Method for constructing a virtual dataset for human foot pose detection based on a supr model

Abstract

The application discloses a kind of based on SUPR model human foot posture detection virtual data set's construction method, comprising: using disclosed SUPR model, human skin map, human shape and posture data, and scene classification picture data set as construction material;Using motion classification and random assignment method generates the foot posture data that conforms to human foot physiological structure and real motion state as construction material;Using commercial software converts SUPR model into editable general 3D skin grid model object;Using prepared material edits model object, builds 3D scene, generates the high-definition RGB picture of edited model foot and the matching 2D bounding box and key point annotation data.The application fills the gap of existing human posture data set in foot posture, greatly simplifies the production process of posture data set on the basis of ensuring availability, reduces the hardware threshold and cost required for dataset production.

Description

Technical Field

[0001] The present invention relates to the technical field of object recognition and posture detection, and in particular to a method for constructing a virtual data set for human foot posture detection based on a SUPR model. Background Art

[0002] Human pose detection is an important research area in computer vision. Its goal is to infer human pose information, including joint positions and postures, by analyzing the relationships between human body parts in images or videos. In recent years, thanks to the rapid development of deep learning technology, human pose detection has made significant progress.

[0003] Training deep learning networks requires a large amount of labeled data. Existing pose detection datasets only provide detailed annotations of key points on the body, hands, and head, but are overly concise when it comes to foot poses. This makes existing pose detection networks unable to detect complex foot poses. A dataset with detailed annotations of complex foot poses is still lacking. Summary of the Invention

[0004] The purpose of this invention is to fill the gaps in the existing posture datasets in terms of human foot posture. A method for constructing a virtual dataset for human foot posture detection based on the SUPR model is proposed, which can generate a practical and usable virtual dataset for human foot posture detection under extremely low hardware conditions.

[0005] To achieve the above-mentioned object, the present invention provides a technical solution: a method for constructing a virtual dataset for human foot posture detection based on a SUPR model, comprising the following steps:

[0006] S1: Set the rotation range of the foot joints of the SUPR model to determine the overall posture space of the foot;

[0007] S2: Use classification methods to reduce the dimension of the overall foot posture space to obtain composite posture classes and the posture space corresponding to each class;

[0008] S3: Traverse the composite posture classes and generate a specified number of random foot posture parameters in the posture space of each class using the random assignment method. The parameters generated in all classes are stored together to form a random foot posture parameter dataset.

[0009] S4: Use the parameters in the random foot posture parameter dataset to replace the values ​​of the corresponding positions in the posture parameters obtained by fitting the existing human body shape and posture data with the SUPR model, and obtain a SUPR model with the real shape, body posture and random foot posture;

[0010] S5: Use the SUPR model obtained in step S4 to build a 3D scene, including model mapping, camera settings, lighting settings, and rendering settings, to obtain a complete rendering pipeline;

[0011] S6: Generate an RGB image of a virtual human foot using the obtained rendering pipeline. Calculate the 2D pixel coordinates of the foot joints and the pixel coordinates of the upper left and lower right corners of the bounding box as label information for the RGB image. An RGB image and its label information constitute an instance of a virtual dataset for human foot posture detection.

[0012] S7: Repeat steps S4 to S6 to generate all instances, and the required virtual dataset for human foot posture detection based on the SUPR model can be obtained, that is, a complete virtual dataset for human foot posture detection.

[0013] Further, in step S1, the foot joints of the SUPR model include the ankle joint J ankle , midfoot jointJ mid and toe joints[J toe1 ,J toe2 ,J toe3 ,J toe4 ,J toe5 ,J toe6 ,J toe7 ,J toe8 ,J toe9 ,J toe10 ], where J toe1 It is the base joint of the big toe, J toe2 The middle joint of the big toe, J toe3 The base joint of the second toe, J toe4 The middle joint of the second toe, J toe5 The base joint of the third toe, J toe6 The middle joint of the third toe, J toe7 The base joint of the fourth toe, J toe8 The middle joint of the fourth toe, J toe9 It is the base joint of the little toe, J toe10 It is the middle joint of the little toe, a total of 12 joints; the posture parameter of each joint is a vector used to represent the Euler rotation angle of the joint Among them, θ x is the rotation angle about the x-axis, θ y is the rotation angle about the y-axis, θ z is the rotation angle on the z-axis; the posture parameter of the whole foot That is, the joint posture parameters are concatenated in sequence through the concat operation. in, It is the ankle joint J ankle The posture parameters, The midfoot joint mid The posture parameters, It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s posture parameters; by setting θ x ,θ y and θ z The value range of is used to limit the rotation range of the joint on the x, y, and z axes. The 3D space determined in this way is the posture space S of the joint; the sum of the posture spaces of all joints is the overall posture space S of the foot. foot =S ankle +S mid +S [toe1,toe2,...,toe10] , where S foot is the overall posture space of the foot, S ankle It is the ankle joint J ankle The posture space, S mid The midfoot joint mid The posture space, S [toe1,toe2,...,toe10] It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s posture space.

[0014] Furthermore, in step S2, the specific operation of using the classification method to reduce the dimension of the overall foot posture space is as follows:

[0015] S21: According to the physiological structure and movement mode of human feet, the initial posture space S T-pose , plantar flexion posture space S Plantarflex , back-bend posture space S Dorsiflex , inversion posture space S Inversion , eversion posture space S Eversion , internal rotation posture space S MedialRotation , external rotation posture space S LateralRotation , toe bending posture space S ToeFlexion , toe curling posture space S ToeExtension , toes open posture space S ToeAbduction and toes gathered posture space S ToeAdduction These 11 mutually independent posture spaces are regarded as the overall foot posture space S set in step S1. foot The principal component basis of S is generated by linearly mixing the posture parameters under these 11 principal component bases. foot Most of the posture parameters under , the posture type represented by these 11 principal component bases is called the basic posture class;

[0016] S22: The posture spaces of the 11 basic posture classes divided in step S21 are further transformed into the overall posture space of the foot S by linear blending. foot Represented as the toes-inward posture space Toe-tightening and outward-turning posture space Tiptoe inversion posture space Tiptoe eversion posture space Tiptoe inversion posture space Toe-turning posture space Toe inward posture space Toe-turning posture space and the supplementary posture space S completion There are 9 bases representing the posture space, among which:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] The posture type represented by the above 9 bases is called a composite posture class.

[0027] Furthermore, in step S3, the specific steps of generating a random foot posture parameter dataset are as follows:

[0028] S31: Traverse the composite posture classes divided in step S2 to obtain the posture space corresponding to each composite posture class, that is, the value range of the rotation angle of each joint on the x, y, and z axes;

[0029] S32: Assign the rotation angle value in the form of random rounding within the range of the rotation angle; concatenate the rotation angle into the random foot posture parameter through the concat operation in, is the foot random posture parameter, It is the ankle joint J ankle The random pose parameters of The midfoot jointmid The random pose parameters of It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s random pose parameters;

[0030] S33: Repeat step S32 according to the specified number to obtain all random foot posture parameters under the specified composite posture class;

[0031] S34: All random foot posture parameters under the composite posture class are stored uniformly as a random foot posture parameter dataset.

[0032] Furthermore, in step S4, the specific operation steps for replacing the values ​​of corresponding positions in the posture parameters obtained by fitting the existing human body shape and posture data using the SUPR model with the random foot posture parameters are as follows:

[0033] S41: The shape parameters of the SUPR model Posture parameters Rotation parameters and displacement parameters Set as optimization parameters;

[0034] S42: Using the existing fitting algorithm based on the LBFGS optimization algorithm, the SUPR model is used to fit the existing human body shape and posture data based on the SMPL-X model, and the optimization parameters in step S41 are iteratively optimized to obtain the rotation parameters with the best fitting effect. Displacement parameters Shape parameters and posture parameters

[0035] S43: Randomly extract the parameters of the random foot posture parameter data set obtained in step S3 According to the joint, the step S42 obtained Replace the values ​​of the corresponding positions in to obtain the posture parameters with real body posture and random foot posture

[0036] S44: and As the input of the SUPR model, a SUPR model with true shape, body pose and random foot pose can be obtained.

[0037] Furthermore, in step S5, the specific steps for building a 3D scene and obtaining a complete rendering pipeline are as follows:

[0038] S51: Use Blender software to load and place the SUPR model at the origin of the 3D scene;

[0039] S52: Use the skin map from the existing SMPL-X model-based skin map dataset to apply skin maps to the SUPR model;

[0040] S53: On a sphere with a specified radius and centered on the midfoot joint, obtain a specified number of evenly distributed world coordinates as camera positions; set the camera's internal and external parameters so that the camera can fully capture the SUPR model's foot at these camera positions evenly distributed on the sphere with the specified radius, and align the field of view with the midfoot joint.

[0041] S54: Set the scene lighting mode to point light source, the position is the same as the camera, and it faces the mid-foot joint;

[0042] S55: Enable the node function of the Blender 3D scene, set the transparency and depth, use the images in the existing scene classification image dataset as node input, set them as the background images for camera rendering, and obtain a complete rendering pipeline.

[0043] Furthermore, in step S6, the specific operation of generating a dataset instance using the rendering pipeline is as follows: using Blender's own rendering engine, the camera view in the 3D scene built in step S5 is rendered into an RGB image, and then the world_to_camera_view function provided by Blender is used to calculate J ankle 、J mid and [J toe1 ,J toe2 ,...,J toe10 ] in the RGB image; the minimum value on the x-axis and the maximum value on the y-axis of the 2D pixel coordinates of the foot joint are used as the pixel coordinates of the upper left corner of the bounding box; the maximum value on the x-axis and the minimum value on the y-axis of the 2D pixel coordinates of the foot joint are used as the pixel coordinates of the lower right corner of the bounding box; the 2D pixel coordinates of the foot joint and the 2D pixel coordinates of the two vertices of the bounding box constitute the label information of the RGB image; the RGB image and its label information constitute an instance of a virtual dataset for human foot posture detection.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] 1. This invention fills the gap in datasets with complex human foot posture annotations.

[0046] 2. The present invention does not require expensive and complex hardware equipment and special sites, is easy to use and has low cost.

[0047] 3. The annotation of the virtual dataset generated by the present invention is very accurate and can overcome the occlusion problem.

[0048] 4. The dataset generated by the present invention can be used to train a detection network for complex human foot postures and achieve good results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a logical flow diagram of the present invention.

[0050] Figure 2 This is a schematic diagram of an example of a rendering scene constructed for the present invention.

[0051] Figure 3 Schematic diagram of the rendering pipeline constructed for the present invention.

[0052] Figure 4 Schematic diagram of data samples of the virtual dataset for human foot posture detection generated by the present invention. DETAILED DESCRIPTION

[0053] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0054] like Figure 1 As shown, this embodiment provides a method for constructing a virtual dataset for human foot posture detection based on a SUPR model, comprising the following steps:

[0055] 1) Set the rotation range of the SUPR model's foot joints to determine the overall foot posture space;

[0056] SUPR model foot joints including ankle joints J ankle , midfoot jointJ mid and toe joints[J toe1 ,J toe2 ,J toe3 ,J toe4 ,J toe5 ,J toe6 ,J toe7 ,J toe8 ,J toe9 ,J toe10 ], where J toe1 It is the base joint of the big toe, J toe2 The middle joint of the big toe, J toe3 The base joint of the second toe, J toe4 The middle joint of the second toe, J toe5 The base joint of the third toe, J toe6 The middle joint of the third toe, J toe7 The base joint of the fourth toe, J toe8 The middle joint of the fourth toe, J toe9 It is the base joint of the little toe, J toe10The toe joint is the medial joint of the little toe, with a total of 12 joints. The toe joints can be divided into the toe base joints with odd subscripts according to the parity of the subscripts. and the mid-phalangeal joints with even subscripts The pose parameter of each joint is a vector representing the Euler rotation angle of the joint. Among them, θ x is the rotation angle about the x-axis, θ y is the rotation angle about the y-axis, θ z Is the rotation angle on the z-axis. The posture parameters of the entire foot That is, the joint posture parameters are concatenated in sequence through the concat operation. in, It is the ankle joint J ankle The posture parameters, The midfoot joint mid The posture parameters, It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ] is the posture parameter. By setting θ x ,θ y and θ z The range of values ​​is used to limit the rotation range of the joint on the x, y, and z axes. The 3D space determined in this way is the joint posture space S. The sum of the posture spaces of all joints is the overall posture space of the foot, where S foot is the overall posture space of the foot, S ankle It is the ankle joint J ankle The posture space, S mid The midfoot joint mid The posture space, S [toe1,toe2,...,toe10] It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s posture space S foot =S ankle +S mid +S [toe1,toe2,...,toe10] During the implementation, the specific rotation ranges of each joint on the x, y, and z axes are shown in Table 1.

[0057] Table 1 Rotation range of joints

[0058]

[0059] 2) Use classification methods to reduce the dimensionality of the overall foot posture space to obtain composite posture classes and the posture space corresponding to each class;

[0060] According to the physiological structure and movement mode of human feet, the initial posture space ST-pose , plantar flexion posture space S Plantarflex , back-bend posture space S Dorsiflex , inversion posture space S Inversion , eversion posture space S Eversion , internal rotation posture space S MedialRotation , external rotation posture space S LateralRotation , toe bending posture space S ToeFlexion , toe curling posture space S ToeExtension , toes open posture space S ToeAbduction and toes gathered posture space S ToeAdduction These 11 mutually independent posture spaces are regarded as the overall foot posture space S set in step S1. foot The principal component basis of S can be generated by linearly mixing the posture parameters under these 11 principal component bases. foot The posture parameters of most of the following are obtained. The posture types represented by these 11 bases are called basic posture classes. The specific posture space divided by each class is shown in Table 2.

[0061] Table 2 Posture space of basic posture class (left foot)

[0062]

[0063]

[0064] By linear mixing, S foot As the overall foot posture space S foot Represented as the toes-inward posture space Toe-tightening and outward-turning posture space Tiptoe inversion posture space Tiptoe eversion posture space Tiptoe inversion posture space Toe-turning posture space Toe inward posture space Toe-turning posture space and the supplementary posture space S completion There are 9 bases representing the posture space, among which:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] The posture type represented by the above 9 bases is called a composite posture class.

[0075] 3) Traverse the composite posture classes and generate a specified number of random foot posture parameters in the posture space of each class using the random assignment method. The parameters generated in all classes are stored together to form a random foot posture parameter dataset.

[0076] Traverse the 9 composite posture classes obtained in step 2) to obtain the posture space corresponding to each composite posture class, that is, the range of the rotation angle of each joint on the x, y, and z axes. The rotation angle on the x axis is obtained by randomly rounding the rotation range of all joints on the x, y, and z axes using the randint function. Rotation angle on the y-axis and the rotation angle on the z axis As the random pose of the joint The rotation angles are then concatenated into random foot pose parameters using the concat operation:

[0077]

[0078] in, is the foot random posture parameter, It is the ankle joint J ankle The random pose parameters of The midfoot joint mid The random pose parameters of It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s random pose parameters.

[0079] Repeat the above steps of random rounding and concatenation according to the specified composite pose class and the specified number to obtain all the random foot pose parameters under the specified composite pose class. The random foot pose parameters under all composite pose classes are stored together as a random foot pose parameter dataset.

[0080] 4) Using the parameters in the random foot posture parameter dataset, replace the corresponding values ​​in the posture parameters obtained by fitting the SUPR model to the existing human body shape and posture data, and obtain a SUPR model with the real shape, body posture and random foot posture;

[0081] The shape parameters of the SUPR model are Posture parameters Rotation parameters and displacement parameters Set as optimization parameters. Use the existing fitting algorithm based on the LBFGS optimization algorithm to fit the existing human body shape and posture data based on the SMPL-X model with the SUPR model, iteratively optimize the above optimization parameters, and obtain the rotation parameters with the best fitting effect. Displacement parameters Shape parameters and posture parameters Randomly extract the parameters of the random foot posture parameter dataset obtained in step 3) According to the joints, Replace the values ​​of the corresponding positions in to obtain the posture parameters with real body posture and random foot posture The corresponding serial numbers of the joints are shown in Table 3.

[0082] Table 3 Corresponding serial numbers of SUPR model joints

[0083]

[0084]

[0085] Will and As the input of the SUPR model, a SUPR model with true shape, body pose and random foot pose can be obtained.

[0086] 5) Using the SUPR model obtained in step S4, build a 3D scene, including model mapping, camera settings, lighting settings, and rendering settings, to obtain a complete rendering pipeline;

[0087] To use the SUPR model in Blender, it's necessary to preprocess the model into a 3D skinned mesh object that can be edited in Blender. This preprocessing involves: first, using the joint positions of the SUPR model as joint heads, generating a 10cm-long skeleton object for each joint. Based on the joint structure of the SUPR model, the skeleton is connected in a tree-like manner to generate an editable skeleton object. Second, according to the SUPR model's skinning calculation method and skinning parameters, the SUPR model's vertices are bound to the skeleton object generated in the previous step to generate an editable 3D skinned mesh object. Finally, based on the UV map of the existing SMPL-X model skin map dataset, a material node is added to the 3D skinned mesh object to facilitate the setting of textures for the 3D skinned mesh object.

[0088] Load and place the 3D skinned mesh object to the world coordinate origin of the Blender 3D scene, and use the skin map of the existing SMPL-X model's skin map data set to apply a skin map to the 3D skinned mesh object. On a sphere with a specified radius and centered on the mid-foot joint of the 3D skinned mesh object, obtain a specified number of evenly distributed world coordinates as the camera position. Set the camera's internal and external parameters so that the camera can fully capture the SUPR model's foot at these camera positions evenly distributed on the sphere with a specified radius, and align the field of view center with the mid-foot joint. Set the scene lighting mode to point light source, with the same position as the camera, facing the mid-foot joint. The entire 3D scene is built like this: Figure 2 shown.

[0089] Enable the node functionality of the Blender 3D scene. Create a CompositorNodeImage node to receive images from the existing scene classification image dataset. Create a CompositorNodeScale node to receive the images and scale the resolution to the same as the camera's resolution. Create a CompositorNodeRLayerd node to receive the camera's rendering results.

[0090] Create a CompositorNodeAlphaOver node to receive the cropped image and camera rendering results.

[0091] Create a CompositorNodeComposite node, receive the output of the CompositorNodeAlphaOver node, and use the image in the scene classification image dataset as the background image of the camera rendering result by setting the transparency and depth. The entire rendering pipeline is as follows Figure 3 shown.

[0092] 6) Generate an RGB image of a virtual human foot using the obtained rendering pipeline. Calculate the 2D pixel coordinates of the foot joints and the pixel coordinates of the bounding box as the label information of the RGB image. An RGB image and its label information constitute an instance of a virtual dataset for human foot posture detection.

[0093] Use Blender's built-in rendering engine, the 3D scene and rendering pipeline built in step 5) to generate the RGB image of the dataset. Then use the world_to_camera_view function provided by Blender to calculate J ankle 、J mid and [J toe1 ,J toe2 ,...,J toe10 ] 2D pixel coordinates in the RGB image. The minimum value on the x-axis and the maximum value on the y-axis of the 2D pixel coordinates of the foot joint are used as the pixel coordinates of the upper left corner of the bounding box. The maximum value on the x-axis and the minimum value on the y-axis of the 2D pixel coordinates of the foot joint are used as the pixel coordinates of the lower right corner of the bounding box. The 2D pixel coordinates of the foot joint and the 2D pixel coordinates of the two vertices of the bounding box constitute the label information of the RGB image. The RGB image and its label information constitute an instance of a virtual dataset for human foot posture detection, such as Figure 4 shown.

[0094] 7) By repeating steps 4) to 6), a complete virtual dataset for human foot posture detection can be generated.

[0095] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for constructing a virtual dataset for human foot posture detection based on the SUPR model, characterized in that: The following steps are involved: S1: Set the rotation range of the foot joints of the SUPR model to determine the overall posture space of the foot; S2: Use classification methods to reduce the dimension of the overall foot posture space to obtain composite posture classes and the posture space corresponding to each class; S3: Traverse the composite posture classes and generate a specified number of random foot posture parameters in the posture space of each class using the random assignment method. The parameters generated in all classes are stored together to form a random foot posture parameter dataset. S4: Use the parameters in the random foot posture parameter dataset to replace the values ​​of the corresponding positions in the posture parameters obtained by fitting the existing human body shape and posture data with the SUPR model, and obtain a SUPR model with the real shape, body posture and random foot posture; S5: Use the SUPR model obtained in step S4 to build a 3D scene, including model mapping, camera settings, lighting settings, and rendering settings, to obtain a complete rendering pipeline; S6: Generate an RGB image of a virtual human foot using the obtained rendering pipeline. Calculate the 2D pixel coordinates of the foot joints and the pixel coordinates of the upper left and lower right corners of the bounding box as label information for the RGB image. An RGB image and its label information constitute an instance of a virtual dataset for human foot posture detection. S7: Repeat steps S4 to S6 to generate all instances, and the required virtual dataset for human foot posture detection based on the SUPR model can be obtained, that is, a complete virtual dataset for human foot posture detection.

2. The method for constructing a virtual dataset for human foot posture detection based on the SUPR model according to claim 1, characterized in that: In step S1, the foot joints of the SUPR model include the ankle joint J ankle , midfoot jointJ mid and toe joints[J toe1 ,J toe2 ,J toe3 ,J toe4 ,J toe5 ,J toe6 ,J toe7 ,J toe8 ,J toe9 ,J toe10 ], where J toe1 It is the base joint of the big toe, J toe2 The middle joint of the big toe, J toe3 The base joint of the second toe, J toe4 The middle joint of the second toe, J toe5 The base joint of the third toe, J toe6 The middle joint of the third toe, J toe7 The base joint of the fourth toe, J toe8 The middle joint of the fourth toe, J toe9 It is the base joint of the little toe, J toe10 It is the middle joint of the little toe, a total of 12 joints; the posture parameter of each joint is a vector used to represent the Euler rotation angle of the joint where θ x is the rotation angle about the x-axis, θ y is the rotation angle about the y-axis, θ z is the rotation angle on the z-axis; the posture parameter of the whole foot That is, the joint posture parameters are concatenated in sequence through the concat operation. in, It is the ankle joint J ankle The posture parameters, The midfoot joint mid The posture parameters, It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s posture parameters; by setting θ x ,θ y and θ z The value range of is used to limit the rotation range of the joint on the x, y, and z axes. The 3D space determined in this way is the posture space S of the joint; the sum of the posture spaces of all joints is the overall posture space S of the foot. foot =S ankle +S mid +S [toe1,toe2,...,toe10] , where S foot is the overall posture space of the foot, S ankle It is the ankle joint J ankle The posture space, S mid The midfoot joint mid The posture space, S [toe1,toe2,...,toe10] It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s posture space.

3. The method for constructing a virtual dataset for human foot posture detection based on the SUPR model according to claim 2, characterized in that: In step S2, the specific operations of reducing the dimension of the overall foot posture space by classification are as follows: S21: According to the physiological structure and movement mode of human feet, the initial posture space S T-pose , plantar flexion posture space S Plantarflex , back-bend posture space S Dorsiflex , inversion posture space S Inversion , eversion posture space S Eversion , internal rotation posture space S MedialRotation , external rotation posture space S LateralRotation , toe bending posture space S ToeFlexion , toe curling posture space S ToeExtension , toes open posture space S ToeAbduction and toes gathered posture space S ToeAdduction These 11 mutually independent posture spaces are regarded as the overall foot posture space S set in step S1. foot The principal component basis of S is generated by linearly mixing the posture parameters under these 11 principal component bases. foot Most of the posture parameters under , the posture type represented by these 11 principal component bases is called the basic posture class; S22: The posture spaces of the 11 basic posture classes divided in step S21 are further transformed into the overall posture space of the foot S by linear blending. foot Represented as the toes-inward posture space Toe-tightening and outward-turning posture space Tiptoe inversion posture space Tiptoe eversion posture space Tiptoe inversion posture space Toe-turning posture space Toe inward posture space Toe-turning posture space and the supplementary posture space S completion There are 9 bases representing the posture space, among which: The posture type represented by the above 9 bases is called a composite posture class.

4. The method for constructing a virtual dataset for human foot posture detection based on the SUPR model according to claim 3, characterized in that: In step S3, the specific steps of generating a random foot posture parameter dataset are as follows: S31: Traverse the composite posture classes divided in step S2 to obtain the posture space corresponding to each composite posture class, that is, the value range of the rotation angle of each joint on the x, y, and z axes; S32: Assign the rotation angle value in the form of random rounding within the range of the rotation angle; concatenate the rotation angle into the random foot posture parameter through the concat operation in, is the foot random posture parameter, It is the ankle joint J ankle The random pose parameters of The midfoot joint mid The random pose parameters of It is the toe joint [J toe1 ,J toe2 ,...,J toe10 ]’s random pose parameters; S33: Repeat step S32 according to the specified number to obtain all random foot posture parameters under the specified composite posture class; S34: All random foot posture parameters under the composite posture class are stored uniformly as a random foot posture parameter dataset.

5. The method for constructing a virtual dataset for human foot posture detection based on the SUPR model according to claim 4, characterized in that: In step S4, the specific operation steps of replacing the values ​​of corresponding positions in the posture parameters obtained by fitting the existing human body shape and posture data using the SUPR model with the random foot posture parameters are as follows: S41: The shape parameters of the SUPR model Posture parameters Rotation parameters and displacement parameters Set as optimization parameters; S42: Using the existing fitting algorithm based on the LBFGS optimization algorithm, the SUPR model is used to fit the existing human body shape and posture data based on the SMPL-X model, and the optimization parameters in step S41 are iteratively optimized to obtain the rotation parameters with the best fitting effect. Displacement parameters Shape parameters and posture parameters S43: Randomly extract the parameters of the random foot posture parameter data set obtained in step S3 According to the joint, the step S42 obtained Replace the values ​​of the corresponding positions in to obtain the posture parameters with real body posture and random foot posture S44: and As the input of the SUPR model, a SUPR model with true shape, body pose and random foot pose can be obtained.

6. The method for constructing a virtual dataset for human foot posture detection based on the SUPR model according to claim 5, characterized in that: In step S5, the specific steps for building a 3D scene and obtaining a complete rendering pipeline are as follows: S51: Use Blender software to load and place the SUPR model at the origin of the 3D scene; S52: Use the skin map from the existing SMPL-X model-based skin map dataset to apply skin maps to the SUPR model; S53: On a sphere with a specified radius and centered on the midfoot joint, obtain a specified number of evenly distributed world coordinates as camera positions; set the camera's internal and external parameters so that the camera can fully capture the SUPR model's foot at these camera positions evenly distributed on the sphere with the specified radius, and align the field of view with the midfoot joint. S54: Set the scene lighting mode to point light source, the position is the same as the camera, and it faces the mid-foot joint; S55: Enable the node function of the Blender 3D scene, set the transparency and depth, use the images in the existing scene classification image dataset as node input, set them as the background images for camera rendering, and obtain a complete rendering pipeline.

7. The method for constructing a virtual dataset for human foot posture detection based on a SUPR model according to claim 6, characterized in that: In step S6, the specific operation of generating a dataset instance using the rendering pipeline is as follows: using Blender's own rendering engine, the camera view in the 3D scene built in step S5 is rendered into an RGB image, and then the J is calculated using the world_to_camera_view function provided by Blender. ankle 、J mid and [J toe1 ,J toe2 ,...,J toe10 ] 2D pixel coordinates in the RGB image; the minimum value on the x-axis and the maximum value on the y-axis of the 2D pixel coordinates of the foot joint are used as the pixel coordinates of the upper left corner of the bounding box; the maximum value on the x-axis and the minimum value on the y-axis of the 2D pixel coordinates of the foot joint are used as the pixel coordinates of the lower right corner of the bounding box; the 2D pixel coordinates of the foot joint and the 2D pixel coordinates of the two vertices of the bounding box constitute the label information of the RGB image; The RGB image and its label information constitute an instance of a virtual dataset for human foot pose detection.

Images