Animal three-dimensional attitude estimation method based on feature screening

Abstract

The invention discloses an animal three-dimensional attitude estimation method based on feature screening, and relates to the field of computer vision and intelligent perception, and the method comprises the following steps: obtaining animal activity multi-view video frames, and generating a three-dimensional volume grid; the method comprises the following steps: constructing an improved 3D U-Net network with an encoder-decoder structure, and carrying out pre-training; inputting the three-dimensional volume grid into the trained network to generate an accurate three-dimensional coordinate of the animal posture; the improved network is specifically characterized in that a context-aware feature screening gate module is arranged on each jump connection path connecting encoder shallow features and decoder deep features; replacing a standard 3D convolution module in the 3D U-Net network encoder with a 3D aggregation transformation residual module; according to the method, the animal three-dimensional attitude estimation precision is improved, and the method can be widely applied to animal behavioristics, neuroscience, ecology and related experimental studies.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent perception, and in particular relates to a method for animal three-dimensional pose estimation based on feature selection. Background Technology

[0002] In multiple scientific fields such as animal behavior, neuroscience, and ecology, the precise quantitative analysis of animal behavior is the cornerstone of research advancement. Since animal behavior is primarily manifested through its whole-body movements, developing a universal method capable of accurately measuring animal three-dimensional (3D) kinematics is of paramount importance. Compared to traditional two-dimensional (2D) tracking techniques, 3D pose estimation provides more comprehensive motion information, overcoming data gaps caused by viewpoint limitations and occlusion. This provides crucial data support for revealing animal movement control mechanisms, diagnosing pathological behaviors, and supporting applications in wildlife conservation. Currently, in the field of computer vision, the mainstream technical approaches for achieving label-free 3D animal pose estimation fall into two main categories. The first category is a multi-stage method based on 2D pose triangulation. This method is one of the most widely used solutions, and its technical origins can be found in various open-source toolkits, such as DeepLabCut proposed by Mathis et al. Its working principle consists of two core steps: First, using 2D convolutional neural networks (2D CNNs), images from each viewpoint are processed independently in video frames acquired by multiple synchronous cameras to detect and locate the 2D coordinates of predefined body keypoints. Then, in the post-processing stage, using pre-calibrated camera intrinsic and extrinsic parameters, triangulation algorithms are used to geometrically reconstruct the 2D coordinates from at least two views, thereby calculating the 3D spatial coordinates of the keypoint. The advantage of this method is its intuitive flow, ease of implementation, and ability to utilize mature 2D pose estimation algorithms. However, this technical solution has fundamental limitations. Its core problem lies in the fact that the feature extraction stages of each camera view are independent, failing to effectively integrate cross-view image information to resolve inherent geometric ambiguity. This independence makes the method particularly vulnerable to occlusion problems commonly encountered during free animal movement. When a key point is occluded due to the animal's posture or environmental objects, resulting in its visibility in fewer than two views, triangulation fails, directly causing the loss of 3D data at that time point or introducing serious errors, thus severely affecting the integrity and accuracy of the final motion trajectory. Although subsequent studies have optimized the results by introducing skeleton constraints and spatiotemporal filters (such as Anipose proposed by Karashchuk et al.), these methods still compensate after 2D prediction and do not fundamentally solve the problem of insufficient information fusion. To overcome the aforementioned limitations, a second, more advanced technological paradigm has emerged in recent years: end-to-end volumetric deep learning methods. A representative framework for this technology is DANNCE, proposed by Dunn et al. in 2021. Its core idea is to fuse information from all camera views before pose prediction. Specifically, it utilizes projection geometry to back-project 2D image information from multiple perspectives, collectively constructing a geometrically aligned 3D feature volume. This 3D volume is then directly input into a 3D convolutional neural network (3DCNN) for end-to-end training and prediction. This design allows the 3D CNN to simultaneously utilize shared features from different views and learned animal pose spatial statistics to directly infer the location of keypoints in 3D space. Because the information is fused at the feature level, this method can perform better geometric inference, thus possessing a natural robustness advantage in handling occlusion and viewpoint changes, with a significantly improved average accuracy compared to triangulation methods. Although volumetric 3D CNN methods represent the current technological forefront, they still face a series of challenges that urgently need to be addressed in practical applications. First, 3D convolutional neural networks are extremely computationally and memory-intensive. As the data dimension expands from two-dimensional images to three-dimensional volumes, the computational cost and number of parameters in the network increase cubically. This characteristic severely restricts the development of deeper and wider network models, limiting further performance improvements. Finally, and most importantly, the standard 3D CNN architecture, which forms the cornerstone of these methods, has inherent design flaws. Currently, these methods commonly employ an encoder-decoder structure, exemplified by 3D U-Net (3D U-shaped convolutional neural network). However, this structure is not specifically optimized for animal pose estimation tasks involving rapid movement and complex postures. First, standard convolutional modules in the network (such as conventional 3D residual blocks) have bottlenecks in feature extraction efficiency and representational capabilities, making it difficult to effectively capture the fine spatiotemporal dynamic features exhibited by fast-moving animals during complex and rapid movements such as jumping and grasping. Second, while the skip connections mechanism in 3D U-Net aims to fuse deep semantic information with shallow detail information, its connection method is indiscriminate. This means that features output by the encoder, whether crucial for keypoint localization—fine anatomical details—or a large amount of irrelevant background textures, lighting variations, and experimental artifacts, are passed to the decoder without filtering. The presence of this noise interferes with the decoder's accurate parsing of features, thus affecting the accuracy and robustness of the final keypoint localization. Summary of the Invention

[0003] The purpose of this invention is to address the problems in existing animal pose estimation algorithms based on three-dimensional convolutional neural networks (3D CNNs), such as insufficient feature extraction capability of standard convolutional modules and the easy introduction of background noise when the network fuses features from different levels. This invention proposes a feature-based method for animal three-dimensional pose estimation.

[0004] This invention is achieved through the following technical solution: A feature-based method for estimating animal 3D pose includes the following steps: S1. Data Acquisition: Acquire multi-view video frames of animal activities and generate a three-dimensional volumetric mesh; S2. Network Construction and Training: Construct an improved 3D U-Net network with an encoder-decoder structure and pre-train it to obtain a trained network; S3. Animal pose estimation: Input the three-dimensional volume grid into the trained network to generate accurate three-dimensional coordinates of the animal pose.

[0005] A storage device that stores instructions and data for implementing a feature-based animal three-dimensional pose estimation method.

[0006] A feature-based animal 3D pose estimation device includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a feature-based animal 3D pose estimation method.

[0007] The present invention has the following beneficial effects: 1. By systematically deploying context-aware feature filtering modules at each level of the U-Net decoder path, the entire network can more efficiently integrate multi-scale information through a coarse-to-fine, hierarchical filtering and focusing process, ultimately generating more accurate and robust 3D pose predictions.

[0008] 2. The aggregation transform residual module architecture successfully combines the concise modularity of ResNet with the advantages of multi-path architecture. By introducing new hyperparameters, it provides a more efficient way to improve model performance than simply increasing depth or width, without incurring prohibitive computational resource overhead. For tasks such as complex 3D animal pose estimation, this means the model is capable of capturing more refined and complex spatiotemporal dynamic information, thereby fundamentally improving the accuracy and depth of analysis. Overall, this invention surpasses existing methods in overall accuracy, improving the precision of animal 3D pose estimation. 3. The method of this invention can be widely applied in animal behavior, neuroscience, ecology, and related experimental research for three-dimensional tracking and quantitative analysis of key points in animal movement under non-contact conditions. In addition to scientific research applications, this method also has potential applications in disease model behavior detection, drug efficacy evaluation, intelligent breeding monitoring, and wildlife conservation. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the process of a feature-based animal 3D pose estimation method according to the present invention; Figure 2 This is a schematic diagram illustrating the implementation process of the present invention; Figure 3 This is a schematic diagram of an improved 3D U-Net network architecture with an encoder-decoder structure; Figure 4 This is a schematic diagram of the context-aware feature filtering gate module; Figure 5 This is a schematic diagram of the three-dimensional aggregation transformation residual module; Figure 6 It is a three-dimensional pose diagram of an animal (mouse); Figure 7 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Example 1 Please see Figures 1-7 As shown, this invention is a method for animal three-dimensional pose estimation based on feature selection, comprising the following steps: S1. Data Acquisition: Acquire multi-view video frames of animal activities and generate a three-dimensional volumetric mesh; It should be noted that step S1 is as follows: S11. Construct a multi-view vision system and calibrate the intrinsic and extrinsic parameters of each camera in the multi-view vision system; S12. Use a multi-view vision system to acquire multi-view video frames of animal activities; As one example, a multi-view vision system consisting of six cameras is first constructed, and the intrinsic and extrinsic parameters of each camera are precisely calibrated. During data acquisition, multi-view video frames of the animal within its activity space are captured simultaneously. Camera calibration can be performed using the traditional Zhang Zhengyou calibration method.

[0013] S13. Based on multi-view video frames, construct a three-dimensional volumetric mesh using the principles of projection geometry.

[0014] Specifically, for each voxel coordinate in three-dimensional space Through the camera intrinsic and extrinsic parameter matrix Calculate its in the first Projected coordinates on the camera image plane This process also requires modeling and correcting lens distortion. Finally, by sampling RGB values ​​from the corresponding pixel positions of each camera view and concatenating these values, a feature vector is generated for each voxel, thus constructing a geometrically aligned 3D feature volume. To ensure computational efficiency and resolution, this 3D volume is dynamically centered around the animal's 3D centroid.

[0015] S2. Network Construction and Training: Construct an improved 3D U-Net network with an encoder-decoder structure and pre-train it to obtain a trained network; It should be noted that in step S2, the three-dimensional volumetric mesh is input into an optimized 3D U-Net network. The basic architecture of this network is an encoder-decoder structure, which combines local details and global contextual information through a series of downsampling and upsampling operations.

[0016] The core improvement of this invention, namely the improved 3D U-Net network with an encoder-decoder structure, is mainly reflected in the following two aspects: On each hop connection path connecting shallow features of the encoder and deep features of the decoder, deploy a context-aware feature filtering gate module; Replace the standard 3D convolution module in the 3D U-Net network encoder with a 3D aggregation transform residual module.

[0017] The present invention will now be described in detail.

[0018] First is the context-aware feature filtering gate module. The specific processing flow within this module is as follows: S211. Obtain the feature map from the l-th shallow layer of the encoder. From the deep layer of the decoder +1 layer upsampled feature map, i.e. g gate signal g; It should be noted that the context-aware feature filtering gate module receives two feature maps from different parts of the network as input: Signals to be filtered ( This is from the encoder's... The feature map of the layer is the traditional skip connection signal. It has high spatial resolution and preserves rich low-level details (such as edges and textures), but lacks high-level semantic context.

[0019] Gating signal ( This is from the decoder. The +1 layer (i.e., a deeper layer) is an upsampled feature map. It has lower spatial resolution but has undergone deeper convolutional processing, thus containing richer semantic information and contextual knowledge. This signal will serve as a "query" or "instruction," guiding the context-aware feature selection gate module to... Which regions are we focusing on?

[0020] S212, Use two independent convolutional layers respectively. and ,right and The transformed feature map is obtained through processing. and gating signals The transformed feature map and gating signals The sums are then passed through a ReLU activation function to obtain a feature map that has undergone non-linear activation. Specifically, to enable interaction between two input signals, they are first subjected to a linear transformation. This is achieved using two separate convolutional layers. and ,right and This step involves processing two devices that may have different numbers of channels (…). and The feature map of ) is projected onto a unified dimensional map. In the intermediate feature space, so that subsequent element-level operations can be performed. Typically, It will be set as To reduce computational complexity.

[0021] After performing a linear transformation, the transformed gate signal Upsampling to The two transformed feature maps are then element-wise added together, sharing the same spatial dimension. This addition operation fuses the information from the two paths. The fused result is then passed through a ReLU (Rectified Linear Unit) activation function to introduce non-linearity and capture the more complex relationships between them.

[0022] S213. Input the non-linearly activated feature map into another convolutional layer for dimensionality reduction, and then pass it through the Sigmoid activation function to obtain the generated selection weight map. ; It should be noted that, next, the feature map after non-linear activation is input into another convolutional layer. The function of this convolutional layer is to increase the number of channels from... Dimensionality reduced to 1. Then, this single-channel feature map is processed through a... Activation function . The function's output range is (0,1), therefore, the final generated tensor This is the required spatial filtering weight map. Each element in the map represents the input feature map. The score represents the importance or relevance of a spatial location. A score closer to 1 indicates a greater relevance of the feature at that location to the current task; a score closer to 0 indicates less relevance. The entire generation and filtering weight map is then used. The process can be summarized by the following formula:

[0023] in Represents the ReLU activation function. This represents the Sigmoid activation function.

[0024] S214. Use the generated filtering weight map α to filter the feature map. Resampling is performed to obtain the final high information density feature map. .

[0025] Finally, the generated filter weight map The spatial dimensions of the feature are adjusted using a resampler operation to match those of the original skip connection feature. If they are completely identical, then perform element-wise multiplication:

[0026] This multiplication operation is the core of the context-aware feature filtering gate. It's equivalent to filtering the weight map. For the original feature map A "soft weighting" or "gating" was performed. Spatially, Regions with higher median values ​​(i.e., regions considered important by the network) The characteristics will be preserved; and Regions with low median values ​​(i.e., irrelevant background or noise regions) The characteristic of will be suppressed, and its value will approach zero.

[0027] Final output This is a "cleaned" feature map with higher information density. It retains high-resolution details from the original skip connections while effectively filtering out irrelevant information. This filtered feature map... It will be used to replace the original. This is then concatenated with the upsampled features of the decoder, thus providing a higher quality input for subsequent pose reconstruction.

[0028] In this invention, the context-aware feature filtering gate module is systematically deployed at each layer of the U-Net decoder path, specifically before each skip connection and upsampled feature concatenation. This hierarchical deployment allows the network to adaptively filter features at different semantic levels: In deeper layers (near the bottleneck layer): the gate signal g has the strongest semantic information and the largest receptive field. The context-aware feature filtering gate learns to focus on coarse, global target regions, performing macroscopic foreground-background separation. In shallower layers (near the output layer): the gate signal g has already incorporated some detailed information. The context-aware feature filtering gate learns to focus on finer structures, such as the boundaries of specific limbs and the precise locations of joints, performing finer-grained feature filtering. Through this coarse-to-fine, hierarchical filtering and focusing process, the entire network can more efficiently integrate multi-scale information, ultimately generating more accurate and robust 3D pose predictions.

[0029] Another key improvement is the 3D aggregation transform residual module, which is an improved structure based on ResNet. The data processing flow of the 3D aggregation transform residual module is as follows: The input tensor of the S221, 3D aggregation transformation residual module is reduced in dimensionality by a 1×1×1 convolutional layer to obtain a low-dimensional feature map; Specifically, this step, also known as "bottleneck projection (Conv3D(1x1x1))," involves the input tensor first passing through a 1x1x1 convolutional layer. The main function of this layer is "dimensionality reduction," significantly reducing the number of channels in the feature map without changing the spatial dimensions. This step is crucial for achieving computational efficiency because it prepares a "narrower" input for subsequent, more computationally intensive convolutional operations.

[0030] S222. The low-dimensional feature map is fed into a 3×3×3 grouped convolutional layer for parallel processing, with the number of groups being C. Specifically, this step, also known as "Convergence Transformation (Conv3D(3x3x3, groups=C))," is the core of the Convergence Transformation residual module. The low-dimensional feature map output from the bottleneck layer is fed into a 3x3x3 grouped convolutional layer. The `groups=C` parameter here is crucial; it instructs the layer to perform grouped convolution operations, where the number of groups equals the cardinality C. This means that the input channels are divided into C groups, and each group is processed in parallel by C independent 3x3x3 convolutional kernels.

[0031] S223. The outputs of each grouped convolutional layer are concatenated and then input into the second 1×1×1 convolutional layer to perform dimensionality upscaling, resulting in an upscaled feature map. Specifically, this step, also known as "Expanded Projection (Conv3D(1x1x1))", involves passing the output of the grouped convolutions (consisting of the results from each group) through a second 1x1x1 convolutional layer. This layer acts as the opposite of the first step, performing a "dimensionality increase" operation, expanding the number of channels in the feature map back to its original width. This step ensures that the final output of the transformed path has the same dimension as the input tensor from the skip connections, allowing for element-wise addition.

[0032] S224. The upgraded feature map is added to the original input tensor through a skip connection to obtain the residual learning result of the 3D aggregation transformation residual module.

[0033] Finally, residual summation is performed. The output after the aggregation transform is added to the original input through a skip connection to complete residual learning. The overall transformation of the aggregation transform residual module can be precisely mathematically represented. The residual function in the module is no longer a single transformation, but... independent transformation functions The sum of the outputs. Among them, It is the cardinality. Representing the There are several transformation paths. In the bottleneck design of the aggregation transformation residual module, each It is itself a small network, typically consisting of a sequence of "1x1 convolution -> 3x3 convolution -> 1x1 convolution". Combining this aggregation transform with residual connections, the final output of the entire aggregation transform residual module... It can be represented as:

[0034] This formula clearly summarizes the core idea of ​​the Aggregate Transform Residual Module: constructing a powerful high-order feature representation by aggregating multiple isomorphic, low-complexity transformations. The Aggregate Transform Residual Module architecture successfully combines the concise modularity of ResNet with the advantages of multi-path architecture, and by introducing new hyperparameters, provides a more efficient way to improve model performance than simply increasing depth or width. The core implementation is grouped convolution, which not only brings significant improvements in computational and parameter efficiency, but more importantly, it encourages the network to learn richer and more decoupled feature representations by forcing channel grouping.

[0035] These advantages are further amplified when extending from 2D image analysis to 3D volumetric data analysis. The computational and memory consumption of 3D convolutions is inherently far greater than that of 2D convolutions, with costs increasing cubically with data dimensionality. Therefore, the aggregation transform residual architecture, through its bottleneck design and the efficiency gains from grouped convolutions, enables the design and training of deeper and more powerful 3D neural networks without incurring prohibitive computational overhead. For tasks such as complex 3D animal pose estimation, this means the model is capable of capturing more refined and complex spatiotemporal dynamics, fundamentally improving the accuracy and depth of the analysis.

[0036] Furthermore, the pre-training in this invention employs supervised learning for end-to-end network training. Specifically, the network-predicted 3D coordinates are compared with the actual 3D coordinates obtained through manual annotation, and the Euclidean distance error between them is calculated using the standard L2 loss function. Through backpropagation, all learnable parameters in the network (including the weights in the 3D aggregation transformation residual module and the context-aware feature selection gate module) are updated based on this loss value. The goal is to minimize the difference between the predicted and actual coordinates, thereby enabling the trained model to accurately locate key points of the animal from the input 3D volume.

[0037] S3. Animal pose estimation: Input the three-dimensional volume grid into the trained network to generate accurate three-dimensional coordinates of the animal pose.

[0038] It should be noted that the improved 3D U-Net network outputs a three-dimensional probability distribution map, i.e., a three-dimensional heatmap G, for each key point to be tracked.

[0039] To regress accurate 3D coordinates with sub-voxel resolution from this probability distribution, a spatial Soft-Argmax operation is applied. This operation obtains the continuous 3D coordinates of each keypoint by calculating the spatial expectation of the heatmap. The calculation formula is as follows: .in It is the first 3D heatmap of key points.

[0040] This method is fully differentiable and provides smoother, more accurate localization results compared to directly taking the maximum point (hard-argmax), and helps stabilize the model training.

[0041] Example 2 This experiment uses the publicly available dataset Rat7M for training and prediction, and compares the results with DLC-3D, DANNCE, and DANNCE-T. The experimental results are shown in Table 1, reporting the performance of the time-related model on the Rat7M dataset.

[0042] Table 1. Schematic diagram of experimental results of the present invention

[0043] As can be seen, this scheme achieves optimal performance in all three core metrics: mean joint position error (MPJPE), mean error after Procrustes alignment (PA-MPJPE), and mean joint velocity error (MPJVE). Specifically: MPJPE: The MPJPE of this solution is 9.87, which is about 4.7% lower than Dannce's 10.36, about 3.5% lower than Dannce-T's 10.23, and a significant 51.0% lower than DLC-3D's 20.14. This indicates that this solution has a significant advantage in absolute positioning accuracy.

[0044] PA-MPJPE: After removing translation, rotation, and scale differences, this scheme achieves a PA-MPJPE score of 7.83, which is superior to Dannce (8.28) and Dannce-T (8.22), and is an improvement of approximately 53.0% compared to DLC-3D's 16.68. This indicates that this scheme also has a stronger capability in reconstructing the relative positional relationships (shape structure) of joints.

[0045] MPJVE: This scheme performs best in terms of timing consistency metrics, with an MPJVE of 1.84, which is 10.2% lower than Dannce's 2.05, 11.1% lower than Dannce-T's 2.07, and a significant 40.1% lower than DLC-3D's 3.07. This result verifies that this scheme has a significant effect on reducing inter-frame jitter and improving the smoothness of predicted trajectories.

[0046] In summary, this solution surpasses existing mainstream methods across all core evaluation dimensions, fully demonstrating the effectiveness and superiority of the proposed improvement strategy.

[0047] Example 3 Please see Figure 7 , Figure 7 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: an animal three-dimensional pose estimation device 401 based on feature screening, a processor 402, and a storage device 403.

[0048] A feature-based animal 3D pose estimation device 401: The feature-based animal 3D pose estimation device 401 implements the feature-based animal 3D pose estimation method.

[0049] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the animal three-dimensional pose estimation method based on feature screening.

[0050] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the animal three-dimensional pose estimation method based on feature screening.

[0051] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0052] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for animal 3D pose estimation based on feature screening, characterized in that: The method comprises the following steps: S1, data acquisition: acquiring multi-view video frames of animal activity and generating a three-dimensional volume grid; S2, network construction and training: constructing an improved 3D U-Net network with an encoder-decoder structure and pre-training to obtain a trained network; S3, animal posture estimation: inputting the three-dimensional volume grid into the trained network to generate three-dimensional coordinates of animal posture. 2.The animal three-dimensional pose estimation method based on feature screening according to claim 1, wherein, Step S1 is specifically as follows: S11, constructing a multi-view vision system and calibrating the internal and external parameters of each camera in the multi-view vision system; S12, acquiring multi-view video frames of animal activity by using the multi-view vision system; S13, constructing a three-dimensional volume grid based on the multi-view video frames and using the principle of projection geometry. 3.The animal three-dimensional pose estimation method based on feature screening according to claim 1, wherein, The improved 3D U-Net network with an encoder-decoder structure is specifically as follows: A context-aware feature screening gate module is deployed on each skip connection path connecting the shallow features of the encoder and the deep features of the decoder; The standard 3D convolution module in the encoder of the 3D U-Net network is replaced by a 3D aggregated transformation residual module.

4. The animal three-dimensional pose estimation method based on feature screening according to claim 3, characterized in that, The processing flow in the context-aware feature screening gate module is specifically as follows: S211、obtaining the feature map from the lth layer of the shallow layer of the encoder , the feature map from the lth layer of the deep layer of the decoder +1 after up-sampling, i.e., the gating signal g; S212, two independent convolution layers are used respectively, and processed to obtain a transformed feature map and processed to obtain a transformed feature map and a gating signal , the transformed feature map and the gating signal are added and then pass through a ReLU activation function to obtain a nonlinearly activated feature map; S213, input the nonlinear activated feature map to another convolution layer for dimension reduction, and pass through a Sigmoid activation function to obtain a generated screening weight map ; S214, generating a screening weight map a to the feature map Resampling is performed to obtain the final output high information density feature map .

5. The animal three-dimensional pose estimation method based on feature screening according to claim 3, characterized in that, The data processing flow of the 3D aggregated transformation residual module is as follows: S221, the input tensor of the 3D aggregated transformation residual module is reduced in dimension by a 1×1×1 convolution layer to obtain a low-dimensional feature map; S222, the low-dimensional feature map is sent to a 3×3×3 grouped convolution layer for parallel processing, and the grouping number is C; S223, the outputs of each grouped convolution layer are spliced and then input into a second 1×1×1 convolution layer for dimension increasing to obtain a dimension-increased feature map; S224, the dimension-increased feature map and the original input tensor are added through a skip connection to obtain the residual learning result of the 3D aggregated transformation residual module.

6. The animal 3D pose estimation method based on feature screening according to claim 1, wherein: The three-dimensional coordinates of animal posture in step S3 are specifically as follows: wherein is a three-dimensional heat map of the key points.

7. The animal three-dimensional pose estimation method based on feature screening according to claim 1, characterized in that, The pre-training in step S2 adopts a supervised learning method to perform end-to-end training on the network.

8. A storage device, characterized by: The storage device stores instructions and data for implementing the animal three-dimensional posture estimation method based on feature screening according to any one of claims 1-7.

9. An animal three-dimensional pose estimation device based on feature screening, characterized by: It comprises: A processor and a storage device; the processor loads and executes the instructions and data in the storage device to implement the animal three-dimensional posture estimation method based on feature screening according to any one of claims 1-7.

Images