A pose estimation method, apparatus and device

Abstract

The application discloses a pose estimation method, device and equipment, and the method comprises the following steps: acquiring a to-be-recognized pose feature map; performing grouped pose recognition estimation on the to-be-recognized pose feature map by using a pre-trained human body pose estimation network model to obtain a conventional feature map recognition result set and a redundant feature map recognition result set; performing fusion processing based on the conventional feature map recognition result set and the redundant feature map recognition result set to obtain a pose estimation result; and the above design solves the problem that the existing human body pose estimation method neglects the effective utilization of the redundant features of the input image.

Description

Technical Field

[0001] This invention relates to the field of human posture estimation, specifically to a posture estimation method, apparatus, and device. Background Technology

[0002] The goal of human pose estimation is to correctly detect and locate key points of the human body in an input image, which is fundamental to other downstream action tasks. With the development of deep neural networks, human pose estimation has become increasingly important in computer vision. High-resolution feature representations can provide higher-performance human pose estimation models, but this typically requires more computational resources, making it difficult to apply these models on mobile devices.

[0003] Current state-of-the-art pose estimation models based on HRNet suffer from large parameter counts and high computational costs. Many attempts have been made to make these models lightweight. However, common methods mainly focus on replacing the network with a lightweight one or eliminating less important filters, neglecting research on the neural architecture design of redundant feature maps in convolutional layers. These hidden details are particularly important for pose estimation tasks.

[0004] To address the aforementioned problems in existing technologies, we propose a pose estimation method. Summary of the Invention

[0005] This invention provides a pose estimation method, apparatus, and device, aiming to solve the problem that existing human pose estimation methods neglect the effective utilization of redundant features of input images.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An attitude estimation method, comprising:

[0008] Obtain the pose feature map to be identified;

[0009] A pre-trained human pose estimation network model is used to perform grouped pose recognition estimation on the pose feature maps to be identified, resulting in a set of regular feature map recognition results and a set of redundant feature map recognition results.

[0010] The pose estimation result is obtained by fusing the results of conventional feature map recognition and redundant feature map recognition.

[0011] Preferably, the human pose estimation network includes:

[0012] Two 3×3 convolutional groups are used to adjust the size of the pose feature map to be identified and extract shallow features;

[0013] The first stage comprises four pyramidal dynamic convolutional units for identifying the shallow features and outputting a feature map with the shallow features; wherein each pyramidal dynamic convolutional unit consists of a PODConv module;

[0014] The second stage consists of a high-resolution branch and a low-resolution branch. The low-resolution branch is used to extract deep features based on the shallow features, and the high-resolution branch is used to output a feature map with the deep features of the second stage. The high-resolution branch and the low-resolution branch each contain 4 GPBlock modules.

[0015] The third stage is constructed by one high-resolution branch and two low-resolution branches. The two low-resolution branches are used to extract the deep features of the object in the third stage based on the feature map with the deep features of the second stage. The high-resolution branch outputs the feature map with the deep features of the third stage. Each of the high-resolution branch and the low-resolution branch contains 4 GPBlock modules.

[0016] The fourth stage is constructed by one high-resolution branch and three low-resolution branches. The three low-resolution branches are used to extract the deep features of the fourth stage based on the feature map with the deep features of the third stage. The high-resolution branch outputs the pose estimation result after the three low-resolution branches are fused. Each of the high-resolution branch and the low-resolution branch contains 4 GPBlock modules.

[0017] A 1×1 convolutional layer is used to perform heatmap operations on the pose estimation results.

[0018] Preferably, the feature maps of the pose to be identified are grouped to obtain regular feature maps and redundant feature maps;

[0019] A pre-trained human pose estimation network model is used to perform grouped pose recognition estimation on the feature maps to be identified, resulting in a set of regular feature map recognition results and a set of redundant feature map recognition results, including:

[0020] The GPConv layer in the GPblock module of the human pose estimation network model is used to perform a conventional convolution operation on the pose feature map to be identified in the 1 / α input channel to obtain a set of conventional feature map recognition results.

[0021] The GPConv layer in the GPblock module of the human pose estimation network model is used to perform a depthwise convolution operation on the pose feature map to be identified in the (α-1) / α input channel to obtain a set of redundant feature map recognition results.

[0022] Preferably, the GPConv layer in the GPblock module is used to perform a depthwise convolution operation on the pose feature map to be recognized in the (α-1) / α input channel to obtain a redundant feature map recognition result set, including:

[0023] The pose feature map to be identified in the (α-1) / α input channel is divided into three groups using the GPConv layer;

[0024] The pose feature maps to be identified in each group are filtered by a filter with progressively increasing kernel size to obtain a set of redundant feature map recognition results.

[0025] Preferably, the pose estimation result is obtained by fusing the recognition results of the conventional feature map set and the recognition results of the redundant feature map set, including:

[0026] The GPConv layer in the GPblock module is used to stitch together the regular feature map recognition result set and the redundant feature map set along the channel dimension.

[0027] The pose estimation result is obtained by fusing the result set of regular feature maps and the result set of redundant feature maps using the first pointwise convolutional layer in the GPblock module.

[0028] Preferably, the training process of the human pose estimation network model includes:

[0029] Step 1: Obtain the human pose estimation dataset, including training set data and test set data, and perform data preprocessing;

[0030] Step 2: Construct the initial human pose estimation network;

[0031] Step 3: Use the training set data obtained in Step 1 to train the initial human pose estimation network;

[0032] Step 4: Test the human pose estimation network model trained in Step 3 on the test set data obtained in Step 1.

[0033] Preferably, the three branches in the third stage are fused four times, and the four branches in the fourth stage are fused three times.

[0034] Preferably, the GPblock module also includes a second pointwise convolutional layer and an SE attention layer.

[0035] An attitude estimation device, comprising:

[0036] The feature map acquisition module is used to acquire the pose feature map to be identified;

[0037] The grouped pose estimation module is used to perform grouped pose recognition estimation on the feature maps to be identified using a pre-trained human pose estimation network model, and obtain a set of regular feature map recognition results and a set of redundant feature map recognition results.

[0038] The pose estimation result generation module is used to fuse the conventional feature map recognition result set and the redundant feature map recognition result set to obtain the pose estimation result.

[0039] An attitude recognition device includes: a memory and a processor, wherein the memory stores instructions and the memory and the processor are interconnected via a circuit; the processor invokes the instructions in the memory.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This invention proposes a pose estimation method, comprising: acquiring a pose feature map to be identified; performing grouped pose recognition estimation on the pose feature map using a pre-trained human pose estimation network model to obtain a set of regular feature map recognition results and a set of redundant feature map recognition results; and fusing the set of regular feature map recognition results and the set of redundant feature map recognition results to obtain a pose estimation result. The method utilizes a lightweight GPblock module to improve the effective utilization of redundant features, thereby assisting the network in learning more subtle and elusive information. This invention applies this module to a baseline model to construct a human pose estimation network. The GPConv layer in the GPblock module first selects a portion of the feature maps for standard convolution operations, while treating the remaining portion as redundant feature maps and applying convolution operations with different kernel sizes to them. This processing method not only helps to obtain implicit detailed information but also obtains multi-scale feature information, thereby helping the entire network to better predict the fine-grained feature information needed for occluded keypoints. Attached Figure Description

[0042] Figure 1 This is a flowchart of the human pose estimation network model of the present invention;

[0043] Figure 2 This is a diagram of the human pose estimation network model of the present invention;

[0044] Figure 3 This is a schematic diagram of the GPblock module of the present invention;

[0045] Figure 4 This is a diagram of the attitude estimation device of the present invention;

[0046] Figure 5 This is a structural diagram of the attitude estimation device of the present invention. Detailed Implementation

[0047] A preferred embodiment of the present invention will now be described in conjunction with the accompanying drawings, providing a clear and complete description of the technical solution in this preferred embodiment.

[0048] Example 1:

[0049] like Figure 1 As shown, an attitude estimation method includes...

[0050] Step 101: Obtain the pose feature map to be identified;

[0051] Step 102: Use a pre-trained human pose estimation network model to perform grouped pose recognition estimation on the pose feature maps to be identified, and obtain a set of regular feature map recognition results and a set of redundant feature map recognition results;

[0052] Step 103: Perform fusion processing on the conventional feature map recognition result set and the redundant feature map recognition result set to obtain the pose estimation result.

[0053] This invention utilizes a human pose estimation network model to first perform standard convolution operations on a portion of the feature maps to obtain a set of conventional feature map recognition results. The remaining portion is treated as redundant feature maps, and convolution operations with different kernel sizes are applied to these redundant feature maps to obtain a set of redundant feature map recognition results. Finally, the set of conventional feature map recognition results and the set of redundant feature map recognition results are fused to obtain a set of pose estimation results. This processing method not only helps to obtain implicit detailed information but also obtains multi-scale feature information, thereby helping the entire network to better predict the fine-grained feature information required for occluded key points. This solves the problem that existing human pose estimation methods ignore the effective utilization of redundant features of the input image.

[0054] like Figure 2 As shown, the human pose estimation network used in the above embodiments includes:

[0055] Two 3×3 convolutional groups are used to adjust the size of the pose feature map to be identified and extract shallow features;

[0056] The first stage, consisting of four pyramidal dynamic convolutional units, is used to identify feature maps that combine dynamic convolution with multi-scale feature map structures; each pyramidal dynamic convolutional unit is composed of a PODConv module.

[0057] The second stage consists of a high-resolution branch and a low-resolution branch. The high-resolution branch maintains the size of the feature map output from the first stage, while the low-resolution branch is used to reduce the size of the pose feature map to be identified by half and increase the width to 64. The high-resolution branch and the low-resolution branch each contain 4 GPBlock modules.

[0058] The third stage consists of one high-resolution branch and two low-resolution branches. The high-resolution branch maintains the size of the feature map output from the first stage, but adjusts the width to 32. The first low-resolution branch is used to reduce the length and width of the feature map to be identified by half and increase the width to 64. The second low-resolution branch reduces the length and width of the feature map by half based on the first low-resolution branch, and increases the width to 128. Each of the high-resolution branch and the low-resolution branch contains 4 GPBlock modules.

[0059] The fourth stage consists of one high-resolution branch and three low-resolution branches. The high-resolution branch maintains the size of the feature map output from the first stage, but adjusts the width to 32. The first low-resolution branch reduces the length and width of the feature map to be identified by half, and increases the width to 64. The second low-resolution branch reduces the length and width of the feature map by half based on the first low-resolution branch, and increases the width to 128. The third low-resolution branch reduces the length and width of the feature map by half based on the second low-resolution branch, and increases the width to 256. The high-resolution branch outputs the pose estimation result after fusing the three low-resolution branches. Both the high-resolution branch and the low-resolution branches contain 4 GPBlock modules.

[0060] A 1×1 convolutional layer is used to perform heatmap operations on the pose estimation results.

[0061] The two 3*3 convolutional groups in the backbone of the human pose estimation network perform downsampling operations on the input pose feature map to be identified, reducing the size of the pose feature map to be identified, and initially extracting shallow features.

[0062] The first stage of a human pose estimation network is the feature map recognition process, which combines dynamic convolution with a multi-scale feature map structure. Dynamic convolution dynamically aggregates multiple parallel convolutional kernels based on attention. The attention dynamically adjusts the weights of each kernel according to the input, thus generating adaptive dynamic convolutions. The multi-scale feature map structure applies dynamic convolution operations with different kernel sizes to different parts of the feature map. This ensures that different parts of the feature map are influenced by filters of different scales, resulting in feature information at different scales.

[0063] The second, third, and fourth stages of the human pose estimation network involve both the identification of conventional feature maps and the identification of redundant feature maps. Each stage consists of an increasing number of GPblock modules. Within each GPblock module, conventional and redundant feature maps are identified, and then fused together.

[0064] The main branch is finally processed by a 1×1 convolutional layer to generate a heatmap of the fused result.

[0065] Specifically, it also includes: grouping the feature maps of the pose to be identified to obtain regular feature maps and redundant feature maps;

[0066] A pre-trained human pose estimation network model is used to perform grouped pose recognition estimation on the feature maps to be identified, resulting in a set of regular feature map recognition results and a set of redundant feature map recognition results, including:

[0067] The GPConv layer in the GPblock module of the human pose estimation network model is used to perform a conventional convolution operation on the pose feature map to be identified in the 1 / α input channel to obtain a set of conventional feature map recognition results.

[0068] The GPConv layer in the GPblock module of the human pose estimation network model is used to perform a depthwise convolution operation on the pose feature map to be identified in the (α-1) / α input channel to obtain a set of redundant feature map recognition results.

[0069] The GPConv layer within the GPblock module of the human pose estimation network model divides all input channels into two paths proportionally. One path uses regular convolution on the 1 / α channel feature map (in this invention, α=4 better matches the model's operations in the channel direction) to obtain clear and representative important information. The other path first performs depthwise convolution on the remaining (α-1) / α channel feature map, thereby improving the overall module's parameter and computational cost. Finally, the two paths yield two types of result feature maps. Since the result feature maps of the two channels have the same size, they can be fused together to obtain a result feature map with the same shape as the input feature map. In this way, after the input feature map passes through the GPConv layer in the GPblock module, it can obtain both hidden detail information and multi-scale feature information, thereby helping the entire network to improve the fine-grained feature information required for predicting occluded keypoints.

[0070] Specifically, the GPConv layer in the GPblock module is used to perform a depthwise convolution operation on the pose feature map of the (α-1) / α input channel to obtain a redundant feature map recognition result set, including:

[0071] The pose feature map to be identified in the (α-1) / α input channel is divided into three groups using the GPConv layer;

[0072] The pose feature maps to be identified in each group are filtered by a filter with progressively increasing kernel size to obtain a set of redundant feature map recognition results.

[0073] The GPConv layer in the GPblock module subdivides the pose feature map to be recognized within the (α-1) / α input channels, dividing it into three groups along the channel dimension. Each group is then filtered using filters with progressively larger kernel sizes (5×5, 7×7, and 9×9), resulting in a redundant feature map recognition set. This redundant feature map utilizes multi-scale convolutional kernels to generate different spatial resolutions, significantly expanding the receptive field of the convolutional layer throughout the convolutional operation and enriching the output functionality.

[0074] Specifically, the pose estimation results are obtained by fusing the results of conventional feature map recognition and redundant feature map recognition, including:

[0075] The GPConv layer in the GPblock module is used to stitch together the regular feature map recognition result set and the redundant feature map set along the channel dimension.

[0076] The pose estimation result is obtained by fusing the result set of regular feature maps and the result set of redundant feature maps using the first pointwise convolutional layer in the GPblock module.

[0077] Since depthwise convolution applies a convolution to each channel separately, there are no connections between the input channels. This means that redundancy is further reduced while also severing connections between input channels. Therefore, the two types of feature maps obtained from the two paths need to be fused. Because the feature maps of the two channels are of the same size, all channel information can be fused by connecting a first pointwise convolutional layer after the GPConv layer. At the same time, the first pointwise convolutional layer can also expand the number of input channels, thereby generating richer intermediate feature maps to extract more information.

[0078] Specifically, this application discloses the training process of a human pose estimation network model, including:

[0079] Step 1: Obtain the human pose estimation dataset, including training set data and test set data, and perform data preprocessing;

[0080] Step 2: Construct the initial human pose estimation network;

[0081] Step 3: Use the training set data obtained in Step 1 to train the initial human pose estimation network;

[0082] Step 4: Test the human pose estimation network model trained in Step 3 on the test set data obtained in Step 1.

[0083] Specifically, the three branches in the third stage are merged four times, and the four branches in the fourth stage are merged three times.

[0084] The third and fourth stages are repeated multiple times, exchanging information at different resolutions to capture richer feature map information.

[0085] This invention provides a complete embodiment of a human pose estimation network:

[0086] Step 1: Obtain the human pose estimation dataset, including training and testing data, and perform data preprocessing. This invention is evaluated on two human pose estimation datasets, the COCO dataset. K=17 heatmaps are used to represent the confidence of keypoint locations. The COCO dataset contains over 200K images and 250K human instances. The dataset annotations include 17 keypoints: 0 for nose, 1 for left eye, 2 for right eye, 3 for left ear, 4 for right ear, 5 for left shoulder, 6 for right shoulder, 7 for left elbow, 8 for right elbow, 9 for left wrist, 10 for right wrist, 11 for left hip, 12 for right hip, 13 for left knee, 14 for right knee, 15 for left ankle, and 16 for right ankle. The human pose estimation model of this invention is trained on the train2017 dataset (including 57K images and 150K human instances) and validated on the val2017 dataset (including 5K images).

[0087] Step 2: Construct the initial human pose estimation network:

[0088] Step 3: Using the training set data obtained in Step 1, train the initial human pose estimation network on a single GeForce RTX 3090 GPU with a minimum batch size of 32 per GPU. The height or width of the human detection bounding box is expanded to a fixed aspect ratio:height:width = 4:3. The human bounding box is then cropped from the image and resized to a fixed size of 256×192. Adam is used as the optimizer during network training, with an initial learning rate of 1E-3. The network is trained for a total of 210 epochs. From epoch 170 to epoch 200, the learning rate decays at a rate of 0.1, meaning it decays to 1E-4 at epoch 170 and to 1E-5 at epoch 210. During training, data augmentation was performed using random image rotation (between -45° and 45°), random scaling (between 0.65 and 1.35), random horizontal flipping, and half-body (with a certain probability of cropping the target to retain only half of the key points, either the upper or lower body).

[0089] Step 4: Test the human pose estimation network model trained in Step 3 on the test set data obtained in Step 1. This invention uses a two-stage top-down paradigm (detecting human instances and predicting keypoints using a human detector): detecting human instances using a human detector, and then predicting the detected keypoints. This invention uses the same human detector as the Simple Baseline for the validation set. A heatmap is calculated by averaging the predicted heatmaps of the original and flipped images. The final position of each keypoint is predicted by adjusting the location of the highest heat value, with a quarter-offset in the direction from the highest response to the second highest response.

[0090] The experimental results of this invention on the COCO validation set are shown in Table 1. The results show that the human pose estimation network model of this invention has less than half the number of parameters and computational complexity of the existing HRNet model, while achieving higher accuracy. Compared with other recent human pose estimation models such as Hourglass, CPN, CPN+OHKM, Simple Baseline, PRTR, and DistilPose-L, the human pose estimation network model achieves improvements of 8.4, 6.9, 5.9, 4.9, 2.4, and 0.9 percentage points in mAP, respectively. Compared with the HRNet model, the SESPA_HRNet model improves mAP by 0.9 percentage points and AP50 by 2.1 percentage points, while maintaining a comparable level to the HRNet model in other validation metrics. Furthermore, the number of parameters in the human pose estimation network model is reduced by 60% compared to the HRNet model, and the computational complexity is reduced by 56%.

[0091] Table 1 shows the performance comparison of various network models on the COCO validation set.

[0092]

[0093] Specifically, in the above embodiments, the GPblock module further includes a second pointwise convolutional layer and an SE attention layer.

[0094] The first pointwise convolutional layer in the GPblock module is followed by a second pointwise convolutional layer for channel dimensionality reduction and controlling the number of channels in the output feature map. To ensure that each channel's feature map receives appropriate attention, an SE channel attention layer is connected after the second pointwise convolutional layer. The SE channel attention layer assigns different weights to different locations in the image from the perspective of the channel domain, obtaining more important feature information. In addition to the above operation layers, a normalization layer and an activation function layer are also included between the first and second pointwise convolutional layers to maintain the diversity of features obtained in the feature map and reduce latency.

[0095] Specifically, in addition to describing the attitude estimation method in detail, this application also discloses, as well as... Figure 4 The aforementioned apparatus, an attitude estimation apparatus, comprises:

[0096] Feature map acquisition module 401 is used to acquire the pose feature map to be identified;

[0097] The grouped pose estimation module 402 is used to perform grouped pose recognition estimation on the pose feature map to be identified using a pre-trained human pose estimation network model, and obtain a set of regular feature map recognition results and a set of redundant feature map recognition results.

[0098] The pose estimation result generation module 403 is used to perform fusion processing based on the conventional feature map recognition result set and the redundant feature map recognition result set to obtain the pose estimation result.

[0099] Specifically, this application also discloses, as follows Figure 5 The posture recognition device includes a memory 502 and a processor 501. The memory stores instructions, and the memory and the processor are interconnected via a circuit. The processor calls the instructions in the memory.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pose estimation method, characterized in that: include Obtain the pose feature map to be identified; The pose estimation network model is used to perform grouped pose recognition estimation on the pose feature map to be identified, so as to obtain a set of regular feature map recognition results and a set of redundant feature map recognition results. The human pose estimation network model includes a GPblock module, which includes a GPConv layer; The step of using a pre-trained human pose estimation network model to perform grouped pose recognition estimation on the pose feature map to be identified includes: using the GPConv layer to divide the input channels into two groups in parallel according to the ratios of 1 / α and (α-1) / α; using the GPConv layer to perform a regular convolution operation on the pose feature map to be identified in the 1 / α input channel to obtain the regular feature map recognition result set; and using the GPConv layer to perform a depthwise convolution operation on the pose feature map to be identified in the (α-1) / α input channel to obtain the redundant feature map recognition result set, where α=4; The pose estimation result is obtained by fusing the conventional feature map recognition result set and the redundant feature map recognition result set. The process includes: concatenating the conventional feature map recognition result set and the redundant feature map recognition result set in the channel dimension; and using the first pointwise convolutional layer in the GPblock module to perform cross-channel weighted linear recombination on all channels after concatenation to obtain the fused pose estimation result. The GPblock module also includes a second pointwise convolutional layer and an SE attention layer.

2. The attitude estimation method according to claim 1, characterized in that, The human pose estimation network includes: Two 3×3 convolutional groups are used to adjust the size of the pose feature map to be identified and extract shallow features; The first stage comprises four pyramidal dynamic convolutional units for identifying the shallow features and outputting a feature map with the shallow features; wherein each pyramidal dynamic convolutional unit consists of a PODConv module; The second stage consists of a high-resolution branch and a low-resolution branch. The low-resolution branch is used to extract deep features based on the feature map with the shallow features, and the high-resolution branch is used to output a feature map with the deep features of the second stage. The high-resolution branch and the low-resolution branch each contain 4 GPBlock modules. The third stage is constructed by one high-resolution branch and two low-resolution branches. The two low-resolution branches are used to extract the deep features of the object in the third stage based on the feature map with the deep features of the second stage. The high-resolution branch outputs the feature map with the deep features of the third stage. Each of the high-resolution branch and the low-resolution branch contains 4 GPBlock modules. The fourth stage is constructed by one high-resolution branch and three low-resolution branches. The three low-resolution branches are used to extract the deep features of the fourth stage based on the feature map with the deep features of the third stage. The high-resolution branch outputs the pose estimation result after the three low-resolution branches are fused. Each of the high-resolution branch and the low-resolution branch contains 4 GPBlock modules. A 1×1 convolutional layer is used to perform heatmap operations on the pose estimation results.

3. The attitude estimation method according to claim 2, characterized in that, Also includes: Based on the pose feature map to be identified, grouping is performed to obtain regular feature maps and redundant feature maps.

4. The attitude estimation method according to claim 3, characterized in that, The step involves performing a depthwise convolution operation on the (α-1) / α input channel's pose feature map using the GPConv layer in the GPblock module to obtain the redundant feature map recognition result set, including: The GPConv layer is used to divide the pose feature map to be identified in the (α-1) / α input channel into three groups; The pose feature maps to be identified in each group are filtered by a filter with progressively increasing kernel size to obtain the set of redundant feature map recognition results.

5. The attitude estimation method according to claim 1, characterized in that, The training process of the human pose estimation network model includes: Step 1: Obtain the human pose estimation dataset, including training set data and test set data, and perform data preprocessing; Step 2: Construct the initial human pose estimation network; Step 3: Use the training set data obtained in Step 1 to train the initial human pose estimation network; Step 4: Test the human pose estimation network model trained in Step 3 on the test set data obtained in Step 1.

6. The attitude estimation method according to claim 2, characterized in that: The three branches in the third stage are fused four times, and the four branches in the fourth stage are fused three times.

7. An attitude estimation device, characterized in that, include: The feature map acquisition module is used to acquire the pose feature map to be identified; The grouped pose estimation module is used to perform grouped pose recognition estimation on the pose feature map to be identified using a pre-trained human pose estimation network model, so as to obtain a set of regular feature map recognition results and a set of redundant feature map recognition results. The pose estimation result generation module is used to perform fusion processing on the conventional feature map recognition result set and the redundant feature map recognition result set to obtain the pose estimation result; Implement the attitude estimation method as described in any one of claims 1 to 6.

8. A posture recognition device, characterized in that, include: A memory and a processor, wherein the memory stores instructions and the memory and the processor are interconnected via a circuit; The processor invokes the instructions in the memory to implement the attitude estimation method as described in any one of claims 1 to 6.

Images