CN117912116B - Wild animal posture estimation method in real scene - Google Patents

Abstract

The invention discloses a wild animal posture estimation method in a real scene, and belongs to the technical field of wild animal monitoring and protection. The invention aims to estimate animal gestures in a complex environment, can be suitable for most real scenes, effectively avoids repeated training of a data set, and specifically comprises the following steps: s1, constructing a data set of an animal posture estimation image, and processing the constructed data set based on style migration; s2, based on the group whitening operation, constructing a free simple baseline attitude estimation model based on heat map generation, generating a heat map by using the model, and completing model training by using the heat map; s3, correcting the model, and designing a coordinate representation method and a heat map code method; s4, adopting a light pose estimation network decoding architecture to finish estimation of the pose of the wild animal in the real scene. The invention takes the wild animal posture estimation method of deep learning as a main body, and finally can realize the monitoring, investigation and protection work of the wild animal in the complex environment.

Description

A method for estimating wild animal posture in real scenes

Technical Field

The present invention relates to the technical field of wildlife monitoring and protection, and in particular to a method for estimating the posture of a wild animal in a real scene.

Background technique

Wildlife monitoring and protection are of great significance. By estimating animal posture, we can better understand the behavior and habits of wild animals and promote the protection of wild animals. At present, the method of estimating wild animal posture is limited by the single scene and species. In the actual complex environment, the performance of the animal posture estimation model is seriously reduced, which limits the actual application of this technology and hinders the efficient implementation of wildlife protection.

At present, the field of ecological research still relies on field workers to collect relevant data on wild animals, and to understand the survival status of wild animals in the current area through expert analysis of the data. With the development of technology, the ability to collect data has been greatly improved by deploying infrared triggered cameras and wireless sensor networks in the wild to monitor wild animals. At the same time, with the development of image recognition technology based on deep learning, the timeliness of wild animal monitoring has been greatly improved, and the ability to protect animal diversity has also been greatly improved. However, the common wildlife monitoring and protection methods currently on the market have not yet fully exerted the ability of visual mining, and lack more detailed information exploration on the characteristics of biodiversity; in view of this, the present invention proposes a method for estimating the posture of wild animals in real scenarios.

Summary of the invention

The purpose of the present invention is to provide a method for estimating the posture of wild animals in real scenarios to solve the technical problems raised in the background technology, realize automatic estimation of the posture of wild animals in different real outdoor scenarios, correct the quantization error of the wild animal posture estimation model, and design a lightweight strategy for the wild animal posture estimation model.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for estimating the posture of wild animals in a real scene comprises the following steps:

S1, constructing a data set of animal posture estimation images, and processing the constructed data set;

S2. Based on the group whitening operation, construct a free simple baseline pose estimation model based on heat map generation, generate a heat map using the model, and complete model training using the heat map;

S3, modifying the model constructed in S2, designing a coordinate representation method and a heat map decoding method;

S4. Combining the operations described in S1 to S3, a lightweight posture estimation network decoding architecture is used to complete the estimation of the posture of wild animals in real scenes.

Preferably, the processing of the data set in S1 specifically includes the following contents:

The public dataset is divided into a test set and a training set. Common wild environmental factors and style transfer factors in real scenarios are added to the test set to simulate complex environmental scenarios in order to observe the distribution differences between the test set and the training set in different scenarios.

Preferably, S2 specifically includes the following contents:

Using non-batch normalization and group whitening methods, and adding squeezing and extraction modules, a free simple baseline wildlife posture estimation model based on heat map generation is constructed to adaptively learn feature weights; the model follows a single architecture-encoder-decoder design. For animal input images containing key points, feature extraction is performed through the backbone network to obtain feature maps. The deconvolution module in the decoder obtains the predicted heat map by upsampling the feature map. Finally, the maximum value of the heat map is obtained as the final key point coordinates through the Argmax function, and the formula is used to calculate the maximum value of the heat map. A target heat map is generated as a supervision object, and training is performed by applying a mean square error loss between the target heat map and the predicted heat map, so that the predicted heat map output by the model is close to the target heat map.

Preferably, the non-batch normalization and group whitening method specifically includes the following contents:

To address the problem of estimation offset accumulation within the constructed model network, a whitening operation is introduced into the group normalization method, and a group whitening method (i.e., non-batch dimension normalization method—group normalization method) is proposed. Neurons are divided into groups, and normalization operations are applied independently to each group of neurons for each sample to avoid estimation offset and alleviate the defect that the error of the batch normalization layer increases rapidly under small batches.

Preferably, the group normalization method specifically includes the following contents:

The whitening operation is introduced into the group normalization method. The neurons of the sample are divided into groups by group whitening so as to standardize the neurons in each group. Then the obtained groups are decorrelated to improve the free simple baseline pose estimation model.

Preferably, in the process of feature extraction, different channels contribute differently to the output results. In order to obtain better feature representation, a squeeze and extract module is introduced. The feature weights are adaptively learned according to the loss through a fully connected network. The squeeze and extract module compresses the feature value of the feature map channel into a real number through a global average pooling operation. The compressed feature map generates different weights for each channel according to the formula e=F _ex (f _c )=σ·G. The channel dimension of the feature map is scaled by multiplying the weight coefficient channel by channel, so as to achieve adaptive adjustment of different feature channels.

Preferably, S3 includes the following contents:

The Gaussian kernel of the Gaussian heat map is coordinate-encoded with the sub-pixel position as the center. Following Taylor's theorem, the activation is approximated by the Taylor second-order expansion formula evaluated at the maximum activation of the predicted heat map to achieve joint prediction with sub-pixel accuracy. A convolution operation is performed using a Gaussian kernel with a similar structure to the training data to smooth multiple peaks in the predicted heat map to obtain a single peak. The smoothed heat map is adjusted by taking the maximum and minimum values of the input feature matrix to ensure the size of the original heat map after processing.

More specifically, it includes the following:

S3.1. The heat map-based posture estimation method uses Gaussian heat map as supervision information. The coordinates of the manually annotated key points need to be generated by the network to achieve this goal. According to the formula The final Gaussian heat map is generated with the quantized coordinates as the center. The quantized coordinates have deviations, which leads to a decrease in training performance. This design allows the Gaussian kernel to be centered on the sub-pixel position, put it on the non-quantized position, and select the sub-pixel as the center for coordinate encoding to avoid quantization errors;

S3.2, the key point coordinates can only correspond to rough position coordinates due to quantization errors in the acquisition process. Now we introduce a principled distribution-aware coordinate decoding method to achieve more accurate joint positioning with sub-pixel accuracy;

To obtain accurate sub-pixel keypoint coordinates, use the formula Represents the predicted heat map of the Gaussian center corresponding to the coordinates of the key points to be predicted; the covariance is a diagonal matrix; with the help of the log-likelihood optimization principle, using the formula Logarithmic transformation of prediction heatmaps is performed to facilitate inference, while keeping the original location of maximum activation unchanged; It shows that the maximum response value point μ of the Gaussian heat map is the extreme value of the function, and the first-order derivative there is 0; the formula follows Taylor's theorem and approximates the activation by evaluating Taylor's second-order expansion formula at the maximum activation of the predicted heat map: Finally, we select m to approximate μ using the formula μ = m-(S″(m)) ^-1 S′(m), which represents a good rough joint prediction close to the maximum activation value;

S3.3. In the animal posture estimation model, the coordinate decoding method based on Taylor expansion does not have a good Gaussian structure, and the predicted heat map usually presents multiple peaks. Now we use a Gaussian kernel K with a similar structure to the training data according to the formula A convolution operation is performed to smooth out multiple peaks in the prediction heatmap.

Preferably, the S4 specifically includes the following contents:

S4.1. A lightweight strategy is adopted in the coding structure. The lightweight architecture splits the deconvolution into point-by-point convolution and depth-wise deconvolution. The deconvolution module is used to reconstruct the high-resolution feature map, and the deconvolution layer of the deconvolution module is split into depth-wise deconvolution and point-by-point convolution.

S4.2, the lightweight model first uses deep convolution to extract features from the input tensor, and then uses point-by-point convolution to compress the feature map to the required number of channels;

S4.3. In order to compensate for the performance degradation that may be caused by lightweight deconvolution operations, a three-branch fusion attention mechanism is introduced to build a structure with three branches: channel attention, spatial attention, and identity mapping.

Compared with the prior art, the present invention provides a method for estimating the posture of wild animals in real scenes, which has the following beneficial effects:

The present invention is based on a wild animal posture estimation method, which can realize the estimation of wild animal posture in real scenarios, and can be deployed on edge devices to conduct real-time monitoring and analysis of wild animals, further strengthening the intelligent development of wildlife protection, and is of great significance to the construction of ecological civilization and species diversity protection.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of a method for estimating the posture of wild animals in a real scene proposed by the present invention;

FIG2 is a comparison diagram of effects under independent and identically distributed conditions mentioned in Example 3 of the present invention;

FIG3 is a schematic diagram of the performance (ResNet101) on the Animal-Pose dataset mentioned in Example 3 of the present invention;

FIG4 is a schematic diagram of the performance (ResNet50) on the Animal-Pose dataset mentioned in Example 3 of the present invention;

FIG5 is a schematic diagram showing the performance of two data sets under small batches mentioned in Example 3 of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

The present invention provides a method for estimating the posture of wild animals in real scenes, and FIG1 is a flow chart of the method for estimating the posture of wild animals in the present invention. As shown in FIG1, it discloses the following steps: constructing a data set for estimating the posture of wild animals in real scenes, and processing the data set based on style migration; using a group whitening operation to construct a free simple baseline method; obtaining a distribution-aware coordinate representation method through unbiased sub-pixel center encoding, Taylor expansion coordinate decoding, and heat map distribution modulation, and constructing a method for estimating the posture of wild animals based on correcting quantization errors; and constructing a lightweight model for the method for estimating the posture of wild animals.

The present invention is based on a model of a method for estimating the posture of wild animals. It uses a free simple baseline method to improve the generalization ability of model recognition. It also takes into account the quantization error problem introduced in the model processing flow of this method, and designs a sub-pixel encoding method and a decoding method based on Taylor's second-order expansion to reduce the error. Finally, by constructing a lightweight model structure, the edge deployment capability is enhanced. Ultimately, accurate estimation of the posture of wild animals in complex environments can be achieved to strengthen the effective implementation of wildlife protection work.

In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

Embodiment 1:

Referring to FIG. 1 , the present invention proposes a method for estimating the posture of wild animals in a real scene, comprising:

step 1:

Construct an animal pose estimation dataset, process the public dataset, add common wild environment factors and style transfer factors in real scenes, and introduce the test part of the dataset.

Step 2:

A free simple baseline pose estimation model is proposed. This model follows a simple architecture - encoder-decoder design. For animal input images containing key points, feature extraction is performed through the backbone network to obtain feature maps. The deconvolution module in the decoder obtains the predicted heat map by upsampling the feature map. Finally, the maximum value of the heat map is obtained as the final key point coordinates through the Argmax function.

The free simple baseline pose estimation model is centered on the coordinates of the key points. According to the formula The generated target heatmap is used as the supervision object, and then the mean square error loss is applied between the predicted heatmap and the target heatmap for training, so that the predicted heatmap output by the model training is close to the target heatmap.

To solve the problem of estimation offset accumulation within the constructed model network, a whitening operation is introduced into the group normalization method. A group whitening method is proposed to divide the neurons of the sample into groups so as to standardize the neurons in each group, and then decorrelate these groups to improve the free simple baseline pose estimation model.

In the process of feature extraction, different channels contribute differently to the output results. In order to obtain better feature representation, a squeezing and extraction module is introduced, and the feature weights are adaptively learned according to the loss through a fully connected network.

The global average pooling operation compresses the eigenvalues of the feature map channels into a real number. The compressed feature map generates different weights for each channel according to the formula e = F _ex (f _c ) = σ · G. The channel dimension of the feature map is scaled by multiplying the weight coefficient channel by channel, so as to achieve adaptive adjustment of different feature channels.

Step 3:

The heat map-based pose estimation method uses Gaussian heat maps as supervision information. The manually annotated key point coordinates need to be generated by the network to achieve this goal. According to the formula The final Gaussian heat map is generated with the quantized coordinates as the center. The quantized coordinates are biased, resulting in a decrease in training performance. This design allows the Gaussian kernel to be centered on the sub-pixel position, placing it on the non-quantized position, and selecting the sub-pixel as the center for coordinate encoding to avoid quantization errors.

Since the key point coordinates can only correspond to rough position coordinates due to quantization errors in the acquisition process, a principled distribution-aware coordinate decoding method is introduced to achieve more accurate joint positioning with sub-pixel accuracy.

To obtain accurate sub-pixel keypoint coordinates, use the formula Represents the predicted heat map of the Gaussian center corresponding to the coordinates of the key points to be predicted. is a diagonal matrix. With the help of log-likelihood optimization principle, using the formula The prediction heatmaps are log-transformed to facilitate inference and keep the original location of the maximum activation unchanged. It shows that the point μ of the maximum response value of the Gaussian heat map is the extreme value of the function, and the first-order derivative there is 0. The formula follows Taylor's theorem and approximates the activation by the Taylor second-order expansion formula evaluated at the maximum activation of the predicted heat map: Finally, we select m to approximate μ using the formula μ=m-(S″(m)) ^- 1S′(m), which represents a good rough joint prediction close to the maximum activation value.

In the animal posture estimation model, the coordinate decoding method based on Taylor expansion does not have a good Gaussian structure, and the predicted heat map usually shows multiple peaks. Now we use a Gaussian kernel K with a similar structure to the training data according to the formula A convolution operation is performed to smooth out multiple peaks in the prediction heatmap.

Step 4:

A lightweight strategy is adopted in the encoding structure. The lightweight architecture splits the deconvolution into point-by-point convolution and depth-wise deconvolution. The deconvolution module is used to reconstruct the high-resolution feature map, and the deconvolution layer of the deconvolution module is split into depth-wise deconvolution and point-by-point convolution.

The lightweight model first uses deep convolution to extract features from the input tensor, and then uses point-by-point convolution to compress the feature map to the required number of channels.

In order to compensate for the performance degradation that may be caused by lightweight deconvolution operations, a three-branch fusion attention mechanism is introduced to build a structure with three branches: channel attention, spatial attention and identity mapping.

Embodiment 2:

Based on Example 1, but different in that,

This embodiment proposes a method for estimating the posture of wild animals in a real scene, which mainly includes:

A wildlife posture estimation training set that meets the design is constructed based on a public dataset and then processed.

This training set is based on the AP-10K dataset and the Animal-Pose dataset. On the basis of the public dataset, common field environment factors and style transfer factors in the real world are introduced into the test part of the dataset to construct a real scene, so that the training set and test set faced by the model are in a non-independent and identically distributed state.

Among them, a real-time style transfer method based on adaptive instance normalization is adopted. A classic wild image is selected as the style image, and the image in the dataset is used as the content image. The content image and style image are encoded using the VGG network. The style transfer operation is performed on the images in the dataset through the adaptive instance normalization layer, and finally the output of the feature space is converted to the image space through the decoder.

Non-batch normalization, whitening operation, squeezing and extraction modules are integrated into the free simple baseline pose estimation model to obtain a pose estimation method based on a heat map generation structure based on the encoding structure. The model is trained using the above training set to verify the performance of the model in real scenarios.

Among them, non-batch normalization divides neurons into groups, and the normalization operation is applied independently to each group of neurons for each sample. The group whitening operation divides the neurons of the sample into groups so that the neurons in each group are normalized and then the groups are decorrelated. The squeezing and extraction module adaptively learns feature weights according to the loss through a fully connected network.

Unbiased sub-pixel center coding, Taylor expansion coordinate coding and heat map distribution modulation method are used to obtain a distribution-aware coordinate representation method.

Unbiased sub-pixel center coding allows the Gaussian kernel to be centered at the sub-pixel position and placed at an unquantized position. The sub-pixel is selected as the center for coordinate coding to avoid the quantization error problem in the heat map coding process. Coordinate decoding based on Taylor expansion achieves more accurate joint positioning with sub-pixel accuracy through a principled distribution-aware coordinate decoding method. Heat map distribution modulation uses a Gaussian kernel K with a similar structure to the training data according to the formula A convolution operation is performed to smooth the multiple peaks in the predicted heat map. The smoothed heat map is adjusted by taking the maximum and minimum values of the input feature matrix so that the maximum activation value of the smoothed Gaussian heat map corresponds to the maximum activation value of the predicted heat map.

By adopting a lightweight deconvolution head module and a three-branch fusion attention mechanism, a wildlife posture estimation method model based on lightweight deconvolution design was obtained.

The deconvolution head module is used to reconstruct high-resolution feature maps. In the deconvolution network, it can learn upsampling parameters from the input data during training and combine upsampling and convolution parameters into one step. Deep deconvolution applies a single deconvolution kernel on each channel to upsample and obtain high-resolution feature representation. The point-by-point convolution layer convolves on all channels of the high-resolution feature representation through a 1×1 kernel, providing a bridge for cross-channel information fusion. The three-branch fusion attention mechanism consists of channel attention, spatial attention, and identity mapping.

Embodiment 3:

Based on Example 1-2, but with the difference that, the method for estimating the posture of wild animals in real scenarios proposed by the present invention is described below in combination with specific examples and designed comparative experiments, which specifically include the following contents.

1. Free Simple Baseline Pose Estimation Method Model

The present invention constructs a wild animal posture estimation model that takes into account the real-world scene offset. The model introduces a non-batch dimensional normalization strategy, which blocks the estimation offset accumulation phenomenon in the model itself and solves the problem of model performance degradation caused by continuous batch normalization layer stacking for complex field scenes. At the same time, the addition of whitening operation further enhances the generalization ability of the model in real scenes. Verified on the AP-10K and Animal-Pose datasets, compared with the SimpleBaseline method, the model has achieved a 3.82% to 8.18% performance improvement in real field scenes such as weather effects, motion blur, and equipment noise, and the accuracy rate has increased by 2.07% to 7.39% when facing different migration scenarios. The specific contents are shown in the following Tables 1-4 and Figures 2-5.

(1) Improvement effect of non-batch dimension normalization

1. The location of non-batch normalization methods in the network (as shown in Table 1)

Table 1 Experimental results of different replacement positions

2. Performance of different non-batch normalization methods (as shown in Table 2)

Table 2 Experimental results of different normalization methods

(2) Model experimental results in real scenarios

1. Experimental results of the model under independent and identical distribution (as shown in Figure 2)

2. Model experimental results in real scenarios (as shown in Table 3-4 and Figure 3-4)

Table 3 Comparison of model effects in real scenarios

Table 4 Comparison of model effects in style transfer scenarios

(3) Experimental results of model performance under small batches (as shown in Figure 5).

2. Posture Estimation Method Based on Distribution-aware Coordinate Representation

The present invention proposes a method for correcting the inherent quantization error of the posture estimation model. The present invention generates an unbiased heat map through sub-pixel coding for training supervision, and corrects the key point coordinate error problem caused by the reduction of image resolution by coordinate decoding based on the Taylor expansion formula and the heat map distribution modulation method. In addition, an unbiased processing method for data enhancement is proposed to redefine the coordinate axis, and the problem of inconsistent results before and after coordinate transformation is eliminated by using the sampling unit distance as the coordinate axis scale. The improvement schemes proposed for two different quantization errors have improved the accuracy indicators by 1.31% and 2.34% on the AP-10K dataset respectively without increasing the complexity of the network. The specific contents are shown in Tables 5-6 below.

(1) Experimental comparison of key point coordinate representation algorithms (as shown in Table 5-6)

Table 5 Comparison of experimental results of key point coordinate representation algorithm

method AP AP ^M AP ^L AR AR ^{M AR} FSB 66.15 50.58 66.47 70.27 51.77 70.46 FSB+ 67.53 51.48 67.88 71.25 53.88 71.45 Simplebaseline 65.04 54.12 65.67 69.53 56.84 69.99 Simplebaseline+ 66.54 50.77 66.85 70.91 52.23 71.11

Table 6 Experimental comparison of different input image resolutions

method Image size AP AR FSB 256×256 66.15 70.27 FSB+ 256×256 67.53 71.25 FSB 192×192 61.37 65.44 FSB+ 192×192 62.8 66.84 FSB 128×128 47.38 51.79 FSB+ 128×128 49.7 53.7

3. Lightweight posture estimation model

The present invention designs a lightweight improvement strategy for the animal posture estimation model. In order to facilitate the deployment and promotion of the model, the model is lightweighted. Drawing on the concept of deep separable convolution, the present invention designs a lightweight deep separable deconvolution module. At the same time, in order to compensate for the performance degradation caused by lightweighting, a three-branch fusion attention mechanism of channel, space and feature identity mapping is proposed. In the end, the performance of the model only decreased by 0.1% on the basis of reducing the number of parameters and the amount of calculation by 18% and 12% respectively. The specific contents are shown in the following tables 7-8.

(1) Performance analysis of lightweight posture estimation model (as shown in Table 7)

Table 7 Experimental comparison of lightweight codec structure

Backbone network Deconvolution Head Parameter quantity Calculation Amount AP AR ResNet101 ordinary 52.99 12.14 66.7 70.67 ResNet101 Lightweight 43.24 10.75 66.6 70.45 ResNet50 ordinary 34 7.28 65.2 69.42 ResNet50 Lightweight 24.25 5.89 63.49 68.51 MobileNetV3 ordinary 8.67 1.96 59.14 63.51 MobileNetV3 Lightweight 3.08 0.78 57.5 62.46 ShuffleNetV2 ordinary 7.8 1.69 54.37 58.21 ShuffleNetV2 Lightweight 2.37 0.59 51.49 56.43

(2) Comparative experimental results of different attention mechanisms (as shown in Table 8)

Table 8 Experimental comparison of different attention mechanisms

Attention Mechanism AP AP ⁵⁰ AP ⁷⁵ AR Parameter quantity None 65.28 91.08 71.58 70.01 43.24 SE 66.15 91.51 72.82 70.67 43.79 CBAM 66.29 92.21 71.52 70.37 43.79 Ours 66.6 91.96 71.66 70.45 44.12

It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims (3)

1. The method for estimating the pose of the wild animal in the real scene is characterized by comprising the following steps of:

S1, constructing a data set of an animal posture estimation image, and processing the constructed data set based on style migration;

S2, based on the group whitening operation, constructing a free simple baseline attitude estimation model based on heat map generation, generating a heat map by using the model, and completing model training by using the heat map; the method specifically comprises the following steps:

utilizing a non-batch normalization and group whitening method, adding an extrusion and extraction module, and constructing a free simple baseline wild animal attitude estimation model based on heat map generation so as to self-adaptively learn characteristic weights; the model follows a single architecture-encoder decoder design, feature extraction is carried out on an animal input image containing key points through a backbone network to obtain a feature image, a deconvolution module in the decoder carries out up-sampling on the feature image extracted by the backbone network feature to obtain a predicted heat image, and finally a heat image maximum value is obtained through Argmax functions to serve as final key point coordinates; marking the coordinates of the key points by using the model, and passing through a formula Generating a target heat map, and training by using a mean square error loss between the target heat map and the predicted heat map;

the non-batch normalization and group whitening method specifically comprises the following steps:

introducing a non-batch dimension normalization method, namely a group normalization method, dividing neurons into groups, and independently applying standardized operation to each group of neurons of each sample to avoid estimating offset and alleviate the defect that errors of a batch normalization layer become large rapidly under small batches;

the group normalization method specifically comprises the following steps:

Introducing a whitening operation into a group normalization method, dividing neurons of a sample by using group whitening so as to normalize the neurons in each group, and then correlating the obtained groups;

The squeezing and extracting module compresses the eigenvalue of the eigenvector channel into a real number through global average pooling operation, and the compressed eigenvector generates different weights for each channel according to the formula e=f _ex(f_c) =σ·g, and uses Multiplying the weight coefficients channel by channel, and scaling the channel dimension of the feature map so as to realize the self-adaptive adjustment of different feature channels;

S3, correcting the model constructed in the S2, and designing a coordinate representation method and a heat diagram code method; the method specifically comprises the following steps:

S3.1, according to the formula Generating a final Gaussian heat map by taking the quantized coordinates as a center; the Gaussian kernel of the Gaussian heat map is subjected to coordinate coding by taking the position of a subpixel as the center, the Taylor theorem is followed, the Taylor second-order expansion formula estimated at the maximum activation of the prediction heat map is used for approximate activation, and the joint prediction is realized with subpixel precision, specifically:

using the formula A prediction heat map of a Gaussian center corresponding to coordinates of a key point to be predicted is represented; wherein covariance isIs a diagonal matrix;

By means of the log-likelihood optimization principle, the formula is utilized Performing logarithmic transformation on the predicted heat map to promote reasoning and keeping the original position of maximum activation unchanged;

The maximum response value point mu of the Gaussian heat map is the extremum of the function, and the first derivative is 0; the formula follows the taylor theorem, approximating activation by a taylor second order expansion formula that evaluates at the maximum activation of the predictive heat map:

Using the formula μ=m- (S "(m)) ^-1 S' (m) to select m to approximate μ, which represents a good rough joint prediction near the maximum activation value;

s3.2, adopting a Gaussian kernel K with a similar structure with training data to calculate the formula Performing convolution operation to smoothly predict a plurality of peaks in the heat map so as to obtain a peak; the smoothed heat map is adjusted by taking the maximum value and the minimum value of the input feature matrix so as to ensure the size of the processed original heat map;

s4, integrating the operations of S1-S3, and adopting a light pose estimation network decoding architecture to finish estimation of the pose of the wild animal in the real scene.

2. The method for estimating the pose of a wild animal in a real scene according to claim 1, wherein said processing the constructed dataset in S1 specifically comprises the following steps:

The data set is divided into a test set and a training set, interference factors of field scenes are introduced to the test set, and complex environment scenes are simulated so as to observe the distribution difference of the test set and the training set under different scenes.

3. The method for estimating the pose of a wild animal in a real scene according to claim 1, wherein S4 specifically comprises the following contents:

deconvolution is split into point-by-point convolution and depth deconvolution to reduce parameters and memory requirements of a model, and a three-branch fusion attention module is added to compensate for performance reduction caused by light deconvolution operation.