A posture estimation method based on improved YOLOV8 algorithm

Abstract

The present invention relates to the field of posture estimation technology, and in particular to a posture estimation method based on an improved YOLOV8 algorithm, comprising: step 1, using a SimDLKA module to simplify the process of a DLKA module containing an attention mechanism; step 2, replacing the original Bottleneck module of the YOLOv8 network structure with a SimDLKA module that improves detection accuracy, step 3, modifying the original loss function, and modifying the CIOU loss function to a DCIOU loss function; step 4, autonomously constructing a large-scale posture estimation data set HP, using a variety of data enhancement strategies, and using open source COCO data sets and MPII data sets for model training. The present invention optimizes the model's selective attention to data input by introducing the SimDLKA attention mechanism, and enhances the model's global semantic fusion capability when processing medium and large target information; it can improve the real-time and accuracy of posture estimation, while reducing the computational complexity of the algorithm, making it suitable for edge devices.

Description

Technical Field

[0001] The present invention relates to the technical field of attitude estimation, and in particular to an attitude estimation method based on an improved YOLOV8 algorithm. Background Art

[0002] Pose estimation has its application prospects in many fields, and the detection of pose estimation is a topic of great concern. The detection of traditional algorithms is not only time-consuming and laborious, but also has the disadvantages of insufficient detection accuracy and objectivity. The YOLO algorithm (You Only Look Once) is a target detection algorithm that can simultaneously complete the tasks of object detection and classification in a single neural network. The YOLOv8 algorithm is a more mature version of the current cutting-edge algorithm. Its algorithm has certain advantages over the traditional version in terms of speed, accuracy and adaptability, and the YOLOv8 algorithm has certain room for improvement in the speed and accuracy of the model.

[0003] In the existing field of pose estimation, although there are many solutions such as OpenPose and DeepCut, these methods have many shortcomings, such as high computational complexity, lack of real-time performance, and low accuracy in complex environments. Especially in multi-person pose estimation, the processing of human body overlap, light and shadow changes, and occlusion of key parts is still not ideal. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings existing in the prior art, and propose a posture estimation method based on an improved YOLOV8 algorithm, which can improve the real-time and accuracy of posture estimation, while reducing the computational complexity of the algorithm, making it suitable for edge devices.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A posture estimation method based on the improved YOLOV8 algorithm, the specific steps are as follows:

[0007] Step 1: Use the SimDLKA module to simplify the process of the DLKA module containing the attention mechanism and maintain the original performance of the DLKA module; merge the convolution layer and the normalization layer of the DLKA module to simplify the structure of the DLKA module;

[0008] Step 2: Replace the original Bottleneck module of the YOLOv8 network structure with the SimDLKA module that improves detection accuracy, and then merge the SimDLKA module with the C2F module to form a C2F-SimDLKA module, which is embedded in and replaces the original C2F module of the YOLOv8 network architecture;

[0009] Step 3: Modify the original loss function and change the CIOU loss function to the DCIOU loss function;

[0010] Step 4: We independently built a large-scale pose estimation dataset HP (Human Pose), adopted a variety of data enhancement strategies, and used the open source COCO dataset and MPII dataset for model training, which provided a rich source of data for the pose estimation task of the model.

[0011] Preferably, in step 1, the following steps are specifically included:

[0012] Step 1.1: For a batch of data, the steps for merging the convolutional layer and the normalization layer are as follows:

[0013]

[0014] x i is the i-th input; is the i-th output; γ is the scaling factor; β is the translation factor; μ is the mean of the input batch; σ 2 is the variance of the input batch; ∈ is a very small quantity to avoid the denominator being 0;

[0015] For C feature maps F, the normalization process is written as follows:

[0016]

[0017] is the Cth output feature map with height i and width j; F C，i，j It is the Cth input feature map with height i and width j;

[0018] The normalized result is directly integrated into the convolution operation, and the equation is as follows:

[0019]

[0020] in, is the weight parameter during convolution; is the bias during normalization; is the weight parameter during normalization; is the bias during convolution; con1×1 is the convolution kernel of size 1×1; C pre is the number of channels of the input layer;

[0021] k is the size of the convolution kernel;

[0022] Decompose the equation to get the updated weight parameter W and bias b:

[0023] W=W conv1×1 ·W BN

[0024] b=W conv1×1 b BN +b conv1×1

[0025] Combine the normalization layer and convolution layer in the above equation, and the above process is called BN-Conv operation;

[0026] Step 1.2, the feature map of the previous layer is calculated using a deformable depth convolution of 5x5 size, which can adaptively adjust the shape of the convolution kernel to adapt to the spatial transformation of the input features, so as to better capture irregularly shaped objects or features; the convolution operation is divided into multiple channel groups, and the size of each group is determined by the dimension of the feature map of the previous layer input; the calculated feature map is calculated using a deformable depth convolution of 7x7 size, and the number of groups of grouped convolution is also equal to the dimension. The expansion rate of the convolution is set to 3, which expands the receptive field of the convolution kernel without increasing the number of parameters; then the feature map of the previous layer is calculated using a convolution kernel of 1x1 size, and the output result is concatenated with the feature map of the initial input for feature concatenation in the residual network ResNet, aiming to integrate multi-scale features, improve information flow, and avoid information loss in deep networks; the above process is called the DLKA module;

[0027] Step 1.3, divide the feature map of the previous layer into two identical parts, perform BN-Conv operation on one part, then perform GELU activation operation, perform the process of DLKA module on the result, and then use 1x1 convolution kernel for calculation; concatenate the other feature map with this result; similarly, divide the result into two identical parts, perform BN-Conv operation on one part, use 3x3, dilation rate 3 hole convolution for convolution, perform GELU activation operation on the result, and then use 1x1 convolution kernel for calculation; concatenate the result with the original feature map; the above process is called SimDLKA module;

[0028] Step 1.4, convolve the feature map of the previous layer with a convolution module with a convolution kernel size of 1x1, a stride of 1, a padding of 0, and an output channel number of c_out; divide the result into two equal parts, one feature map is not processed; the other part is divided into two equal parts after the Split operation, and the number of channels is reduced to half of the original, and the result is divided into two equal parts, one feature map is not processed; the other is passed to the SimDLKA module for feature fusion operation; then divide the result into two equal parts, one feature map is not processed; the other is then passed to the SimD The feature fusion operation is performed in the LKA module; the subsequent operations are the same as above; each SimDLKA module will separate into two channels, one is to pass the processed features to the next SimDLKA module, and the other is retained for subsequent splicing; finally, after passing through n SimDLKA modules, all unprocessed feature maps are spliced; finally, the result is convolved with a convolution module with a convolution kernel size of 1x1, a stride of 1, a padding of 0, and an output channel number of c_out; the above process is called the C2F-SimDLKA module.

[0029] Preferably, in step 3, the following steps are specifically included:

[0030] Step 3.1, the original YOLO-pose based on pose estimation test uses the CIOU loss function, which is as follows:

[0031]

[0032]

[0033]

[0034] Among them, w gt 、h gt and w, h are the width and height of the target box and the predicted box respectively; a is a weight coefficient used to balance the importance of aspect ratio consistency; v is a term that measures the difference in aspect ratio of the bounding box;

[0035] Step 3.2: Since the CIOU loss function also has certain limitations in some aspects, when the height and width of the target box and the prediction box are exactly the same, that is, w gt =w,h gt =h, then v becomes zero, and the CIOU loss function will degenerate into the DIOU loss function, which reduces the convergence ability of the function in some aspects; here we use a new loss function DCIOU (Double-Complete-IOU), which retains the advantages of the CIOU loss function and avoids the disadvantages of the CIOU loss function degradation, and is defined as follows:

[0036]

[0037]

[0038]

[0039] Here, ∈ is a very small quantity, and the v value of the DCIOU loss function will inherit the advantage of the original v value in measuring the difference in the aspect ratio of the bounding box.

[0040] and They can never be 0 at the same time, that is, the model will consider the importance of aspect ratio consistency at any time, which increases the convergence ability of the model and avoids the risk of loss function degradation in some cases.

[0041] Preferably, in step 4, the following steps are specifically included:

[0042] Step 4.1, HP (Human Pose) dataset is a self-constructed human pose estimation dataset with a large amount of data, in order to provide a high-quality and diverse resource to test the performance of the model. Here, more than 5,000 images are captured and collected, covering individuals of different ages, genders, postures and body shapes, to ensure the wide adaptability and robustness of the algorithm;

[0043] Step 4.2, the above images cover various life scenes, including indoor and outdoor environments, different lighting conditions and various activity states (such as running, dancing, sitting, etc.); the position of each key point is manually annotated by professionals, and each image is accurately annotated with 20 key points of the human body, such as head, shoulders, knees and feet, to ensure the accuracy of the annotation;

[0044] Step 4.3: In order to improve the robustness and generalization ability of the model, the HP dataset is then subjected to a variety of data augmentation strategies, including random rotation, scaling, flipping, and color adjustment (by adjusting contrast, brightness, and saturation). The HP training set images are rotated at random angles (-30° to 30°) to enable the model to learn to recognize objects at different angles, and the images are randomly scaled (0.8 to 1.2 times) to simulate the effect of distance changes of objects in the visual scene. Some images are horizontally flipped to change their visual appearance without losing their semantic content. The contrast, brightness, and saturation of all images are adjusted to simulate the visual effects under different lighting conditions and increase the model's adaptability to lighting changes.

[0045] By adopting the above technical solutions: By adding a proposed SimDLKA attention mechanism on the basis of the original YOLOv8, the model selectively pays attention to the data input, improving the decoupling and flexibility of the model. For the feature fusion module of YOLOv8, the SimDLKA module replaces the original Bottleneck module of YOLOv8, and is fused with the C2F module to form C2F-SimDLKA, which can more effectively integrate global semantics, especially medium and large target information. At the same time, based on the loss function in YOLOv8, a new loss function DCIOU is proposed to improve the forward propagation of model training, providing a clearer optimization direction signal for gradient descent.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. By introducing the SimDLKA attention mechanism, the present invention optimizes the model's selective attention to data input and enhances the model's global semantic fusion capability when processing medium and large-scale target information.

[0048] 2. The present invention adopts the improved C2F-SimDLKA module and the new DCIOU loss function to further optimize the model training process, providing a clearer optimization direction, so that the model can converge faster and improve the generalization ability.

[0049] 3. The present invention adopts an integrated SimDLKA attention mechanism to enable the model to selectively focus on important features of the input data, thereby improving its flexibility and decoupling ability. The new loss function DCIOU introduced provides a clearer optimization direction signal for gradient descent and improves the forward propagation of model training. This enhancement enables the model to converge 3 to 4 epochs earlier than the original CIOU loss function.

[0050] 4. The present invention optimizes the model's ability to handle image changes and complex scenarios by introducing large-core attention and deformable convolution; the optimization of its model reduces the computational load, allowing the YOLOv8 algorithm to be embedded in devices with limited processing capabilities, thereby ensuring real-time detection.

[0051] 5. The present invention can improve the real-time and accuracy of posture estimation, while reducing the computational complexity of the algorithm, making it suitable for edge devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is the network architecture diagram of YOLOv8 in the present invention;

[0053] Figure 2 It is the structural diagram of C2F-SimDLKA module in the present invention;

[0054] Figure 3 It is a convergence curve diagram of CIOU loss function and DIOU loss function in the present invention;

[0055] Figure 4 The state diagram shows the posture estimation results under the YOLOv8 algorithm in the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the protection scope of the present invention. The embodiments described in the present invention are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0057] Reference Figure 1-Figure 4 , a posture estimation method based on the improved YOLOV8 algorithm, the specific steps are as follows:

[0058] Step 1: Use the SimDLKA module to simplify the process of the DLKA module containing the attention mechanism and maintain the original performance of the DLKA module; merge the convolution layer and the normalization layer of the DLKA module to simplify the structure of the DLKA module;

[0059] Step 2: Replace the original Bottleneck module of the YOLOv8 network structure with the SimDLKA module that improves detection accuracy, and then merge the SimDLKA module with the C2F module to form a C2F-SimDLKA module, which is embedded in and replaces the original C2F module of the YOLOv8 network architecture;

[0060] Step 3: Modify the original loss function and change the CIOU loss function to the DCIOU loss function;

[0061] Step 4: We independently built a large-scale pose estimation dataset HP (Human Pose), adopted a variety of data enhancement strategies, and used the open source COCO dataset and MPII dataset for model training, which provided a rich source of data for the pose estimation task of the model.

[0062] Specifically, in step 1, the following steps are specifically included:

[0063] Step 1.1: For a batch of data, the steps for merging the convolutional layer and the normalization layer are as follows:

[0064]

[0065] xi is the i-th input; is the i-th output; γ is the scaling factor; β is the translation factor; μ is the mean of the input batch; σ 2 is the variance of the input batch; ∈ is a very small quantity to avoid the denominator being 0;

[0066] For C feature maps F, the normalization process is written as follows:

[0067]

[0068] is the Cth output feature map with height i and width j; T C，i，j It is the Cth input feature map with height i and width j;

[0069] The normalized result is directly integrated into the convolution operation, and the equation is as follows:

[0070]

[0071] in, is the weight parameter during convolution; is the bias during normalization; is the weight parameter during normalization; is the bias during convolution; con1×1 is the convolution kernel of size 1×1; C pre is the number of channels of the input layer;

[0072] k is the size of the convolution kernel;

[0073] Decompose the equation to get the updated weight parameter W and bias b:

[0074] W=W conv1×1 ·W BN

[0075] b=W conv1×1 b BN +b conv1×1

[0076] Combine the normalization layer and convolution layer in the above equation, and the above process is called BN-Conv operation;

[0077] Step 1.2, the feature map of the previous layer is calculated using a deformable depth convolution of 5x5 size, which can adaptively adjust the shape of the convolution kernel to adapt to the spatial transformation of the input features, so as to better capture irregularly shaped objects or features; the convolution operation is divided into multiple channel groups, and the size of each group is determined by the dimension of the feature map of the previous layer input; the calculated feature map is calculated using a deformable depth convolution of 7x7 size, and the number of groups of grouped convolution is also equal to the dimension. The expansion rate of the convolution is set to 3, which expands the receptive field of the convolution kernel without increasing the number of parameters; then the feature map of the previous layer is calculated using a convolution kernel of 1x1 size, and the output result is concatenated with the feature map of the initial input for feature concatenation in the residual network ResNet, aiming to integrate multi-scale features, improve information flow, and avoid information loss in deep networks; the above process is called the DLKA module;

[0078] Step 1.3, divide the feature map of the previous layer into two identical parts, perform BN-Conv operation on one part, then perform GELU activation operation, perform the process of DLKA module on the result, and then use 1x1 convolution kernel for calculation; concatenate the other feature map with this result; similarly, divide the result into two identical parts, perform BN-Conv operation on one part, use 3x3, dilation rate 3 hole convolution for convolution, perform GELU activation operation on the result, and then use 1x1 convolution kernel for calculation; concatenate the result with the original feature map; the above process is called SimDLKA module;

[0079] Step 1.4, convolve the feature map of the previous layer with a convolution module with a convolution kernel size of 1x1, a stride of 1, a padding of 0, and an output channel number of c_out; divide the result into two equal parts, one feature map is not processed; the other part is divided into two equal parts after the Split operation, and the number of channels is reduced to half of the original, and the result is divided into two equal parts, one feature map is not processed; the other is passed to the SimDLKA module for feature fusion operation; then divide the result into two equal parts, one feature map is not processed; the other is then passed to the SimD The feature fusion operation is performed in the LKA module; the subsequent operations are the same as above; each SimDLKA module will separate into two channels, one is to pass the processed features to the next SimDLKA module, and the other is retained for subsequent splicing; finally, after passing through n SimDLKA modules, all unprocessed feature maps are spliced; finally, the result is convolved with a convolution module with a convolution kernel size of 1x1, a stride of 1, a padding of 0, and an output channel number of c_out; the above process is called the C2F-SimDLKA module.

[0080] Specifically, in step 3, the following steps are specifically included:

[0081] Step 3.1, the original YOLO-pose based on pose estimation test uses the CIOU loss function, which is as follows:

[0082]

[0083]

[0084]

[0085] Among them, w gt 、h gt and w, h are the width and height of the target box and the predicted box respectively; a is a weight coefficient used to balance the importance of aspect ratio consistency; v is a term that measures the difference in aspect ratio of the bounding box;

[0086] Step 3.2: Since the CIOU loss function also has certain limitations in some aspects, when the height and width of the target box and the prediction box are exactly the same, that is, w gt =w,h gt =h, then v becomes zero, and the CIOU loss function will degenerate into the DIOU loss function, which reduces the convergence ability of the function in some aspects; here we use a new loss function DCIOU (Double-Complete-IOU), which retains the advantages of the CIOU loss function and avoids the disadvantages of the CIOU loss function degradation, and is defined as follows:

[0087]

[0088]

[0089]

[0090] Here, ∈ is a very small quantity, and the v value of the DCIOU loss function will inherit the advantage of the original v value in measuring the difference in the aspect ratio of the bounding box.

[0091] and They can never be 0 at the same time, that is, the model will consider the importance of aspect ratio consistency at any time, which increases the convergence ability of the model and avoids the risk of loss function degradation in some cases.

[0092] Specifically, in step 4, the following steps are specifically included:

[0093] Step 4.1, HP (Human Pose) dataset is a self-constructed human pose estimation dataset with a large amount of data, in order to provide a high-quality and diverse resource to test the performance of the model. Here, more than 5,000 images are captured and collected, covering individuals of different ages, genders, postures and body shapes, to ensure the wide adaptability and robustness of the algorithm;

[0094] Step 4.2, the above images cover various life scenes, including indoor and outdoor environments, different lighting conditions and various activity states (such as running, dancing, sitting, etc.); the position of each key point is manually annotated by professionals, and each image is accurately annotated with 20 key points of the human body, such as head, shoulders, knees and feet, to ensure the accuracy of the annotation;

[0095] Step 4.3: In order to improve the robustness and generalization ability of the model, the HP dataset is then subjected to a variety of data augmentation strategies, including random rotation, scaling, flipping, and color adjustment (by adjusting contrast, brightness, and saturation). The HP training set images are rotated at random angles (-30° to 30°) to enable the model to learn to recognize objects at different angles, and the images are randomly scaled (0.8 to 1.2 times) to simulate the effect of distance changes of objects in the visual scene. Some images are horizontally flipped to change their visual appearance without losing their semantic content. The contrast, brightness, and saturation of all images are adjusted to simulate the visual effects under different lighting conditions and increase the model's adaptability to lighting changes.

[0096] In this embodiment, by adding a proposed SimDLKA attention mechanism on the basis of the original YOLOv8, the model selectively pays attention to the data input, thereby improving the decoupling and flexibility of the model. For the feature fusion module of YOLOv8, the SimDLKA module replaces the original Bottleneck module of YOLOv8, and is fused with the C2F module to form C2F-SimDLKA, which can more effectively fuse global semantics, especially medium and large target information. At the same time, based on the loss function in YOLOv8, a new loss function DCIOU is proposed to improve the forward propagation of model training, providing a clearer optimization direction signal for gradient descent.

[0097] In practical applications, the present invention utilizes multiple datasets including COCO, MPII, and customized HP datasets to improve the generalization and accuracy of the model. The laboratory environment is equipped with high-performance hardware and software to support complex deep learning tasks. The system runs on the Windows 10 operating system and is equipped with an NVIDIA RTX 3060TI graphics card, which can effectively process large-scale datasets.

[0098] In addition, in order to maximize the performance of the GPU, CUDA version 11.4 was used, which can greatly accelerate the training and reasoning of the model by using parallel processing. In terms of programming language, Python 3.8.8 was selected. In addition, PyTorch 1.10.0 was used as the main deep learning framework.

[0099] In summary, the present invention can improve the real-time and accuracy of posture estimation, while reducing the computational complexity of the algorithm, making it suitable for edge devices.

[0100] The description and practice disclosed in the present invention are easy to think and understand for ordinary technicians in the technical field, and several improvements and modifications can be made without departing from the principles of the present invention. Therefore, modifications or improvements made without departing from the spirit of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A posture estimation method based on an improved YOLOV8 algorithm, characterized in that: The specific steps are as follows: Step 1: Use the SimDLKA module to simplify the process of the DLKA module containing the attention mechanism and maintain the original performance of the DLKA module; merge the convolution layer and the normalization layer of the DLKA module to simplify the structure of the DLKA module; Step 2: Replace the original Bottleneck module of the YOLOv8 network structure with the SimDLKA module that improves detection accuracy, and then merge the SimDLKA module with the C2F module to form a C2F-SimDLKA module, which is embedded in and replaces the original C2F module of the YOLOv8 network architecture; Step 3: Modify the original loss function and change the CIOU loss function to the DCIOU loss function; Step 4: Independently build a large-scale pose estimation dataset HP, adopt multiple data enhancement strategies, and use the open source COCO dataset and MPII dataset for model training.

2. A posture estimation method based on an improved YOLOV8 algorithm according to claim 1, characterized in that: In step 1, the following steps are specifically included: Step 1.1: For a batch of data, the steps for merging the convolutional layer and the normalization layer are as follows: x i is the i-th input; is the i-th output; γ is the scaling factor; β is the translation factor; μ is the mean of the input batch; σ 2 is the variance of the input batch; ∈ is a very small quantity to avoid the denominator being 0; For C feature maps F, the normalization process is written as follows: is the Cth output feature map with height i and width j; F C，i，j It is the cth input feature map with height i and width j; The normalized result is directly integrated into the convolution operation, and the equation is as follows: in, is the weight parameter during convolution; is the bias during normalization; is the weight parameter during normalization; is the bias during convolution; con1×1 is the convolution kernel of size 1×1; C pre is the number of channels in the input layer; k is the size of the convolution kernel; Decompose the equation to get the updated weight parameter W and bias b: W=W conv1×1 ·IN BN b=W conv1×1 ·b BN +b conv1×1 Combine the normalization layer and convolution layer in the above equation, and the above process is called BN-Conv operation; Step 1.2, the feature map of the previous layer is calculated using a deformable depth convolution of 5x5 size, which can adaptively adjust the shape of the convolution kernel to adapt to the spatial transformation of the input features, so as to better capture irregularly shaped objects or features; the convolution operation is divided into multiple channel groups, and the size of each group is determined by the dimension of the feature map of the previous layer input; the calculated feature map is calculated using a deformable depth convolution of 7x7 size, and the number of groups of grouped convolution is also equal to the dimension. The expansion rate of the convolution is set to 3, which expands the receptive field of the convolution kernel without increasing the number of parameters; then the feature map of the previous layer is calculated using a convolution kernel of 1x1 size, and the output result is concatenated with the feature map of the initial input for feature concatenation in the residual network ResNet, aiming to integrate multi-scale features, improve information flow, and avoid information loss in deep networks; the above process is called the DLKA module; Step 1.3, divide the feature map of the previous layer into two identical parts, perform BN-Conv operation on one part, then perform GELU activation operation, perform the process of DLKA module on the result, and then use 1x1 convolution kernel for calculation; concatenate the other feature map with this result; similarly, divide the result into two identical parts, perform BN-Conv operation on one part, use 3x3, dilation rate 3 hole convolution for convolution, perform GELU activation operation on the result, and then use 1x1 convolution kernel for calculation; concatenate the result with the original feature map; the above process is called SimDLKA module; Step 1.4, convolve the feature map of the previous layer with a convolution module with a convolution kernel size of 1x1, a stride of 1, a padding of 0, and an output channel number of c_out; divide the result into two equal parts, one feature map is not processed; the other part is divided into two equal parts after the Split operation, and the number of channels is reduced to half of the original, and the result is divided into two equal parts, one feature map is not processed; the other is passed to the SimDLKA module for feature fusion operation; then divide the result into two equal parts, one feature map is not processed; the other is then passed to the SimD The feature fusion operation is performed in the LKA module; the subsequent operations are the same as above; each SimDLKA module will separate into two channels, one is to pass the processed features to the next SimDLKA module, and the other is retained for subsequent splicing; finally, after passing through n SimDLKA modules, all unprocessed feature maps are spliced; finally, the result is convolved with a convolution module with a convolution kernel size of 1x1, a stride of 1, a padding of 0, and an output channel number of c_out; the above process is called the C2F-SimDLKA module.

3. A posture estimation method based on an improved YOLOV8 algorithm according to claim 1, characterized in that: In step 3, the following steps are specifically included: Step 3.1, the original YOLO-pose based on pose estimation test uses the CIOU loss function, which is as follows: Among them, w gt 、h gt and w, h are the width and height of the target box and the predicted box respectively; a is a weight coefficient used to balance the importance of aspect ratio consistency; v is a term that measures the difference in aspect ratio of the bounding box; Step 3.2: Use the new loss function DCIOU, which retains the advantages of the CIOU loss function while avoiding the disadvantages of the CIOU loss function degradation. It is defined as follows: Among them, ∈ is a very small quantity, and the v value of the DCIOU loss function will inherit the advantage of the original v value in measuring the difference in the aspect ratio of the bounding box.

4. A posture estimation method based on an improved YOLOV8 algorithm according to claim 1, characterized in that: In step 4, the following steps are specifically included: Step 4.1, HP dataset is a self-constructed human posture estimation dataset with a large amount of data. By capturing and collecting more than 5000 images, it covers individuals of different ages, genders, postures and body shapes to ensure the wide adaptability and robustness of the algorithm. Step 4.2, the above images cover various life scenes, including indoor and outdoor environments, different lighting conditions and various activity states; the position of each key point is manually annotated by professionals, and each image is accurately annotated with 20 key points of the human body to ensure the accuracy of the annotation; Step 4.3: Then, the HP dataset is subjected to a variety of data augmentation strategies, including random rotation, scaling, flipping, and color adjustment. The HP training set images are rotated at random angles so that the model can learn to recognize objects at different angles, and the images are randomly scaled to simulate the effect of distance changes of objects in the visual scene. Some images are horizontally flipped to change their visual appearance without losing their semantic content. The contrast, brightness, and saturation of all images are adjusted to simulate the visual effects under different lighting conditions and increase the model's adaptability to lighting changes.

Images