A feature point calibration method for an ICP algorithm-based no-identity human upper body point cloud model

Abstract

The application discloses a feature point calibration method of an unmarked human upper body point cloud model based on an ICP algorithm and belongs to the technical field of human parameter measurement. In a human template model library in which feature points have been calibrated, parameters of an uncalibrated human point cloud target model are taken as input to determine a template model, the upper body of the template model is taken as a source point cloud, the upper body of the target model is taken as a target point cloud, non-rigid ICP registration is performed, local source point cloud sets and local target point cloud sets are determined according to a registration result, rigid ICP registration is further performed, and a feature point set of the target model is determined through nearest point searching according to a registration result. The application solves the feature point calibration problem of a human point cloud model acquired by a general depth camera without setting an identification point, the feature point can be located at any part of the human body, and no feature information such as curvature, edge and girth is needed, thereby providing technical support for online clothing customization and individual customization in the field of human engineering.

Description

Technical Field

[0001] The invention belongs to the field of human body parameter measurement, in particular to a feature point calibration method of an unmarked human upper body point cloud model based on an ICP algorithm. Background Art

[0002] The integration of the apparel industry with the internet and the Internet of Things has enabled online sales to address the fit of clothing purchased by customers through new models such as virtual fitting and online customization. A key technology for this is the measurement of human body parameters and the determination of human projection points (feature points) for clothing. Under the current mature online sales model, if the location data of human body parameters and feature points for clothing can be provided to customers at their convenience, even if the data is partial or less accurate, it can help customers choose clothing that fits them better. Customized clothing services are a development trend in the clothing industry. Future customized services that combine online and offline customization or fully online customization services will require relatively accurate human body measurement data and feature point location data.

[0003] The acquisition of human body parameters and feature point locations is inseparable from effective anthropometry technology. Traditional contact measurement technology is suitable for offline customization models, but not for online customization and online and offline combined customization methods. With the development of measurement technology, image processing technology, 3D scanning technology, and computer graphics technology, non-contact measurement technology will be the development trend of future clothing customization measurement technology. In particular, for sales models that require remote measurement operations such as auxiliary selection of clothing models, online and offline combined clothing customization, and completely online clothing customization, non-contact measurement will be an irreplaceable and effective method. Non-contact measurement technology includes two-dimensional measurement technology and three-dimensional measurement technology. Three-dimensional measurement technology has high measurement accuracy and has become a hot spot in the current research and application of anthropometry.

[0004] To improve the fit of custom clothing, in addition to measuring basic body parameters such as length, width, and girth, it is also necessary to obtain the following body shape measurements and positional information: the slope of specific areas, the locations of key feature points, and geodesics between arbitrary feature points. The determination of body feature points and the acquisition of body parameters complement each other. Methods for determining body feature points include two main categories: marker-based and non-marker-based methods. Marker-based methods require the subject to wear specific clothing or place markers at the subject's feature points to accurately locate them. Non-marker-based methods do not require markers to be placed on the subject. Instead, specialized algorithms are designed to determine the locations of the required feature points based on scanned or photographed body shape data (such as point cloud data). While marker-based methods can obtain relatively accurate feature point location information, they require the placement of markers on the subject's body equal to the number of feature points during measurement. This typically requires a professional technician to configure the markers, and the process can cause discomfort to the subject. Non-marker-based methods are simple to use and do not impose any additional burden on the subject, but their measurement accuracy is relatively low.

[0005] Markerless methods are clearly more suitable for measurement technology used in online clothing customization. Existing markerless methods can be divided into two categories: one utilizes the proportional relationships between human body parts on the body to determine the location of each feature point on the human body model; the other utilizes surface variation patterns of 3D point cloud models or 2D images, such as curvature, normal vectors, and the convex hull of the projection, to design algorithms to determine the location of each feature point on the human body model. These methods all have certain restrictions on the feature points that can be calibrated, and cannot accurately calibrate feature points at any location on the human body surface. Summary of the Invention

[0006] The present invention aims to address the above-mentioned problems in the existing technology by providing a method for calibrating feature points of an unmarked upper torso point cloud model based on the ICP algorithm. Based on a 3D point cloud model of the human body, the present invention uses the Iterative Closest Point (ICP) algorithm twice to calibrate multiple key feature points for clothing on a 3D point cloud model of the upper torso without markers.

[0007] The technical solution for achieving the purpose of the present invention is: a feature point calibration method for an unmarked upper body point cloud model based on the ICP algorithm, the method comprising: calculating a similarity function value in a template model library with calibrated feature points using easily accessible human body parameters of a target model to be calibrated; determining a matching template model based on the similarity function value, and applying a non-rigid ICP registration algorithm to preliminarily determine a feature point set of the target model to be calibrated; and performing rigid ICP registration again on the local point clouds around each feature point of the template model and the target model to be calibrated to obtain feature point position information of the target model to be calibrated.

[0008] Furthermore, the method comprises the following steps:

[0009] Step 1: In a template model library of 3D human point clouds of different body types with calibrated feature points, the similarity function value is calculated using the readily available human body parameters of the 3D human point cloud model to be calibrated, i.e., the target model, and the template model with the highest similarity is selected.

[0010] Step 2: Select the upper body point clouds of the target model and the template model selected in step 1, and determine the sequence number of each feature point in the point cloud array on the upper body point cloud of the template model;

[0011] Step 3: Using the upper body point cloud of the template model selected in step 2 as the source point cloud and the upper body point cloud of the target model as the target point cloud, perform non-rigid ICP registration to obtain the transformed source point cloud and its feature point sequence number set;

[0012] Step 4: On the source point cloud obtained after non-rigid ICP registration in step 3, the corresponding feature point coordinates are found. These coordinate values ​​are used as input to find the nearest point of each feature point on the target model in step 1 to obtain the feature point set of the target model.

[0013] Step 5: For each feature point on the source point cloud after non-rigid ICP registration, select the surrounding points as the local source point cloud; for each preliminarily calibrated feature point on the target point cloud, select the surrounding points as the local target point cloud;

[0014] In step 6, for each pair of corresponding local source point clouds and local target point clouds obtained in step 5, rigid ICP registration is performed to obtain the coordinate values ​​of each feature point on the transformed local source point cloud. Based on these coordinates, the calibrated feature point set is found on the target point cloud.

[0015] Furthermore, the feature points calibrated in step 1 are key projection points of the upper garment design pattern prototype, including: front neck point, vertebral arch, right shoulder and neck point, left shoulder and neck point, left breast point, right breast point, front right front armpit point, front left front armpit point, right acromion, left acromion, front fossa, center point of right acromion to right front armpit point, center point of left acromion to left front armpit point, back right back armpit point, back left back armpit point, center point of right acromion to right back armpit point, center point of left acromion to left back armpit point, back vest, front waistline point, right waistline point, left waistline point and back waistline point.

[0016] Furthermore, the human body parameters that are easily obtained in step 1 include any combination of the following parameters: gender, age, height, weight, chest circumference, waist circumference, and hip circumference.

[0017] Furthermore, the calculation formula for calculating the similarity function value in step 1 is:

[0018]

[0019] Where S is the similarity value between the template model and the target model, n is the number of human body parameters, α i represents the coefficient, p i is the i-th human body parameter of the target model, p m,i is the i-th human body parameter of template model m.

[0020] Furthermore, in step 2, the topological information of the template model point cloud and the calibrated feature point positions are applied to select the upper body point cloud models of the target model and the template model. The point set and triangle face set of the selected upper body point cloud models are subsets of the point set and triangle face set of the respective original models.

[0021] Furthermore, in step 3, for the source point cloud and the target point cloud, the distance error and the smoothing error are used to form the objective function, and the transformation matrix is ​​calculated for each pair of points to perform non-rigid ICP registration.

[0022] Furthermore, in step 4, the coordinates of each feature point are found by the serial number, and based on the coordinates of the feature points on the source point cloud, the respective nearest points are found on the target point cloud before non-rigid ICP registration to obtain the feature point set for the preliminary calibration of the target model.

[0023] Furthermore, in step 5, on the target point cloud before non-rigid ICP registration, N local point clouds within N circular domains are selected as local target point cloud sets with the N feature points preliminarily calibrated as the circle center and r as the radius; on the source point cloud after non-rigid ICP registration, N local point clouds within N circular domains are selected as local source point cloud sets with the N feature points as the circle center and r as the radius, and at the same time, the coordinate values ​​of the N feature points on each local source point cloud are used to determine their respective serial numbers on the local point cloud; where N is the number of feature points.

[0024] Furthermore, step 6 includes:

[0025] Using distance error as the objective function, each point cloud in the local source point cloud set is used as the source point cloud, and the corresponding point cloud in the local target point cloud set is used as the target point cloud to perform rigid ICP registration.

[0026] On each source point cloud after registration, determine the corresponding feature point coordinate value according to the sequence number of each feature point;

[0027] Based on the coordinate values ​​of the N feature points, the nearest points of each are found on the target point cloud in step 3 to obtain the final set of N feature points on the calibrated target point cloud; where N is the number of feature points.

[0028] Compared with the prior art, the present invention has the following significant advantages:

[0029] (1) Based on a human body 3D point cloud model, the present invention uses the Iterative Closest Point (ICP) algorithm twice to calibrate multiple key feature points for clothing on a human upper body 3D point cloud model without markers. The present invention can solve the problem of determining the human upper body feature points required for a garment pattern prototype based on unmarked human body 3D point cloud data obtained by an ordinary depth camera from an untrained customer in an everyday environment.

[0030] (2) The feature points calibrated by the present invention can be located at any part of the human body and do not need to have geometric features and boundary features. The flexibility of feature point calibration is greatly improved within an acceptable range of accuracy error, providing technical support for personalized customization in human factors engineering related fields such as online clothing customization.

[0031] (3) The three-dimensional point cloud model of the human body processed by the present invention can come from the scanning results of a scanning device, the calculation results of a modeling software, and the data collection results of any other recording equipment. The calibrated feature points can be extended to any part of the body.

[0032] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the feature point calibration method of the unmarked upper body point cloud model based on the ICP algorithm of the present invention.

[0034] Figure 2 Schematic diagram of the complete point cloud and 22 feature point positions of a template model in one embodiment.

[0035] Figure 3 Schematic diagram of the upper body point cloud and 22 feature point numbers of a template model in one embodiment.

[0036] Figure 4 Schematic diagram of a complete point cloud of a target model in one embodiment.

[0037] Figure 5 Schematic diagram of the upper body point cloud of the target model in one embodiment.

[0038] Figure 6 Schematic diagram of the target point cloud and source point cloud (red is the source point cloud, blue is the target point cloud) after non-rigid ICP registration in one embodiment.

[0039] Figure 7 Schematic diagram of a source point cloud after non-rigid ICP registration in one embodiment.

[0040] Figure 8Schematic diagram of the location of the anterior fossa point of the source point cloud and its local point cloud after non-rigid ICP registration in one embodiment.

[0041] Figure 9 Schematic diagram of the location of the anterior fossa point of the target point cloud and its local point cloud in one embodiment.

[0042] Figure 10 Schematic diagram of the source point cloud and target point cloud after the second rigid ICP registration of the front center local area with feature points in one embodiment.

[0043] Figure 11 Schematic diagram of the target model point cloud and calibrated feature points in one embodiment. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0045] It should be noted that if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0046] The ideal scenario for online clothing customization is that ordinary customers can use easy-to-use terminal devices (such as mobile phones equipped with depth cameras, etc.) to obtain multiple three-dimensional point clouds of the human body through a multi-angle shooting method with relatively loose shooting conditions, upload them to the server side of the customization platform, and obtain a complete three-dimensional point cloud model of the clothing parts by applying the server-side splicing and optimization algorithm. The present invention is based on the three-dimensional point cloud model of the human body. By using the iterative closest point (ICP) algorithm twice, it realizes the calibration of 22 key feature points for clothing on the three-dimensional point cloud model of the upper body of the human body under unmarked conditions. Compared with the existing methods, the feature points calibrated by the present invention can be located at any part of the human body, and do not need to have geometric features and boundary features. The flexibility of feature point calibration is greatly improved within an acceptable range of accuracy error. The three-dimensional point cloud model of the human body processed by the present invention can come from the scanning results of the scanning equipment, the calculation results of the modeling software and the data acquisition results of any other recording equipment. The calibrated feature points can be extended to any part of the body.

[0047] In one embodiment, the present invention provides a method for calibrating feature points of an unmarked upper body point cloud model based on an ICP algorithm, specifically comprising: in a template model library where feature points have been calibrated, using easily accessible human body parameters of a target model to be calibrated, calculating a similarity function value, determining a matching template model based on the similarity function value, applying a non-rigid ICP registration algorithm for the first time to preliminarily determine a feature point set of the target model, performing a second rigid ICP registration on the local point clouds around the above-mentioned feature points of the template model and the target model, and obtaining accurate feature point position information of the target model to be calibrated.

[0048] Furthermore, in one embodiment, in combination Figure 1 (The text on the left side of the arrow line in the figure represents the key data of the input or output of each step). The method specifically includes:

[0049] Step 1: In a template model library of 3D human point clouds of different body types with calibrated feature points, the similarity function value is calculated using the readily available human body parameters of the 3D human point cloud model to be calibrated, i.e., the target model, and the template model with the highest similarity is selected.

[0050] Here, the feature points calibrated in step 1 are the 22 key projection points of the upper garment design pattern prototype, including: front neck point, vertebral column, right shoulder and neck point, left shoulder and neck point, left breast point, right breast point, front right front armpit point, front left front armpit point, right acromion, left acromion, front fossa, center point between right acromion and right front armpit, center point between left acromion and left front armpit, back right back armpit point, back left back armpit point, center point between right acromion and right back armpit, center point between left acromion and left back armpit, back waist, front waistline point, right waistline point, left waistline point, and back waistline point. The feature points and their meanings are shown in Table 1.

[0051] Table 1 Feature points and their significance

[0052]

[0053] It should be noted that the feature points that can be determined by the present invention are not limited to the 22 feature points of the upper body, and can be any other feature points of the upper body, or feature points of any other part of the body.

[0054] Step 2: Select the upper body point clouds of the target model and the template model selected in step 1, and determine the sequence number of each feature point in the point cloud array on the upper body point cloud of the template model;

[0055] Here, the point cloud data of the model to be calibrated is obtained without special identification of the feature point positions. The point cloud model in the template model library is a model with 22 feature points calibrated using other methods.

[0056] Step 3: Using the upper body point cloud of the template model selected in step 2 as the source point cloud and the upper body point cloud of the target model as the target point cloud, perform non-rigid ICP registration to obtain the transformed source point cloud and its feature point sequence number set;

[0057] Step 4: On the source point cloud obtained after non-rigid ICP registration in step 3, the corresponding feature point coordinates are found. These coordinate values ​​are used as input to find the nearest point of each feature point on the target model in step 1 to obtain the feature point set of the target model.

[0058] Step 5: For each feature point on the source point cloud after non-rigid ICP registration, select the surrounding points as the local source point cloud; for each preliminarily calibrated feature point on the target point cloud, select the surrounding points as the local target point cloud;

[0059] In step 6, for each pair of corresponding local source point clouds and local target point clouds obtained in step 5, rigid ICP registration is performed to obtain the coordinate values ​​of each feature point on the transformed local source point cloud. Based on these coordinates, the calibrated feature point set is found on the target point cloud.

[0060] Furthermore, in one embodiment, in step 1, when selecting the template model, the human body parameters used include (but are not limited to) all or part of the following parameters, and the basis for selection is that the corresponding parameters of the target model are easy to obtain (same gender): age, height, weight, chest circumference, waist circumference, hip circumference, thigh circumference, calf circumference, shoulder height, leg length, etc.

[0061] Furthermore, in one embodiment, the calculation formula for calculating the similarity function value in step 1 is:

[0062]

[0063] Where S is the similarity value between the template model and the target model, n is the number of human body parameters, α i represents the coefficient, p i is the i-th human body parameter of the target model, p m,i is the i-th human body parameter of the template model m. The smaller the S value, the higher the similarity between the target model to be calibrated and the template model.

[0064] Calculate the S value for each template model in the model library, and select the template model with the smallest S value as the source model for ICP registration.

[0065] Furthermore, in one embodiment, in step 2, the upper body point clouds of the target and template human point cloud models are selected as the target and source point clouds, respectively, for the first non-rigid ICP registration based on the model's topological structure and the coordinate values ​​of the upper body point cloud region. On the upper body point cloud model, the retained upper body point set and corresponding triangular facets are calculated using the triangular facet information of the original point cloud model, ensuring that the point set and triangular facets of the upper body model remain consistent with those of the original models.

[0066] On the complete human body point cloud template model, such as Figure 2 The red asterisks in the middle indicate the positions of the 22 feature points that have been determined. The corresponding points on the upper body template model are found by using the coordinates of the feature points to obtain the feature point set of the upper body template model (such as Figure 3 The numbers shown are the locations of the feature points with corresponding numbers in Table 1. Figure 3 It is taken from Figure 2 Upper body point cloud of the point cloud model): The subscript i is the number of the feature point, i∈{1,2,3,…22}, and s indicates that these feature points belong to the source point cloud during ICP registration.

[0067] Furthermore, in one embodiment, in step 3, for the source point cloud and the target point cloud, the distance error and the smoothing error are used to form an objective function, and a transformation matrix is ​​calculated for each pair of points to perform non-rigid ICP registration. Specifically:

[0068] The surface composed of the source point cloud of the ICP registration algorithm using the template model is The surface composed of the target point cloud of the ICP registration algorithm with the target model is For each point in the source point cloud Find a 4×4 affine transformation matrix T i , so that the points in the source point cloud S are gradually matched to the corresponding points in the target point cloud D.

[0069] In order to measure the matching degree between the source point cloud and the target point cloud after each transformation, two types of error functions are defined, which represent the distance error of corresponding points between the source point cloud and the target point cloud and the smoothing error between adjacent points of the source point cloud.

[0070] For the distance error, the cumulative sum of the Euclidean distances between each pair of corresponding points between the source point cloud and the target point cloud is used, as shown in the following formula:

[0071]

[0072] Where n is the number of points in the source point cloud.

[0073] The smoothing error is the difference between the transformation matrices of adjacent points in the source point cloud, which is used to ensure that the transformation matrices of adjacent points are close. The definition of the smoothing error is as follows:

[0074]

[0075] Among them, S.ε represents the set of edges of the source point cloud, |||| F represents the F-norm, v i ,v j Represent the adjacent points i and j in the source point cloud, T i 、T j Represents the 4×4 affine transformation matrix corresponding to adjacent points i and j respectively.

[0076] Combining the above two types of errors, the objective function is defined as follows:

[0077] E∶=αE d +βE s

[0078] Among them, α and β are weights, indicating the importance of the two types of errors. In the application, they can be adaptively adjusted according to the characteristics of the point cloud and the effect of the registration.

[0079] Furthermore, in one embodiment, in step 4, the coordinates of each feature point are obtained by looking up the serial number, and based on the coordinates of the feature points on the source point cloud, the nearest points of each feature point are found on the target point cloud before non-rigid ICP registration to obtain a feature point set for the target model (here, on the source point cloud after non-rigid ICP registration, the coordinates of the 22 feature points have changed, but the serial numbers of the feature points in the point cloud array have not changed, and the coordinates of each feature point are found by looking up the serial number). In step 5, on the target point cloud before non-rigid ICP registration, N local point clouds within N circular domains are selected as local target point cloud sets with the N initially calibrated feature points as the center and r as the radius; on the source point cloud after non-rigid ICP registration, N local point clouds within N circular domains are selected as local source point cloud sets with the N feature points as the center and r as the radius, and the serial numbers of the respective feature points on each local source point cloud are determined by the coordinate values ​​of the N feature points; where N is the number of feature points.

[0080] Specifically:

[0081] On the source point cloud after non-rigid ICP registration, 22 local point clouds are selected with the 22 feature points as the center and r as the radius, and they are used as the source point cloud set; on the target model point cloud of the upper body, the coordinate values ​​of the 22 feature points are used as input to find the 22 nearest points, which are defined as the 22 approximate feature points on the target model. With these approximate feature points as the center and r as the radius, 22 local point clouds are selected on the target model and they are used as the target point cloud set.

[0082] Suppose {S i} is the local source point cloud, {D i Where} is the local target point cloud set. The rigid ICP registration algorithm uses the same rotation and translation matrix for all points in the source point cloud during a registration process. The distance error function is used to measure the degree of match between the source and target point clouds after each transformation, as shown in the following equation.

[0083]

[0084] Among them, r is a 3×3 rotation matrix, t is a 3×1 translation matrix, and (rt) constitutes a 3×4 affine transformation matrix X.

[0085] Furthermore, in one embodiment, step 6 includes:

[0086] Using distance error as the objective function, each point cloud in the local source point cloud set is used as the source point cloud, and the corresponding point cloud in the local target point cloud set is used as the target point cloud. Rigid ICP registration is performed on 22 pairs of local point clouds around the 22 feature points.

[0087] On each source point cloud after registration, determine the corresponding feature point coordinate value according to the sequence number of each feature point;

[0088] Based on the coordinate values ​​of the N feature points, the nearest points of each are found on the target point cloud in step 3 to obtain the final set of N feature points on the calibrated target point cloud; where N is the number of feature points.

[0089] Here, each point cloud in the local source point cloud set is already close to the corresponding point cloud in the target point cloud set in shape and position. These point cloud pairs are the initial state of the second rigid ICP registration. After the second rigid ICP registration of each pair of point clouds, the source local point cloud will be closer to the corresponding target point cloud in shape and position, and the sequence number of the feature points in the point cloud array has not changed. Using the sequence number of the feature point, the coordinates of the feature point are obtained in the registered source local point cloud (the feature point of the second preliminary calibration), and the nearest point is found on the corresponding local target point cloud as the final calibrated feature point. The coordinate value of the feature point is used to search in the upper body point cloud or the whole body point cloud of the target model to determine the sequence number of the feature point in the upper body point cloud or the whole body point cloud, thus obtaining the feature point set on the target point cloud.

[0090] In one embodiment, a feature point calibration system for an unmarked upper body point cloud model based on an ICP algorithm is provided, the system comprising:

[0091] The first module is used to calculate the similarity function value from the template model library of 3D human point clouds of different body shapes with calibrated feature points, using the easily available human body parameters of the 3D human point cloud model to be calibrated, i.e., the target model, and select the template model with the highest similarity;

[0092] The second module is used to select the upper body point cloud of the target model and the template model selected by the first module respectively, and determine the sequence number of each feature point in the point cloud array on the upper body point cloud of the template model;

[0093] The third module is used to perform non-rigid ICP registration on the upper body point cloud of the template model selected in the second module as the source point cloud and the upper body point cloud of the target model as the target point cloud, to obtain the transformed source point cloud and its feature point sequence number set;

[0094] The fourth module is used to find the coordinates of the corresponding feature points on the source point cloud after non-rigid ICP registration obtained in the third module, and use these coordinate values ​​as input to find the nearest point of each feature point on the target model of the first module to obtain the feature point set of the target model;

[0095] The fifth module is used to select the surrounding points as the local source point cloud for each feature point on the source point cloud after non-rigid ICP registration, and to select the surrounding points as the local target point cloud for each preliminarily calibrated feature point on the target point cloud;

[0096] The sixth module is used to perform rigid ICP registration on each pair of corresponding local source point clouds and local target point clouds obtained in the fifth module, obtain the coordinate values ​​of each feature point on the transformed local source point cloud, and use these coordinates as the basis to find the calibrated feature point set on the target point cloud.

[0097] Regarding the specific limitations of the feature point calibration system for the unmarked upper body point cloud model based on the ICP algorithm, please refer to the limitations of the feature point calibration method for the unmarked upper body point cloud model based on the ICP algorithm above, which will not be repeated here. Each module in the above-mentioned feature point calibration system for the unmarked upper body point cloud model based on the ICP algorithm can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0098] In one embodiment, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following is achieved:

[0099] Step 1: In a template model library of 3D human point clouds of different body types with calibrated feature points, the similarity function value is calculated using the readily available human body parameters of the 3D human point cloud model to be calibrated, i.e., the target model, and the template model with the highest similarity is selected.

[0100] Step 2: Select the upper body point clouds of the target model and the template model selected in step 1, and determine the sequence number of each feature point in the point cloud array on the upper body point cloud of the template model;

[0101] Step 3: Using the upper body point cloud of the template model selected in step 2 as the source point cloud and the upper body point cloud of the target model as the target point cloud, perform non-rigid ICP registration to obtain the transformed source point cloud and its feature point sequence number set;

[0102] Step 4: On the source point cloud obtained after non-rigid ICP registration in step 3, the corresponding feature point coordinates are found. These coordinate values ​​are used as input to find the nearest point of each feature point on the target model in step 1 to obtain the feature point set of the target model.

[0103] Step 5: For each feature point on the source point cloud after non-rigid ICP registration, select the surrounding points as the local source point cloud; for each preliminarily calibrated feature point on the target point cloud, select the surrounding points as the local target point cloud;

[0104] In step 6, for each pair of corresponding local source point clouds and local target point clouds obtained in step 5, rigid ICP registration is performed to obtain the coordinate values ​​of each feature point on the transformed local source point cloud. Based on these coordinates, the calibrated feature point set is found on the target point cloud.

[0105] For the specific limitations of each step, please refer to the limitations of the feature point calibration method for the unmarked upper body point cloud model based on the ICP algorithm above, which will not be repeated here.

[0106] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the computer program implements:

[0107] Step 1: In a template model library of 3D human point clouds of different body types with calibrated feature points, the similarity function value is calculated using the readily available human body parameters of the 3D human point cloud model to be calibrated, i.e., the target model, and the template model with the highest similarity is selected.

[0108] Step 2: Select the upper body point clouds of the target model and the template model selected in step 1, and determine the sequence number of each feature point in the point cloud array on the upper body point cloud of the template model;

[0109] Step 3: Using the upper body point cloud of the template model selected in step 2 as the source point cloud and the upper body point cloud of the target model as the target point cloud, perform non-rigid ICP registration to obtain the transformed source point cloud and its feature point sequence number set;

[0110] Step 4: On the source point cloud obtained after non-rigid ICP registration in step 3, the corresponding feature point coordinates are found. These coordinate values ​​are used as input to find the nearest point of each feature point on the target model in step 1 to obtain the feature point set of the target model.

[0111] Step 5: For each feature point on the source point cloud after non-rigid ICP registration, select the surrounding points as the local source point cloud; for each preliminarily calibrated feature point on the target point cloud, select the surrounding points as the local target point cloud;

[0112] In step 6, for each pair of corresponding local source point clouds and local target point clouds obtained in step 5, rigid ICP registration is performed to obtain the coordinate values ​​of each feature point on the transformed local source point cloud. Based on these coordinates, the calibrated feature point set is found on the target point cloud.

[0113] For the specific limitations of each step, please refer to the limitations of the feature point calibration method for the unmarked upper body point cloud model based on the ICP algorithm above, which will not be repeated here.

[0114] As a specific example, in one of the embodiments, the present invention is further verified and explained in detail.

[0115] The feature point calibration method of the unmarked upper body point cloud model based on the ICP algorithm specifically includes:

[0116] Step 1: Use the female 3D human body point cloud model to calibrate the feature points. The template model library contains 55 female point cloud models of different ages and body shapes. 22 feature points of each model have been determined, such as Figure 2 As shown. This embodiment uses the three-dimensional point cloud of a 23-year-old female as the target model. The human body parameters used when selecting a similar template model are: height (P1), age (P2), weight (P3), chest circumference (P4), waist circumference (P5), and hip circumference (P6), and their values ​​are: 168cm, 23 years old, 66kg, 96cm, 76cm, and 104cm. It should be noted that the female point cloud model selected in the embodiment is arbitrary, and the parameters used in selecting the template model are not necessarily limited to the above 6 types of human body parameters, and can be any number of human body parameters that are relatively easy to obtain or have already been obtained. Substitute the above six parameters into the following formula (1), where the coefficient α i In the embodiment, the value of is based on age and is assigned a value of 1. The other coefficients α iThe value of is set to the quotient of the average value of the parameter divided by the age value (rounded and can be adjusted moderately through experiments).

[0117]

[0118] After expansion, we get the equation shown in formula (2), P m,i is the corresponding parameter of the template model, and the value of m is the serial number of the template model.

[0119]

[0120] In the template model library, the human body parameters of the template model with the smallest S value are retrieved and calculated as follows: height 170 cm, age 22 years old, weight 66 kg, chest circumference 94 cm, waist circumference 76 cm, and hip circumference 103 cm.

[0121] Step 2: Select the upper body point cloud model of the template model and the target model according to the area of ​​the top pattern prototype. Figure 3 The upper body point cloud and 22 feature point numbers of the template model, the feature point set L s ={(1) front neck point, (2) vertebral prominence, (3) right shoulder and neck point, (4) left shoulder and neck point, (5) left breast point, (6) right breast point, (7) front right front axillary point, (8) front left front axillary point, (9) right acromion, (10) left acromion, (11) front fossa, (12) center point between right acromion and right front axillary point, (13) center point between left acromion and left front axillary point, (14) back right back axillary point, (15) back left back axillary point, (16) center point between right acromion and right back axillary point, (17) center point between left acromion and left back axillary point, (18) back vest, (19) front waistline point, (20) right waistline point, (21) left waistline point, (22) back waistline point}. Figure 4 is the complete point cloud of the target model, and its point density is Figure 2 The template model shown is obviously different. Figure 5 is the upper body point cloud of the target model, Figure 5 It is taken from Figure 4 The upper body part of the point cloud model.

[0122] Step 3, Figure 5 The target model shown is the target point cloud. Figure 3 The template model shown is the source point cloud, and the E defined in formula (3) is used as the objective function, where the coefficients α and β are both 1, and the non-rigid ICP algorithm is executed.

[0123] E∶=αE d +βE s (3)

[0124] in,

[0125]

[0126]

[0127] In formula (4) and formula (5), Dist represents the Euclidean distance function between two points. is the surface formed by the source point cloud, For the surface of the target point cloud, each point in the source point cloud T i is a 4×4 affine transformation matrix, S.ε represents the set of edges of the source point cloud, |||| F represents the F-norm.

[0128] Figure 6 This is the result of non-rigid ICP registration of the target point cloud and the source point cloud together, where the red one is the source point cloud after registration and the blue one is the target point cloud after registration. Figure 7 Display the deformed source point cloud separately for comparison Figure 7 and Figure 3 The number of points, sequence numbers and triangle definitions of the two point clouds are exactly the same, but the positions of the points have changed and the shapes are closer. Figure 5 The target point cloud is shown.

[0129] Step 4: Determine the local target point cloud set and local source point cloud set required for the second ICP registration on the target point cloud and the deformed source point cloud. Figure 7 On the source point cloud shown, 22 feature points (the number of the points is the same as Figure 3 ) as the center and r (preferably 15 mm in this embodiment) as the radius, 22 local point clouds are selected as the source point cloud set. Figure 8 is the location of the center point of the source point cloud and the local point cloud selected with it as the center. Figure 5 On the target model point cloud shown in Figure 7 The coordinate values ​​of the 22 feature points are input, and the 22 points closest to them are found and defined as the 22 approximate feature points on the target model. With the 22 approximate feature points as the center and r (15mm in this embodiment) as the radius, 22 local point clouds are selected on the target model and used as the target point cloud set, as shown in the following example: Figure 9 The position of the front heart point of the target point cloud and the local point cloud selected with it as the center.

[0130] Step 5: Perform a second rigid ICP registration on the corresponding point clouds in the local point cloud set. Figure 7 The local source point cloud set selected with 22 feature points as the center on the source point cloud model shown is {S i},exist Figure 5The local target cloud cluster selected with 22 approximate feature points as the center on the target point cloud model shown is {D i}. i Each local point cloud in {D i The corresponding local point cloud in} is the target point cloud, and the rigid ICP registration is performed with formula (6) as the objective function. A total of 22 rigid ICP registrations are performed to obtain 22 source point clouds S after transformation by formula (7). i '. Figure 10 This is the result of the registration of the local point cloud of the anterior fossa. The red one is the source point cloud after the local area registration of the anterior heart, and the blue one is the target point cloud. The red * with a green circle in the middle is the position of the reference anterior fossa point of the original template on the source point cloud after two ICP registrations. The blue * with a purple-red circle is the position of the preliminarily calibrated anterior fossa point on the target point cloud.

[0131]

[0132] S i ′=rS i +t (7)

[0133] Among them, r is a 3×3 rotation matrix, t is a 3×1 translation matrix, Dist represents the Euclidean distance function between two points, D is a local target point cloud, v i for each point in the local source point cloud.

[0134] Step 6: Use the coordinates of the feature points on the 22 local source point clouds after the second rigid ICP registration to obtain the set of 22 feature points on the target point cloud. During the first non-rigid ICP registration process, the upper body point cloud of the human template model is the source model. After the registration is completed, its shape will be fitted to the upper body point cloud of the approximate human target model. Therefore, the position of the point cloud changes, but the sequence number of the point does not change, and therefore the sequence number of the feature point does not change. The coordinates of the corresponding feature point can be determined using the sequence number of the feature point. After the second rigid ICP registration of the local point cloud set, the source local point cloud will be closer in shape and position to the corresponding target point cloud, and the sequence number of the feature point will not change. Similarly, the feature point serial number can be used to obtain the coordinates of the feature point in the source local point cloud after registration, and the nearest point can be found on the corresponding local target point cloud as the final calibrated feature point. The coordinate value of the feature point can be used to search in the upper body point cloud or the whole body point cloud of the target model, so that the position of the feature point in the upper body point cloud or the whole body point cloud can be determined, thus completing the calibration of the feature point of the target point cloud.

[0135] The calibration results of 22 feature points of the upper body of the unmarked target point cloud are as follows: Figure 11As shown. In this embodiment, in order to measure the accuracy of the calibration results, the feature points of the target point cloud have also been calibrated by other methods, which are called benchmark feature points. The distance errors between the feature points calibrated by the present invention and the corresponding benchmark feature points are shown in Table 2, of which 10 points are completely consistent with the benchmark feature points, and the average distance error of 22 points is 0.81 cm. The same method was used to calibrate 36 target models, with an average distance error of 1.20 cm, and an average number of feature points that are completely consistent with the benchmark feature points of 7.6. It should be noted that the target model includes abnormal body shapes such as obesity. For example, among the calibrated feature points of 6 target models, the number of feature points that are completely consistent with the benchmark is less than or equal to 3.

[0136] Table 2 Distance error between the calibrated feature points and the corresponding reference feature points

[0137] Unit: cm

[0138]

[0139] The present invention can solve the problem of determining the characteristic points of the upper body of the human body required for the prototype of the jacket pattern based on the unmarked three-dimensional point cloud data of the human body obtained by an ordinary depth camera in an everyday environment by an untrained customer. The characteristic points calibrated by the present invention do not need to have characteristic information such as curvature, edge, circumference, etc., and can be any characteristic points of any part of the human body, and can be arbitrarily specified according to the application requirements. The accuracy of the characteristic points obtained by the present invention and the error are within a reasonable range acceptable for practical applications. Taking into account the convenience, effect and efficiency of the measurement method, the present method is practical and reliable in the field of fully online clothing customization or other human factor engineering applications. As the scope of application expands, the template model library used by the present invention will be richer and more sufficient in the number of models and body types, which will help to improve the accuracy of characteristic point calibration.

[0140] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only illustrative of the principles of the present invention. Without departing from the spirit and scope of the present invention, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for calibrating feature points of an unmarked upper body point cloud model based on an ICP algorithm, characterized in that: The method comprises: calculating a similarity function value using readily available human body parameters of a target model to be calibrated in a template model library with calibrated feature points; determining a matching template model based on the similarity function value, and applying a non-rigid ICP registration algorithm to preliminarily determine a feature point set of the target model to be calibrated; and performing rigid ICP registration again on local point clouds around each feature point of the template model and the target model to be calibrated to obtain feature point position information of the target model to be calibrated; The method comprises the following steps: Step 1: In a template model library of 3D human point clouds of different body types with calibrated feature points, the similarity function value is calculated using the readily available human body parameters of the 3D human point cloud model to be calibrated, i.e., the target model, and the template model with the highest similarity is selected. Step 2: Select the upper body point clouds of the target model and the template model selected in step 1, and determine the sequence number of each feature point in the point cloud array on the upper body point cloud of the template model; Step 3: Using the upper body point cloud of the template model selected in step 2 as the source point cloud and the upper body point cloud of the target model as the target point cloud, perform non-rigid ICP registration to obtain the transformed source point cloud and its feature point sequence number set; Step 4: On the source point cloud obtained after non-rigid ICP registration in step 3, the corresponding feature point coordinates are found. These coordinate values ​​are used as input to find the nearest point of each feature point on the target model in step 1 to obtain the feature point set of the target model. Step 5: For each feature point on the source point cloud after non-rigid ICP registration, select the surrounding points as the local source point cloud; for each preliminarily calibrated feature point on the target point cloud, select the surrounding points as the local target point cloud; In step 6, for each pair of corresponding local source point clouds and local target point clouds obtained in step 5, rigid ICP registration is performed to obtain the coordinate values ​​of each feature point on the transformed local source point cloud. Based on these coordinates, the calibrated feature point set is found on the target point cloud.

2. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: The feature points calibrated in step 1 are the key projection points of the upper garment design pattern prototype, including: front neck point, vertebral arch, right shoulder and neck point, left shoulder and neck point, left breast point, right breast point, front right front armpit point, front left front armpit point, right acromion, left acromion, front fossa, center point between right acromion and right front armpit point, center point between left acromion and left front armpit point, back right back armpit point, back left back armpit point, center point between right acromion and right back armpit point, center point between left acromion and left back armpit point, back vest, front waistline point, right waistline point, left waistline point and back waistline point.

3. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: The human body parameters that can be easily obtained in step 1 include any combination of the following parameters: gender, age, height, weight, chest circumference, waist circumference, and hip circumference.

4. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: The calculation formula for calculating the similarity function value in step 1 is: Where S is the similarity value between the template model and the target model, n is the number of human body parameters, α i represents the coefficient, p i is the i-th human body parameter of the target model, p m,i is the i-th human body parameter of template model m.

5. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: In step 2, the topological information of the template model point cloud and the calibrated feature point positions are used to select the upper body point cloud models of the target model and the template model. The point set and triangle face set of the selected upper body point cloud models are subsets of the point set and triangle face set of the respective original models.

6. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: In step 3, for the source point cloud and the target point cloud, the distance error and the smoothing error are used to form the objective function, and the transformation matrix is ​​calculated for each pair of points to perform non-rigid ICP registration.

7. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: In step 4, the coordinates of each feature point are found by the serial number, and based on the coordinates of the feature points on the source point cloud, the nearest points of each feature point are found on the target point cloud before non-rigid ICP registration to obtain the feature point set for the preliminary calibration of the target model.

8. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: In step 5, on the target point cloud before non-rigid ICP registration, N local point clouds within N circular domains are selected as local target point cloud sets with the N initially calibrated feature points as the circle center and r as the radius; on the source point cloud after non-rigid ICP registration, N local point clouds within N circular domains are selected as local source point cloud sets with the N feature points as the circle center and r as the radius, and at the same time, the coordinate values ​​of the N feature points on each local source point cloud determine their respective serial numbers on the local point cloud; where N is the number of feature points.

9. The method for calibrating feature points of an unmarked upper body point cloud model based on the ICP algorithm according to claim 1, characterized in that: Step 6 includes: Using distance error as the objective function, each point cloud in the local source point cloud set is used as the source point cloud, and the corresponding point cloud in the local target point cloud set is used as the target point cloud to perform rigid ICP registration. On each source point cloud after registration, determine the corresponding feature point coordinate value according to the sequence number of each feature point; Based on the coordinate values ​​of the N feature points, the nearest points of each are found on the target point cloud in step 3 to obtain the final set of N feature points on the calibrated target point cloud; where N is the number of feature points.

Images