Improved multi-person attitude estimation method for edge computing equipment

Abstract

The invention discloses an improved multi-person attitude estimation method for edge computing equipment. Comprises: preprocessing a video input by a video input device; a YoLoPose prediction target frame and skeleton point coordinates are improved; according to the appearance characteristics of the targets, using Deep Oc-port to calibrate tracking frame serial numbers for different targets; according to the tracking frame sequence number, inputting skeleton point data into TSSTG to realize multi-person attitude estimation; the feature extraction network in the skeleton point detection model YoLoPose and the Residual part of the last C2f module of the neck are fused into DAtention; adding an LSKA in front of a final convolution output layer of the SPPF module; the method comprises the following steps of: changing a PANet network of a neck into a PANet-P6 network, and replacing up-sampling of an original edition by using Dysample; and an Anchor-Based detection head is used, and an SEAM attention module is added before the output convolution of the detection head, so that an improved YoLoPose model is formed. According to the invention, multi-person posture estimation can be realized and the recognition precision of human body actions can be improved.

Description

Technical Field

[0001] The present invention relates to the field of vision and target detection technology, and in particular to an improved multi-person posture estimation method for edge computing devices. Background Art

[0002] Multi-person pose estimation has always been an important research direction in the development of computer vision and has been widely used in the fields of medicine, sports training, autonomous driving, and film and television animation. In the field of pose estimation, the difficulties and hotspots faced by researchers include the ability to identify occluded targets, perform pose estimation on multiple targets at the same time, reduce computing costs, and increase detection speed to meet the needs of edge computing devices. It is usually hoped that multi-person pose estimation will be deployed on edge computing devices to meet the requirements of different real-time application scenarios.

[0003] A lot of experience shows that skeleton point information, as a higher-level visual information, can reflect motion feature vectors and is suitable for posture estimation tasks, and the quality of skeleton point information directly affects the effect of subsequent posture estimation tasks. However, traditional skeleton point detection methods are difficult to apply on edge devices with limited computing power due to their high computational cost and slow detection speed. They also only consider posture estimation for a single person, which results in recognition errors when multiple people appear in the image, thereby reducing the accuracy of multi-person posture estimation.

[0004] Therefore, the prior art still needs to be improved and enhanced. Summary of the invention

[0005] In view of the above-mentioned shortcomings of the prior art, the purpose of the present invention is to provide an improved multi-person posture estimation method for edge computing devices, aiming to solve the problem that the skeleton point detection method in the prior art only considers posture estimation of a single person, resulting in recognition errors when multiple people appear in the image, thereby reducing the accuracy of multi-person posture estimation.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, an improved multi-person posture estimation method for edge computing devices is provided, the method comprising the following steps: Step 1: Based on the improved YoLoPose network, an improved multi-person pose estimation model is established; Step 2: training the improved multi-person pose estimation model, using the COCO KeyPoints dataset to train the improved YoLoPose network, and obtaining a trained improved multi-person pose estimation model; Step 3: Input the video collected by the camera in real time into the trained improved multi-person posture estimation model, perform multi-person posture estimation using the improved multi-person posture estimation model, and output the multi-person posture estimation result.

[0007] Furthermore, the establishment of an improved multi-person posture estimation model in step 1 specifically includes the following steps: Step 1.1: Based on YoLoPose, DAttention is integrated into the feature extraction network and the residual part of the last C2f module of the neck to detect tiny bone points; LSKA is integrated into the SPPF module to eliminate irrelevant background interference; the PANet network of the neck is changed to the PANet-P6 network, and Dysample is used to replace the original upsampling to ensure efficient upsampling performance while avoiding complex structural design and high computational cost; Anchor-Based detection head is used to compensate for the problem of overlapping bone points, and the SEAM attention module is added before the output convolution of the detection head to enhance the detection of occluded bone points; Step 1.2: Input the appearance features extracted by the improved YoLoPose into Deep Oc-sort to calibrate the tracking frame numbers of different targets to achieve multi-target tracking; Step 1.3: Save the target skeleton point information according to the tracking frame serial number of different targets, input the target skeleton point information into TSSTG to realize posture estimation, and complete the establishment of the improved multi-person posture estimation model.

[0008] Furthermore, the training of the improved multi-person posture estimation model in step 2 includes the following steps: Step 2.1: Use the COCO KeyPoints dataset to train the improved YoLoPose network, and use the OKS loss instead of Iou in target detection to perform regression prediction of skeleton points; calculate OKS for each individual skeleton point and accumulate it to the final OKS loss; the total skeleton point loss function formula is as follows: in, , , , is a hyperparameter to balance the loss between different sizes, They are the losses of classification, target box, skeleton point and skeleton point confidence respectively; Step 2.2: Train the Deep Oc-sort network using the Market-1501 dataset. Step 2.3: Select 30 fall sequences from the Fall Detection Dataset, and increase the number of data sets to 90 by flipping and adding noise; use the trained improved YoLoPose to extract joint points, and form a skeleton sequence in groups of 30 frames; use the processed skeleton sequence to train the TSSTG network.

[0009] Further, in step 3, performing multi-person posture estimation using the improved multi-person posture estimation model includes: Preprocess the camera input video; Use the improved YoLoPose network to predict the target box and skeleton point coordinates; Use Deep Oc-sort to calibrate the tracking frame numbers of different targets according to the target appearance features; According to the tracking box sequence number output by Deep Oc-sort, the skeleton point data detected by YoLoPose is improved to establish and save the skeleton sequences of different targets, and then the skeleton sequences are input into TSSTG to realize multi-person pose estimation.

[0010] Furthermore, the target box and skeleton point coordinates are predicted using the improved YoLoPose network including: The DAttention is integrated into the Residual part of the last C2f module in the feature extraction network of the YoLoPose network and the last C2f module of the neck, so that the feature map input to the Residual part is first processed by the DAttention and then added to the convolution part to detect tiny bone points; LSKA is added before the last convolution output layer of the SPPF module to eliminate irrelevant background interference; the PANet network of the neck is changed to the PANet-P6 network and Dysample is used to replace the original upsampling; the Anchor-Based detection head is used to compensate for the problem of bone point overlap, and the SEAM attention module is added before the output convolution of the detection head to enhance the detection of occluded bone points.

[0011] Furthermore, the preprocessing of the camera input video includes preprocessing the image, using Opencv to adjust the camera video size to , the frame rate is adjusted to 24 frames / s.

[0012] Furthermore, the method of using the improved YoLoPose network to predict the target frame and the skeleton point coordinates also includes: Use the feature extraction network to extract features from the image and obtain a feature map; Use SPPF_LSKA to eliminate irrelevant background information interference in the feature map and capture more effective feature information: Add Large Separable Kernel Attention before the last convolution output layer of the SPPF module to eliminate irrelevant background information interference and capture more effective feature information; Through the PANet-P6 path aggregation mechanism, top-down path aggregation PA and bottom-up path aggregation PAM are introduced to capture and integrate multi-scale information and output feature maps of four different sizes; The target box and skeleton points are predicted for the obtained feature map, and the improved Anchor-Based detection head is used to perform regression prediction of the target box and skeleton points.

[0013] Furthermore, the step of extracting features from an image using a feature extraction network to obtain a feature map includes: The image captured by the camera is in color and naturally has three channels: R, G, and B. Therefore, multiple convolution operations are performed on the image on the three RGB channels to extract features and obtain three feature maps of different sizes. , , ; The input feature map is mapped to the query weight to generate the query , and after Generate offsets and use The generated uniform grid Reference Points are combined to generate Deformed Points. The formula is as follows: In the formula yes Generate a network, is the weight of the query; Generate simple features using DP and bilinear interpolation , and projected onto the weights of the keys and values, thus generating the shift key Sum , the formula is as follows: in is bilinear interpolation; Will Transferred together to the multi-head attention to generate the weight of the offset , the formula is as follows: In the formula For multiple attention; By offset weight Dynamically adjust the input feature map The sampling position gets the new feature map .

[0014] In the second aspect, an edge computing device includes a memory, a processor, a camera, and a computer program stored in the memory and executable on the processor. When the processor executes the program stored in the memory, an improved multi-person posture estimation method for edge computing devices as described above is implemented.

[0015] In the third aspect, an edge computing device-readable storage medium stores a computer program, and when the computer program is executed by a processor, an improved multi-person posture estimation method for edge computing devices as described above is implemented.

[0016] In a fourth aspect, a video input device is connected to the edge computing device via a serial communication protocol. When the computer program is executed by the processor, the video input device provides a real-time video stream to the edge computing device to implement an improved multi-person posture estimation method for edge computing devices as described above.

[0017] The technical solution adopted by the present invention has the following beneficial effects: Taking into account the limited computational power of edge devices, the improved YoLoPose of the present invention has a lower computational power. Taking into account that when the target is small, the connection between tiny bone points will be ignored, the improved YoLoPose of the present invention obtains more details, ensures that tiny bone points can be detected, and improves the accuracy of bone point detection. Taking into account that human bone points are easily obscured by the background or overlap with bone points in other parts, the improved YoLoPose of the present invention eliminates the interference of irrelevant background information and improves the accuracy of bone point detection. Taking into account that the action is time-related, the present invention uses TSSTG to complete the posture estimation task to increase the accuracy of posture estimation; taking into account the situation where there are multiple targets in the image, it is proposed to use the DeepOc-sort network for multi-target tracking, and combine the three networks to achieve multi-person posture estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the algorithm structure diagram of the present invention; Figure 2 It is a multi-person posture estimation flow chart of the present invention; Figure 3 It is the C2f_DAttention structure diagram; Figure 4 It is the DAttention structure diagram; Figure 5 It is the structure diagram of SPPF_LSKA; Figure 6 This is the LSKA structure diagram; Figure 7 This is the PANet-P6 structure diagram; Figure 8 This is the Dysample upsampling structure diagram; Fig. 9 Flowchart of SEAM module Fig.10 This is the effect diagram of human skeleton point detection; Fig.11This is the effect diagram of posture estimation. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0020] Example 1 An improved multi-person posture estimation method for edge computing devices comprises the following steps: Step 1: Establish an improved multi-person pose estimation model; Step 2: training the improved multi-person pose estimation model to obtain a trained improved multi-person pose estimation model; Step 3: Input the video collected by the camera in real time into the trained improved multi-person pose estimation model, perform multi-person pose estimation using the improved multi-person pose estimation model, and output the multi-person pose estimation result.

[0021] Among them, Figure 1 As shown, the algorithm of the improved multi-person posture estimation model established by the present invention is shown in the figure, and establishing the improved multi-person posture estimation model specifically includes the following steps: Based on YoLoPose, DAttention is integrated into the feature extraction network and the residual part of the last C2f module of the neck to detect tiny bone points; LSKA is integrated into the SPPF module to eliminate irrelevant background interference; the PANet network of the neck is changed to the PANet-P6 network and Dysample is used to replace the original upsampling to ensure efficient upsampling performance while avoiding complex structural design and high computational cost; Anchor-Based detection head is used to make up for the problem of overlapping bone points, and the SEAM attention module is added before the output convolution of the detection head to enhance the detection of occluded bone points. According to the extracted appearance features, Deep Oc-sort is used to calibrate the tracking box sequence of different targets, and the target bone point information is saved, and the bone point information of different targets is input into TSSTG to realize multi-person pose estimation.

[0022] The specific steps of training the improved multi-person posture estimation model include: Step 1: Use the COCO KeyPoints dataset to train the improved YoLoPose network, and use the OKS loss instead of Iou in target detection to perform regression prediction of skeleton points. Calculate OKS for each individual skeleton point and add them to the final OKS loss. The total skeleton point loss function formula is as follows: in, , , , is a hyperparameter to balance the loss between different sizes, They are the losses of classification, target box, skeleton point and skeleton point confidence respectively.

[0023] Step 2: Use the Market-1501 dataset to train the Deep Oc-sort network; Step 3: Select 30 fall sequences from the Fall Detection Dataset, and increase the number of data sets to 90 by flipping and adding noise. Use the trained improved YoLoPose to extract joint points and form a skeleton sequence in groups of 30 frames. Use the processed skeleton sequence to train the TSSTG network.

[0024] The flowchart of using the improved multi-person posture estimation model to perform multi-person posture estimation is as follows: Figure 2 As shown, the following steps are included: Step 1: Preprocess the image: Use Opencv to resize the camera video to , the frame rate is adjusted to 24 frames / s.

[0025] Step 2: Use the feature extraction network to extract features from the image; Step 2 includes the following steps 2.1 to 2.3: Step 2.1: The image acquired by the camera is in color and naturally has three channels: R, G, and B. Therefore, multiple convolution operations are performed on the image on the three channels of RGB to extract features and obtain three feature maps of different sizes. , , ; Step 2.2: Use the improved C2f module to extract the feature map Get more details and detect tiny bone points. Figure 3 As shown, the feature map In passing After the convolution, the feature maps are evenly divided into two feature maps according to the number of channels and stored in a list. The last feature map in the list is input into the Residual part of the C2f module, and the result is added to the list. Finally, the feature maps in the list are concatenated and convolved to obtain the feature map. The last element of the list is processed by DAttention in the Residual part, and then added to the convolution part to obtain more details and ensure that tiny bone points can be detected. Figure 4As shown, it includes the following steps 2.2.1~2.2.4: Step 2.2.1: Map the input feature map to the query weights to generate the query , and after Generate offsets and use The generated uniform grid RP (Reference Points) is combined to generate DP (Deformed Points), the formula is as follows: In the formula yes Generate a network, is the weight of the query.

[0026] Step 2.2.2: Generate simple features using DP and bilinear interpolation , and projected onto the weights of the keys and values, thus generating the shift key Sum , the formula is as follows: in is bilinear interpolation.

[0027] Step 2.2.3: Transferred together to the multi-head attention to generate the weight of the offset , the formula is as follows: In the formula For multiple attention.

[0028] Step 2.2.4: Pass the offset weight from step 2.2.3 Dynamically adjust the input feature map The sampling position gets the new feature map .

[0029] Step 2.3: Use SPPF_LSKA to eliminate the feature map obtained in step 2.2 Add LSKA (Large Separable Kernel Attention) before the last convolution output layer of the SPPF module to eliminate the interference of irrelevant background information and capture more effective feature information. Its structure diagram is as follows Figure 5 As shown, it includes the following steps 2.3.1~2.3.3: Step 2.3.1: For the feature map obtained in step 2.2 conduct Convolution operation, then three maximum pooling operations, and the feature maps after different operations are added point by point to obtain the feature map .

[0030] Step 2.3.2: Use LSKA (Large Separable Kernel Attention) to eliminate irrelevant background information interference, improve semantic fusion between different feature layers, and capture more effective feature information.

[0031] The LSKA structure diagram in step 2.3.2 is as follows Figure 6 As shown, the entire The convolution is split into 2D deep convolution, The two-dimensional depth expansion convolution and point-by-point convolution are combined, and the two-dimensional convolution is split into a cascade of horizontal and vertical one-dimensional convolution combinations, and finally all the one-dimensional convolutions are connected in series. It includes the following steps 2.3.2.1~2.3.2.2: Step 2.3.2.1: Perform the feature maps in order , , , , Convolution operation obtains feature map .

[0032] Step 2.3.2.2: Feature map With feature map Perform Hadamard product to get the feature map Step 2.3.3: Feature map conduct Convolution to get feature map Step 3: Through the PANet-P6 path aggregation mechanism, top-down path aggregation (PA) and bottom-up path aggregation (PAM) are introduced to capture and integrate multi-scale information and output four feature maps of different sizes. Figure 7 As shown, it includes the following steps 3.1~3.3: Step 3.1: Use the top-down path aggregation module (PA) to classify the feature map obtained in step 2 Upsampling is performed sequentially to pass information to the upper layer layer by layer. At each level, the low-resolution feature map and the high-resolution feature map are added point by point. The formula is as follows: In the formula Representative Layer feature map , ; Indicates the upsampling operation. The upsampling is completed by the Dysample state upsampling module, and its structure diagram is as follows Figure 8 As shown, it includes the following steps 3.1.1~3.1.3: Step 3.1.1: Feature map Use two output channels for The linear layer generates two The offset and .right After multiplying by the dynamic range factor, the sigmoid function is used for smoothing to ensure uniform distribution of the initial sampling positions without overlapping the sampling positions. The formula is as follows: The dynamic range factor is set to 0.5. represents a linear layer, Represents the sigmoid function.

[0033] Step 3.1.2: Use Pixel Shuffle to offset and Resize to , and with a two-dimensional grid Add together to generate a size of The original sampling grid set ,in , , the formula is as follows: In the formula For Pixel Shuffel, is the upsampling scale factor, the two-dimensional grid Using functions and parameters in the Pytorch library generate.

[0034] Step 3.1.3: Use the grid sample function and the sampling grid set , for size Feature map Perform bilinear interpolation to generate a value of Upsampling feature maps , the formula is as follows; In the formula Represents the grid_sample function.

[0035] Step 3.2: Using the bottom-up Path Aggregation Module (PAM), starting from the high-resolution feature map, the feature map obtained in step 3.1 Aggregation operations are performed to pass information downward layer by layer. Similarly, at each level, high-resolution feature maps are added point by point with low-resolution feature maps. The formula is as follows: In the formula Representative Layer feature map , , Represents a convolution operation.

[0036] Step 3.3: Output feature maps of different sizes .

[0037] Step 4: The feature map obtained in step 3 Predict the target box and skeleton points. Use the improved Anchor-Based detection head to perform regression prediction of the target box and skeleton points, including the following steps 4.1~4.2: Step 4.1: Use the SEAM (Separated and Enhancement Module) attention module to enhance the relationship between the occluded and unoccluded skeleton points in each feature fusion layer. The flowchart is as follows Fig. 9 The following steps are included as shown in 4.1.1~4.1.3: Step 4.1.1: Extract appearance information and use the AE (Appearance Extraction) module with ordinary convolution and point-by-point convolution with residual connection to fuse the information. The formula is as follows: In the formula It is the appearance feature extraction module (Path Embedding). is a normal convolution, It is pointwise convolution.

[0038] Step 4.1.2: Use the average pooling layer and two fully connected layers to fuse the information of each channel to enhance the information connectivity. The formula is as follows: In the formula is the average pooling layer, is a fully connected layer.

[0039] Step 4.1.3: Input the sigmoid function for normalization, increase the regression error tolerance for the occluded position, and multiply the original features with the processed features to handle the occluded bone points.

[0040] Step 4.2: Convolve the feature map and output the target box and skeleton point coordinates. The result is as follows Fig.10 shown.

[0041] To demonstrate the advantages of the present invention in the skeleton point detection task, skeleton point detection is performed on the COCO KeyPoints test set and compared with the original network. The detection results are shown in Table 1. While GFLOPS is reduced by 8.3, mAP@0.5(%) is increased by 6.41%, and mAP0.5:0.95(%) is increased by 7.49%. Therefore, the improved skeleton point detection network of the present invention is suitable for running on mainstream edge computing platforms.

[0042] Table 1 Comparison of skeleton point detection Step 5: Use Deep Oc-sort to implement multi-person tracking based on the target appearance features obtained in step 4, and calibrate the tracking frame numbers for different targets.

[0043] Step 6: According to the tracking box sequence number output by Deep Oc-sort, save the skeleton point data of different targets detected by the improved YoLoPose, and then input the skeleton point data of different targets into TSSTG respectively to realize multi-person pose estimation. Fig.11 This is the effect diagram of posture estimation.

[0044] Based on YoLoPose, the present invention integrates DAttention attention into the Residual part of the last C2f module of the feature extraction network and the neck network to detect tiny bone points; integrates LSKA into the SPPF module to eliminate irrelevant background interference; changes the PANet network of the neck to the PANet-P6 network and uses Dysample to replace the original upsampling to ensure efficient upsampling performance on the basis of avoiding complex structural design and high computational cost; uses Anchor-Based detection head to make up for the problem of overlapping bone points, and adds SEAM attention module before the output convolution of the detection head to enhance the detection of occluded bone points; the improved YoLoPose network of this method can effectively improve the accuracy of bone point detection and reduce the computational cost. This method uses Deep Oc-sort to calibrate the tracking frame number of different targets according to the target appearance features; according to the tracking frame number, the bone point data is respectively input into TSSTG to realize multi-person posture estimation, which can effectively improve the posture estimation accuracy.

[0045] Example 2 An embodiment of the present invention provides an edge computing device, comprising a memory, a processor, a camera, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, an improved multi-person posture estimation step for the edge computing device is implemented.

[0046] An embodiment of the present invention provides an edge computing device-readable storage medium, which stores a computer program. When the computer program is executed by a processor, an improved multi-person posture estimation step for the edge computing device is implemented.

[0047] An embodiment of the present invention provides a video input device, which is connected to the edge computing device through a serial communication protocol. When a computer program is executed by a processor, the video input device provides a real-time video stream to the edge computing device to implement an improved multi-person posture estimation step for the edge computing device.

[0048] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the schemes disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed in this disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the claims.

Claims

1. An improved multi-person posture estimation method for edge computing devices, characterized in that: The method comprises the following steps: Step 1: Based on the improved YoLoPose network, an improved multi-person pose estimation model is established; Step 2: training the improved multi-person pose estimation model, using the COCO KeyPoints dataset to train the improved YoLoPose network, and obtaining a trained improved multi-person pose estimation model; Step 3: Input the video collected by the camera in real time into the trained improved multi-person posture estimation model, perform multi-person posture estimation using the improved multi-person posture estimation model, and output the multi-person posture estimation result.

2. An improved multi-person posture estimation method for edge computing devices according to claim 1, characterized in that: The establishment of an improved multi-person pose estimation model in step 1 specifically includes the following steps: Step 1.1: Based on YoLoPose, DAttention is integrated into the feature extraction network and the residual part of the last C2f module of the neck to detect tiny bone points; LSKA is integrated into the SPPF module to eliminate irrelevant background interference; the PANet network of the neck is changed to the PANet-P6 network, and Dysample is used to replace the original upsampling to ensure efficient upsampling performance while avoiding complex structural design and high computational cost; Anchor-Based detection head is used to compensate for the problem of overlapping bone points, and the SEAM attention module is added before the output convolution of the detection head to enhance the detection of occluded bone points; Step 1.2: Input the appearance features extracted by the improved YoLoPose into Deep Oc-sort to calibrate the tracking frame numbers of different targets to achieve multi-target tracking; Step 1.3: Save the target skeleton point information according to the tracking frame serial number of different targets, input the target skeleton point information into TSSTG to realize posture estimation, and complete the establishment of the improved multi-person posture estimation model.

3. According to the improved multi-person posture estimation method for edge computing devices described in claim 1, it is characterized in that: Training the improved multi-person posture estimation model in step 2 includes the following steps: Step 2.1: Use the COCO KeyPoints dataset to train the improved YoLoPose network, and use the OKS loss instead of Iou in target detection to perform regression prediction of skeleton points; calculate OKS for each individual skeleton point and accumulate it to the final OKS loss; the total skeleton point loss function formula is as follows: in, , , , is a hyperparameter to balance the loss between different sizes, They are the losses of classification, target box, skeleton point and skeleton point confidence respectively; Step 2.2: Train the Deep Oc-sort network using the Market-1501 dataset. Step 2.3: Select 30 fall sequences from the Fall Detection Dataset, and increase the number of data sets to 90 by flipping and adding noise; use the trained improved YoLoPose to extract joint points, and form a skeleton sequence in groups of 30 frames; use the processed skeleton sequence to train the TSSTG network.

4. An improved multi-person posture estimation method for edge computing devices according to claim 1, characterized in that: In step 3, performing multi-person posture estimation using the improved multi-person posture estimation model includes: Preprocess the camera input video; Use the improved YoLoPose network to predict the target box and skeleton point coordinates; Use Deep Oc-sort to calibrate the tracking frame numbers of different targets according to the target appearance features; According to the tracking box sequence number output by Deep Oc-sort, the skeleton point data detected by YoLoPose is improved to establish and save the skeleton sequences of different targets, and then the skeleton sequences are input into TSSTG to realize multi-person pose estimation.

5. An improved multi-person posture estimation method for edge computing devices according to claim 4, characterized in that: Using the improved YoLoPose network to predict target boxes and skeleton point coordinates includes: The DAttention is integrated into the Residual part of the last C2f module in the feature extraction network of the YoLoPose network and the last C2f module of the neck, so that the feature map input to the Residual part is first processed by the DAttention and then added to the convolution part to detect tiny bone points; LSKA is added before the last convolution output layer of the SPPF module to eliminate irrelevant background interference; the PANet network of the neck is changed to the PANet-P6 network and Dysample is used to replace the original upsampling; the Anchor-Based detection head is used to compensate for the problem of bone point overlap, and the SEAM attention module is added before the output convolution of the detection head to enhance the detection of occluded bone points.

6. An improved multi-person posture estimation method for edge computing devices according to claim 4, characterized in that: The preprocessing of the camera input video includes preprocessing the image, using Opencv to adjust the camera video size to , the frame rate is adjusted to 24 frames / s.

7. An improved multi-person posture estimation method for edge computing devices according to claim 4, characterized in that: The use of the improved YoLoPose network to predict the target frame and the skeleton point coordinates also includes: Use the feature extraction network to extract features from the image and obtain a feature map; Use SPPF_LSKA to eliminate irrelevant background information interference in the feature map and capture more effective feature information: Add Large Separable Kernel Attention before the last convolution output layer of the SPPF module to eliminate irrelevant background information interference and capture more effective feature information; Through the PANet-P6 path aggregation mechanism, top-down path aggregation PA and bottom-up path aggregation PAM are introduced to capture and integrate multi-scale information and output feature maps of four different sizes; The target box and skeleton points are predicted for the obtained feature map, and the improved Anchor-Based detection head is used to perform regression prediction of the target box and skeleton points.

8. An improved multi-person posture estimation method for edge computing devices according to claim 7, characterized in that: The feature extraction network is used to extract features from the image to obtain a feature map, which includes: The image captured by the camera is in color and naturally has three channels: R, G, and B. Therefore, multiple convolution operations are performed on the image on the three RGB channels to extract features and obtain three feature maps of different sizes. , , ; The input feature map is mapped to the query weight to generate the query , and after Generate offsets and use The generated uniform grid Reference Points are combined to generate Deformed Points. The formula is as follows: In the formula yes Generate a network, is the weight of the query; Generate simple features using DP and bilinear interpolation , and projected onto the weights of the keys and values, thus generating the shift key Sum , the formula is as follows: in is bilinear interpolation; Will Transferred together to the multi-head attention to generate the weight of the offset , the formula is as follows: In the formula For multiple attention; By offset weight Dynamically adjust the input feature map The sampling position gets the new feature map .

Images