To address this, we divide the features into three parameter-independent branches, granting each viewpoint ample freedom of expression, enabling refined learning and semantic capture tailored to its specific characteristics. Simultaneously, within each branch, a lightweight ECA-Lite channel attention mechanism is introduced to rapidly capture the correlation between feature channels, thereby improving feature selectivity and representational stability. Furthermore, a GateFC gate layer is employed to dynamically generate gating coefficients for each channel using a fully connected structure, effectively suppressing redundant features and highlighting key information. The synergistic effect of these attention and gating modules allows for dynamic adjustment of branch feature responses, achieving enhanced expression of key features and noise suppression, thus improving the compactness and discriminative power of feature representation. To enhance information interaction between different branches, the model introduces the CrossBranch-ECA efficient cross-branch channel attention mechanism in the mid- and high-level feature stages, effectively promoting information exchange and complementarity between viewpoint branches, thereby enhancing feature diversity and correlation. Meanwhile, between different feature levels, we use the Context Gate module to generate adaptive weights by fusing multi-stage global features to dynamically recalibrate the responses of higher-level channels. This effectively mitigates semantic drift of higher-level features and ensures the consistency and stability of multi-scale feature representations. Finally, the three higher-level features interact at the output, and branch-scale attention dynamically allocates weights based on the global representation of each branch to achieve adaptive weighted fusion of multi-view features, ensuring that the contribution of key perspectives in decision-making is fully reflected. Subsequently, the fused unified features are fed into a lightweight and adaptive classification head to complete the discrimination, further integrating local details and global semantics while controlling computational overhead, and generating the final multi-class prediction.

[0020] Step 3) uses the DynamicTanhECA module as follows: To enhance channel interactions between mid-level features in the model, we designed a lightweight and learnable module called DynamicTanhECA. This module enhances the dynamic nonlinear representation of features and the modeling of channel dependencies. The module is based on channel normalization and a parameterized hyperbolic tangent activation function. It achieves efficient channel attention gating, adjustable residual fusion coefficients, and a random dropout mechanism during training through one-dimensional convolution, improving the robustness and expressive power of the model without significantly increasing the number of parameters. The specific process is as follows: First, layer LN is applied to the input feature x to obtain the normalized feature x. n Then, the learnable vector α and the feature x are used n Element-wise multiplication is performed, and a non-linear output is obtained through Tanh, as shown in the following formula.

[0021] (1) Where y1 represents the nonlinear output obtained after the above operation. This represents the learnable vector obtained by transforming the learnable parameter raw_alpha using the softplus function, where Tanh represents the hyperbolic tangent activation function, and x... n represents the normalized eigenvector, and ʘ represents element-wise multiplication.

[0022] Subsequently, the mean of y1 in the spatial dimension is calculated to obtain the channel feature description vector. Then, it is processed by a 1D convolution with adaptive kernel size and the Sigmoid function to generate dynamic gating coefficients g. y1 is then gated using g to obtain the output y2. This process is the core process of the dynamic ECA gating mechanism. By dynamically adjusting the weights of each channel through adaptive local convolution, the inter-channel dependencies are effectively modeled. The formula is as follows.

[0023] (2) Where g represents the dynamic gating coefficient. Represents the Sigmoid function, and Conv1D is a 1D convolution with the kernel size adaptively based on the number of channels. y1 represents the mean function, y2 represents the non-linear output, y3 represents the output obtained after the above series of operations, and ʘ represents element-wise multiplication.

[0024] Next, a learnable affine transformation is applied to y2 and LN is applied again to obtain y3. Finally, the output is obtained by controlling the residual fusion using learnable coefficients λ∈[0,1] and using the DropPath random discarding mechanism during training. The formula is as follows.

[0025] (3) Where Y represents the final output, y3 represents the output after layer normalization of inter-channel dependencies, x represents the initial input, and DP represents DropPath.

[0026] The DynamicTanhECA module enhances nonlinear expressiveness through a parameterized Tanh activation function, effectively captures inter-channel dependencies by combining a dynamic ECA gating mechanism based on adaptive one-dimensional convolution, and improves training stability and model generalization ability by utilizing adjustable residual fusion and DropPath. The module's design balances feature expressiveness with lightweight design, making the model more robust and efficient in fine-grained feature modeling and complex backgrounds.

[0027] The principle of Attn-GeM-Lite in step 3) is as follows: To address the shortcomings of traditional classification heads, such as difficulty in modeling channel dependencies, lack of discriminative margins and robust pooling, and susceptibility to feature scale fluctuations, we designed the lightweight and learnable Attn-GeM-Lite classification head. This head uses GhostConv for efficient feature generation and channel recalibration, combined with ECA-Lite's adaptive one-dimensional convolutional channel attention, learnable exponential GeM, and a "dual-branch classifier + Softmax gated fusion" strategy, to improve the robustness and expressive power of the model while maintaining stable model parameters. The specific process of the module is as follows: First, intermediate features are obtained from the input feature x using GhostConv, and then the feature channels are recalibrated once using SE-Lite to obtain feature x. se The next step is to sequentially perform GAP, ECA-Lite, and Sigmoid on the recalibrated features to obtain channel weights w, and then multiply them channel by channel with the recalibrated features to obtain weighted features x. a The formula is as follows.

[0028] (4) (5) Where x represents the input feature, w represents the channel weights obtained after ECA-Lite and Sigmoid, σ represents the Sigmoid function, and GAP represents global average pooling. ʘ represents ghost convolution, and ʘ represents element-wise multiplication. ECA-lite is a model that uses "GAP channel description + adaptive kernel-length 1D convolution + Sigmoid" to model channel dependencies without full connectivity, and removes the BN and extra activation functions of the original ECA, retaining only a single convolution kernel to further compress parameters and memory access. SE-Lite uses GAP, 1×1 (stride=1) convolution for dimensionality increase / decrease, Sigmoid, etc., to replace the full connectivity of the standard SE.

[0029] Then xa GeM aggregation and GAP with learnable exponent p are used, and their results are summed. Based on this result, BN and Dropout are applied in sequence to obtain the feature vector f, as shown in the following formula.

[0030] (6) (7) Where f gem Indicates x a The global feature vector obtained after GeM aggregation, x a This indicates a weighted feature map that is multiplied channel by channel, where h and w represent the spatial dimensions of the feature map: height and width. This represents a very small constant to prevent numerical underflow, p represents the learnable exponent of GeM, and Dropout represents the random zeroing operation of the feature elements. represents batch normalization, GAP represents global average pooling, and f represents the feature vector obtained after the above operations.

[0031] Next, input the cosine metric classifier and the linear classifier respectively to obtain two logits, as shown in the following formula.

[0032] (8) Z cos This represents the logits output of the cosine metric branch, where s represents the scaling factor used to amplify the cosine score. This indicates that the eigenvector f has been normalized using the L2 norm. W represents the weight matrix of the cosine classifier. c After L2 norm normalization, T represents the transpose operation, m represents the angular margin, and Z... fc W represents the logits of the linear branch output. l Let f represent the weight matrix of the linear classifier, and let f represent the input feature vector.

[0033] Finally, the gating coefficients are obtained through SoftmaxGate and the two logits are fused to obtain the final output F, as shown in the following formula.

[0034] (9) (10) Where Softmax represents a set of real numbers mapped to a probability distribution that sums to 1, and FC1 and FC2 represent two fully connected layers. Z represents the ReLU activation function, g1 and g2 represent the two gating coefficients of the SoftmaxGate output, ʘ represents element-wise multiplication, F represents the final logits after weighted fusion of the two logits according to the gating coefficients, and Z represents the ReLU activation function.cos Z represents the logits of the cosine metric branch output. fc Let represent the logits of the linear branch output, and f represent the input feature vector.

[0035] Compared with the prior art, the present invention has the following advantages: This invention proposes a multi-pose estimation method for chickens based on a three-view collaborative approach and a lightweight TriFuseNet-OSDA model. The model uses an optimized lightweight OverLoCK-xt-light feature extraction network to reduce network parameters, and sequentially introduces a self-developed dynamic multi-branch complementary fusion structure to enhance feature interaction. It also incorporates the invention's designed DynamicTanhECA attention, SRFD-enhanced shallow feature enhancement, and Attn-GeM-Lite lightweight classification head. This invention's model achieves high-precision pose estimation with a file size of 2.36 MB and 0.72 GFLOPs. TriFuseNet-OSDA achieves P-values ​​of 95.89%, R-values ​​of 95.74%, F1 scores of 95.81%, and ACC of 96.40% across eight chicken poses, outperforming existing multi-pose estimation models in all metrics. The eating and resting poses achieve perfect scores, while the preening and flapping poses achieve F1 scores exceeding 98%. Even the walking pose, which has relatively lower pose estimation accuracy, still achieves an F1 score of 87.22%. Module ablation experiments demonstrate that the collaborative use of modules can simultaneously improve accuracy and reduce computational cost without increasing parameters. Viewpoint ablation experiments show that the model's pose estimation accuracy gradually improves with the increase in the number of views, with three-view input contributing 7.07% to the accuracy compared to single-view input. Comparative experiments of the DMCF structure before and after different model introductions demonstrate that this structure can achieve significant model lightweighting with only a slight loss in accuracy. Comparisons of different pose estimation methods show that the method of this invention outperforms other methods in both the types of estimated poses and the accuracy of pose estimation. In summary, the TriFuseNet-OSDA model proposed in this invention achieves a good balance between the types of pose estimation, accuracy, and model lightweighting. This invention provides an efficient, reliable, and easily deployable technical foundation for chicken pose estimation, while also contributing to intelligent farming technology. Attached Figure Description

[0036] Figure 1 This is a diagram of the overall framework of the TriFuseNet-OSDA model in a specific embodiment of the present invention.

[0037] Figure 2 This is a diagram of the DMCF network structure in a specific embodiment of the present invention.

[0038] Figure 3This is a diagram of the OverLoCK-xt-light network structure in a specific embodiment of the present invention.

[0039] Figure 4 This is the Attn-GeM-Lite lightweight classification head structure in a specific embodiment of the present invention.

[0040] Figure 5 This is a structural diagram of the SRFD-enhanced module in a specific embodiment of the present invention.

[0041] Figure 6 This is an example diagram of a specific embodiment of the present invention. Detailed Implementation

[0042] The key to a chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA is to construct a multi-view pose estimation model that achieves both high accuracy and low volume. This mainly includes the following steps: Existing visual chicken pose estimation methods suffer from limitations such as restricted acquisition viewpoints, insufficient pose variety, and low accuracy. This invention proposes the TriFuseNet-OSDA model, which enables multi-view pose estimation while maintaining high accuracy and a lightweight overall structure. The model utilizes a self-developed multi-branch DMCF structure, with an optimized OverLoCK-xt-light backbone network as the basic feature extraction network. The traditional Stage 1 (PatchEmbed1 and Basic block1) is replaced with an optimized lightweight SRFD-enhanced module. This module enhances early texture extraction and orientation awareness while maintaining relatively unchanged parameters through depthwise separable convolutions, coordinate attention, and a hierarchical Cut-Fusion structure. Furthermore, the three inputs (different viewpoints) are processed through the SRFD-enhanced feature extraction module, PatchEmbed2, and Basic block2, respectively, with completely shared parameters, achieving lightweight basic feature extraction. Secondly, in order to improve the model's ability to acquire detailed features, after the model enters Stage 3 (each branch has independent PatchEmbed3 and Dynamic block3), the three features are divided into three parameter-independent branches. At the same time, DynamicTanhECA modules are inserted into each branch to perform fine-grained channel reweighting and dynamic discarding, and the one-dimensional ECA module is used to capture global channel dependencies, highlighting key semantics while suppressing noise. At the end of Stage 3, the cross-branch efficient channel attention (CrossBranch-ECA) module is used to achieve information complementarity and fusion of the three branch features, enhancing the diversity of feature expression and the correlation between branches

[30] . Next, the GateFC module generates fine-grained channel gating coefficients for the features of each branch of Stage 3, and inputs them into Stage 4 (each branch has independent PatchEmbed4 and dynamic block4) after channel weighting. Meanwhile, the Context gate module concatenates the global feature vectors of Stage 2 and Stage 3, and generates dynamic adaptive weights through fully connected mapping, which are used to dynamically recalibrate the channel responses of each branch of Stage 4, further enhancing the high-level semantic expression capability. Subsequently, the three high-level features output by Stage4 are further complemented and integrated between branches through the CrossBranch-ECA module, enhancing the expressive power of the high-level features.Next, the weights of different branches are calculated and weighted concatenated using the branch scalar attention module, and then fed into the Attn-GeM-Lite classification head. Efficient features are extracted using GhostConv and ECA-Lite, and then enhanced by fusing learnable generalized mean pooling and global average pooling. The fused features are normalized and Dropout-processed before being input into the CosineClassifier and a regular fully connected classifier, respectively. Finally, the models are dynamically fused using SoftmaxGate to obtain their multi-class prediction results. The model structure is as follows. Figure 1 As shown.

[0043] The overall design of the DMCF architecture adopts a strategy of "shared low-level features + independent high-level branches," enabling the model to achieve both high accuracy and lightweight efficiency. In the low-level stage, the model primarily captures basic texture information, edges, colors, and simple shapes from the input image. These features exhibit strong commonality and consistency across different viewpoints, making them suitable for unified extraction using parameter sharing. The model achieves parameter sharing across three different viewpoint inputs during the low-level feature extraction stage, completing the low-level feature encoding through unified feature extraction. Compared to designing independent low-level branches for each viewpoint, parameter sharing significantly reduces the number of parameters in the model while promoting the alignment and consistency of basic information across different viewpoints, avoiding feature distribution drift and redundancy. This design not only improves computational efficiency but also enhances the model's feature representation.

[0044] In the mid- and high-level stages, features gradually shift towards more abstract semantic representations, incorporating complex spatial relationships, local pose details, and viewpoint-specific information. To address this, we divide the features into three parameter-independent branches, granting each viewpoint ample freedom of expression, enabling refined learning and semantic capture tailored to its specific characteristics. Simultaneously, within each branch, a lightweight ECA-Lite channel attention mechanism is introduced to rapidly capture the correlation between feature channels, thereby improving feature selectivity and representational stability. Furthermore, a gated GateFC layer is employed to dynamically generate gating coefficients for each channel using a fully connected structure, effectively suppressing redundant features and highlighting key information. The synergistic effect of these attention and gating modules allows for dynamic adjustment of branch feature responses, achieving enhanced expression of key features and noise suppression, thus improving the compactness and discriminative power of feature representation. To enhance information interaction between different branches, the model introduces the CrossBranch-ECA efficient cross-branch channel attention mechanism in the mid- and high-level feature stages, effectively promoting information exchange and complementarity between viewpoint branches, thereby enhancing feature diversity and correlation. Meanwhile, between different feature levels, we use the Context Gate module to generate adaptive weights by fusing multi-stage global features to dynamically recalibrate the responses of higher-level channels, effectively mitigating semantic drift of higher-level features and ensuring the consistency and stability of multi-scale feature representations. Finally, the three higher-level features interact at the output, and through branch-scale attention, weights are dynamically allocated based on the global representation of each branch to achieve adaptive weighted fusion of multi-view features, ensuring that the contribution of key perspectives in decision-making is fully reflected. Subsequently, the fused unified features are fed into a lightweight and adaptive classification head to complete the discrimination, further integrating local details and global semantics while controlling computational overhead, and generating the final multi-class prediction. This structure is as follows: Figure 2 As shown.

[0045] The OverLoCK-xt-light network operates as follows: Inspired by the human visual system's "overview first, focus later" mechanism, the original OverLoCK network comprises three sub-networks: a base network, an overview network, and a focus network. The base network progressively downsamples the input image through three levels of embedding layers, generating mid-level feature maps and simultaneously providing feature support for the overview and focus networks, reducing overhead. The overview network further downsamples mid-level features, quickly generating global feature maps with semantic information but coarser precision. This context prior serves as feedback input to the focus network, allowing it to incorporate contextual information, progressively refine mid-level features, and expand the receptive field to obtain more accurate high-level representations. Structurally, both the overview and base networks use basic blocks as their basic units. This module includes 3×3 depthwise separable convolutions with residual connections, as well as layer normalization, Dilated RepConv, SE, and ConvFFN. The focus network employs a more complex dynamic block structure, including 3×3 DWConv, GDSA, and ConvFFN. The GDSA module uses ContMix as its core token and utilizes a gating mechanism to remove contextual noise interference.

[0046] Compared to previous methods, the OverLoCK-XT-Light lightweight feature extraction proposed in this invention mainly optimizes the following aspects: To significantly reduce the number of network parameters, we replaced the standard DWConv in the original base and overview networks with the more efficient Ghost-DWConv. Compared to traditional DWConv, Ghost-DWConv uses a main branch of depthwise convolution and a lightweight pointwise convolution ghost branch to generate "pseudo-features," achieving similar expressive power with fewer parameters. Furthermore, by introducing channel width scaling and layer depth scaling factors, the overall model complexity is reduced, further decreasing the network's computational cost.

[0047] (1) Among them, w m d represents the width scaling factor. m Pa represents the depth scaling factor. original Pa represents the number of parameters in the original DWConv module. light This represents the lightweight parameters after compression.

[0048] For the initial global context prior P0 generated by the original overview network, we designed a more efficient channel compression and spatial matching mechanism, the specific formula of which is as follows.

[0049] (2) Here, Compress represents channel compression, and Upsample represents spatial upsampling to match the features of the focus network input resolution, thereby further reducing the computational cost of fusion. Indicates original features, This indicates the updated features.

[0050] In the Dynamic Block of the Focus network, we streamline and optimize the original GDSA by employing a smaller spatial mixing kernel to further reduce the computational burden of dynamic convolution. Specifically, the context prior and feature map of the i-th module of the Focus network are labeled as P. i and Z i Before entering the network, the two are spliced ​​and fused, and then the dynamic gating G is calculated. i Adjusting the feature map allows for refined guidance of the feature layer.

[0051] (3) (4) (5) Among them, P i Z represents the context prior feature map. i Feature map representing the main trunk, Indicates channel splicing operation, U i The Conv represents the features formed by the fusion and splicing operation. 1x1 G represents a 1x1 convolution operation, σ represents the SiLU activation function, and G... i Branch(Z) represents the dynamic gating of the i-th layer. i ) represents the original feature Z i Apply a parallel convolution branch, This represents the refined features.

[0052] On the other hand, to achieve guidance at the weight level, context prior P i It will be injected into the dynamic convolution, utilizing P through a simplified ContMix mixer. iTo generate dynamic convolutional kernel weights, the fused feature map is re-split into context prior and feature map before the module output. The split feature maps are treated as decoupled and updated context prior and feature map, respectively. Simultaneously, to prevent dilution of the context prior, the initial context prior P0 is added back to the updated feature map. The formula is as follows.

[0053] (6) Where α and β are learnable scalars, initialized to 1 before training, and P0 is the original context prior formed after channel compression and spatial upsampling. This method saves computation while making its resolution match the input of the focus network. This represents the updated feature map. This represents the feature map after the operation.

[0054] In classification tasks, predictions in the OverLoCK-xt-light network are always performed by the focus network; in prediction tasks, the base network is utilized. and Features at resolution, and focus network and Features at different resolutions collectively construct the feature pyramid. These four stages of features were chosen because they effectively cover multi-scale information from finer to coarser granularity, encompassing rich spatial details while also taking into account higher-level semantic representations. These four sets of features correspond to stages 1 through 4 of our design, and the network structure is as follows: Figure 3 .

[0055] The principle of the SRFD-enhanced module is as follows: We use an improved and lightweight SRFD-enhanced module to extract preliminary image features. This improved module enhances the model's early texture extraction and orientation awareness capabilities while keeping the parameters essentially unchanged through DWConv, coordinate attention, and a hierarchical Cut-Fusion structure. In the SRFD module, the original feature map x is first processed using a depthwise convolution with a kernel size of 7×7 and a stride of 1. Then, it is sequentially processed through BN, a sigmoid linear unit (SiLU), and a 1×1 (stride=1) convolution to reduce the number of channels to 1 / 4 of the original, and then processed again through BN and SiLU to obtain the initial features. Next, we use coordinate attention to enhance the orientation information and combine DropPath and LayerScale to complete a lightweight residual update, obtaining feature x0. Then, we enter the first downsampling stage. On one hand, x0 is first decomposed into four sub-feature maps c1, c2, c3, and c4 according to adjacent pixels using a cut operation. Then, a concatenation (Concat) operation is used to merge them into a new feature c. On the other hand, for x0, a 2x downsampling is achieved by sequentially applying a 3×3 (stride=1) DWConv, a 1×1 (stride=1) pointwise convolution (PWConv), and a 3×3 (stride=2) DWConv, followed by BN and SiLU activation to obtain feature x1. Then, feature c and x1 are concatted to obtain x. 1c Then, a 1×1 (stride=1) convolution is applied to the concatenated features, along with coordinate attention, drop path, and layer scale to obtain the updated feature map. The formula is as follows.

[0056] (7) (8) DW 3,1 DWConv and PW represent 3×3 (stride=1) 1,1 PWConv,DW represents 1×1 (stride=1) 3,2 DWConv and Conv represent 3×3 (stride=2) 1,1 This represents a 1×1 (stride=1) convolution, CA represents coordinated attention, DP represents DropPath, LS represents LayerScale, and x represents the number of convolutions.1c c1, c2, c3, and c4 represent the features after concatenation, and c4 represent the sub-feature maps after decomposition. This indicates the updated sub-diagnosis diagram. represents the input features, Contact represents the channel concatenation operation, SiLU represents the activation function, and BN represents batch normalization.

[0057] The above describes the transformation from the original feature map x to the feature map after a series of operations. This achieves downsampling from the original sample to 2x. Subsequently, it undergoes downsampling from 2x to 4x, first... The feature r is obtained by performing a Cut operation combined with a Concat operation. Next, ... We obtain the following by sequentially using 3×3 (stride=1)DWConv and 1×1 (stride=1)PWConv. In the The two-branch operation involves sequentially applying 2×2 (stride=2) max pooling and BN to obtain feature m on one branch, and applying 3×3 (stride=2) DWConv to obtain the second downsampled feature n on the other branch. The concatenation operation of n, r, and m yields feature y, which is then processed by 1×1 (stride=1) convolutions, along with operations such as Coordinate Attention, DropPath, LayerScale, lightweight multi-layer perceptron (Light-MLP), and squeeze-and-excitation lite (SE-Lite) to obtain the output feature Y. Light-MLP replaces large-size fully connected / convolutional stacks with two 1×1 (stride=1) convolutions combined with a Gaussian error linear unit (GELU), significantly reducing parameter and memory overhead while maintaining channel mixing and non-linear representation. SE-Lite replaces the fully connected bottleneck of standard SE with operations such as global average pooling (GAP), 1×1 (stride=1) convolution for dimensionality increase / decrease, and Sigmoid, thereby removing redundant activations and bias compression computations while retaining channel recalibration capabilities. The above feature processing formulas are as follows.

[0058] (9) (10) (11) Where LM represents Light-MLP operation, SEL represents SE-Lite, and DW3,1 DWConv and PW represent 3×3 (stride=1) 1,1 PWConv and DW represent 1×1 (stride=1). 3,2 DWConv,Conv represent 3×3 (stride=2) 1,1 This represents a 1×1 (stride=1) convolution, CA represents coordinated attention, DP represents DropPath, LS represents LayerScale, and BN represents batch normalization. Indicates to , Y represents the features after the connection is performed, and Y represents the final feature. This represents the input features.

[0059] In addition, this module includes a residual / shortcut path. The original input feature x is processed through 4x4 (stride=4) average pooling to match the spatial mechanism, then adjusted using 1×1 (stride=1) convolution and BN, and finally added element-wise to the main branch Y to obtain the final feature map Z. This method preserves low-level information and enhances gradient propagation stability. The module structure is as follows: Figure 5 .

[0060] The principle of DynamicTanhECA is as follows: To enhance channel interactions between mid-level features in the model, we designed a lightweight and learnable DynamicTanhECA module. This module enhances the dynamic nonlinear representation of features and the modeling of inter-channel dependencies. The module is based on channel normalization and a parameterized hyperbolic tangent activation function. It achieves efficient channel attention gating, adjustable residual fusion coefficients, and a random dropout mechanism during training through one-dimensional convolution, improving the robustness and expressive power of the model without significantly increasing the number of parameters. The specific process is as follows: First, LN is applied to the input feature x to obtain the normalized feature x. n Then, the learnable vector α and the feature x are used n Element-wise multiplication is performed, and a non-linear output is obtained through Tanh, as shown in the following formula.

[0061] (12) Where y1 represents the nonlinear output obtained after the above operation. This represents the learnable vector obtained by transforming the learnable parameter raw_alpha using the softplus function, where Tanh represents the hyperbolic tangent activation function, and x... n represents the normalized eigenvector, and ʘ represents element-wise multiplication.

[0062] Subsequently, the mean of y1 in the spatial dimension is calculated to obtain the channel feature description vector. Then, it is processed by a 1D convolution with adaptive kernel size and the Sigmoid function to generate dynamic gating coefficients g. y1 is then gated using g to obtain the output y2. This process is the core process of the dynamic ECA gating mechanism. By dynamically adjusting the weights of each channel through adaptive local convolution, the inter-channel dependencies are effectively modeled. The formula is as follows.

[0063] (13) Where g represents the dynamic gating coefficient. Represents the Sigmoid function, and Conv1D is a 1D convolution with the kernel size adaptively based on the number of channels. y1 represents the mean function, y2 represents the non-linear output, y3 represents the output obtained after the above series of operations, and ʘ represents element-wise multiplication.

[0064] Next, a learnable affine transformation is applied to y2 and LN is applied again to obtain y3. Finally, the output is obtained by controlling the residual fusion using learnable coefficients λ∈[0,1] and using the DropPath random discarding mechanism during training. The formula is as follows.

[0065] (14) Where Y represents the final output, x represents the initial input, DP represents DropPath, and y3 represents the output after layer normalization of inter-channel dependencies.

[0066] The DynamicTanhECA module enhances nonlinear expressiveness through a parameterized Tanh activation function, effectively captures inter-channel dependencies by combining a dynamic ECA gating mechanism based on adaptive one-dimensional convolution, and improves training stability and model generalization ability by utilizing adjustable residual fusion and DropPath. The module's design balances feature expressiveness with lightweight design, making the model more robust and efficient in fine-grained feature modeling and complex backgrounds.

[0067] The Attn-GeM-Lite principle is as follows: To address the shortcomings of traditional classification heads, such as difficulty in modeling channel dependencies, lack of discriminative margins and robust pooling, and susceptibility to feature scale fluctuations, we designed the lightweight and learnable Attn-GeM-Lite classification head. This head uses GhostConv for efficient feature generation and channel recalibration, combined with ECA-Lite's adaptive one-dimensional convolutional channel attention, learnable power-law GeM, and a "dual-branch classifier + Softmax gated fusion" strategy, improving the model's robustness and expressive power while maintaining stable model parameters. The specific module flow is as follows: First, GhostConv is used to obtain intermediate features from the input feature x. Then, SE-Lite is used to recalibrate the feature channels to obtain feature x. se The next step is to sequentially perform GAP, ECA-Lite, and Sigmoid on the recalibrated features to obtain channel weights w, and then multiply them channel by channel with the recalibrated features to obtain weighted features x. a The formula is as follows.

[0068] (15) (16) Where w represents the channel weights obtained after ECA-Lite and Sigmoid, σ represents the Sigmoid function, and GAP represents global average pooling. ʘ represents ghost convolution, and ʘ represents element-wise multiplication. ECA-lite is a model that uses "GAP channel description + adaptive kernel-length 1D convolution + Sigmoid" to model channel dependencies without full connectivity, and removes the BN and extra activation functions of the original ECA, retaining only a single convolution kernel to further compress parameters and memory access. SE-Lite uses GAP, 1×1 (stride=1) convolution for dimensionality increase / decrease, Sigmoid, etc., to replace the full connectivity of the standard SE, where x represents the input feature.

[0069] Then x a GeM aggregation and GAP with learnable exponent p are used, and their results are summed. Based on this result, BN and Dropout are applied in sequence to obtain the feature vector f, as shown in the following formula.

[0070] (17) (18) Where f gem Indicates x a The global feature vector obtained after GeM aggregation, x a This indicates a weighted feature map that is multiplied channel by channel, where h and w represent the spatial dimensions of the feature map: height and width. This represents a very small constant to prevent numerical underflow, p represents the learnable exponent of GeM, and Dropout represents the random zeroing operation of the feature elements. represents batch normalization, GAP represents global average pooling, and f represents the feature vector obtained after the above operations.

[0071] Next, input the cosine metric classifier and the linear classifier respectively to obtain two logits, as shown in the following formula.

[0072] (19) Z cos This represents the logits output of the cosine metric branch, where s represents the scaling factor used to amplify the cosine score. This indicates that the eigenvector f has been normalized using the L2 norm. W represents the weight matrix of the cosine classifier. c After L2 norm normalization, T represents the transpose operation, m represents the angular margin, and Z... fc W represents the logits of the linear branch output. l Let f represent the weight matrix of the linear classifier, and let f represent the input feature vector.

[0073] Finally, the gating coefficients are obtained through SoftmaxGate and the two logits are fused to obtain the final output F, as shown in the following formula.

[0074] (20) (twenty one) Where Softmax represents a set of real numbers mapped to a probability distribution that sums to 1, and FC1 and FC2 represent two fully connected layers. Z represents the ReLU activation function, g1 and g2 represent the two gating coefficients of the SoftmaxGate output, ʘ represents element-wise multiplication, F represents the final logits after weighted fusion of the two logits according to the gating coefficients, and Z represents the ReLU activation function. cos Z represents the logits of the cosine metric branch output. fc The output logits of the linear branch are represented by f, and the input feature vector is represented by f. The module structure is as follows: Figure 4 Pose model estimation example Figure 6 .

Claims

1. A chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA, characterized in that, Includes the following steps: 1) Use multi-view depth cameras in a multi-camera system to collect chicken video data and form a raw dataset of chicken videos; 2) Extract RGB images from three perspectives from the original chicken video dataset and construct an RGB image dataset; 3) Construct the TriFuseNet-OSDA model; The TriFuseNet-OSDA model includes: Input terminal; The shallow feature enhancement module is connected to the input terminal; The feature extraction network connected to the shallow feature enhancement module; A dynamic multi-branch complementary fusion structure connected to the feature extraction network; The classification head is connected to the dynamic multi-branch complementary fusion structure; The output terminal connected to the classification head; 4) The TriFuseNet-OSDA model is trained using an RGB image dataset. After training, the RGB images of chickens from three perspectives to be detected are input into the trained TriFuseNet-OSDA model to estimate the multiple poses of the chickens and output the chicken pose classification results.

2. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 1, characterized in that, In the shallow feature enhancement module, the feature extraction stage replaces the Conv convolution module with the DWConv module and adds a Coordinate Attention module. In the downsampling stage, the hierarchical Cut-Fusion structure replaces the single cascaded downsampling structure.

3. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 1, characterized in that, The feature extraction network adopts the OverLoCK-xt framework, which includes three parts: base, overview, and focus. The DWCony convolutional modules in the base and overview parts are replaced with Ghost-DWConv convolutional modules, and channel compression and upsampling alignment operations are added to the overview part.

4. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 1, characterized in that, In step 3), the dynamic multi-branch complementary fusion structure includes: A feature extraction layer that shares basic features is used to obtain basic features; A feature extraction layer that separates complex features and is connected to the feature extraction layer that shares basic features is used to obtain complex features; A complementary layer, wherein the basic features and complex features are exchanged and complemented to obtain three-branch feature information; The fusion layer combines the feature information from the three branches to form a fused feature output.

5. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 4, characterized in that, The basic features include: basic texture information, edges, colors, and simple shapes.

6. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 4, characterized in that, The complex features include complex spatial relationships, local posture details, and viewpoint-specific information.

7. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 4, characterized in that, The complementary layer exchanges and complements the basic features and complex features to obtain three-branch feature information, specifically including: 3.1) Weighted optimization is performed on complex features to obtain features after removing redundancy; 3.2) The redundant features are then processed through the interaction of each branch to obtain complete branch feature information; 3.3) The complete branch feature information is interactively weighted with the basic features to obtain the three-branch feature information.

8. The chicken multi-pose estimation method based on a multi-view imaging system and TriFuseNet-OSDA according to claim 1, characterized in that, The classification head is used to classify and output the fused features, specifically including: First, the fused features are subjected to GhostConv convolution and channel recalibration. Then, channel weighting is performed through global statistics. Subsequently, a global representation is obtained by learningable generalized mean and average aggregation. Finally, the classification result is output by gating fusion of cosine metric classifier and linear classifier.

Images