Method for digitizing hub workpiece on basis of implicit three-dimensional reconstruction

Abstract

Disclosed is a method for digitizing a hub workpiece on the basis of implicit three-dimensional reconstruction. The present invention comprises the following steps: step 1, constructing a neural dista

Description

[0002] The present disclosure relates to the fields of a Radiation Field (RF), a Neural Signed Distance field (NSdf) and Point Cloud Registration (PCR), and in particular to a digitization method of wheel hub workpieces based on implicit three-dimensional reconstruction. BACKGROUND

[0003] In the field of three-dimensional reconstruction, it is researched that an original three-dimensional model is restored by other information, such as an image, without an original model.

[0004] A neural signed distance field obtains an implicit space model by fitting the three-dimensional structure of the real space. The essence of the neural signed distance field is to store the closest distance from each point to the surface of the object. The distance value of the outer side of the object is greater than 0, and the distance value of at the inner side of the object is less than 0. Specifically, the signed distance can be predicted through a Multi-Layer Perceptron (MLP) as a model.

[0005] A Neural Radiance Field (NeRF) realizes the three-dimensional reconstruction in a voxel space. The MLP inputs the coordinate (x, y, z) of the spatial point and the view direction (9, (p) of observing the spatial point, and outputs the voxels on the spatial point and the corresponding color Red, Green, Blue (RGB) values. The camera pose of a certain picture (consisted of the camera coordinate and the view direction) is known. Under this pose, for each pixel, the inferred picture is acquired by integrating the color and the voxel output by the MLP, and the loss is acquired by comparing the inferred picture with the real picture pixel by pixel for reverse optimization. By using pictures from multiple views as true values to train the process, the MLP can be fitted to a complex function, so that appropriate voxel and corresponding color RGB values can be output for any point in space. There is also a lot of work to improve the operation, such as adding distance cutoff as a constraint to the accumulation process of voxels. These improvements have effectively improved the accuracy of implicit reconstruction, and the work has effectively improved the fineness of reconstructed pictures.

[0006] Point cloud registration refers to finding a rigid spatial transformation including rotation and translation within the scope of this application, so that the overlap of two set of point clouds should be as high as possible. Point cloud registration is divided into two steps: coarse registration and fine registration. Coarse registration refers to carrying out rough registration when the transformation between the two set of point clouds is completely unknown to obtain the corresponding relationship between the associated points of the two set of point clouds. Fine registration refers to optimizing the transformation under the corresponding relationship of the point clouds after roughly registered. Fine registration can be completed by an Iterative Closest Point (ICP) algorithm.

[0007] The problems existing in the existing methods are as follows.

[0008] 1. A Signed Distance Field (SDF) process does not learn texture information. The Signed Distance Field (SDF) process, whether it is the traditional calculation method or the neural network fitting algorithm, relies heavily on the effective three-dimensional structural information. In the region where the original spatial structure is missing, the model is often inaccurate, resulting in a wrong void.

[0009] 2. NeRF has a poor learning effect on the surface with inconspicuous texture. Strictly speaking, the main job of the NeRF is not to carry out three-dimensional reconstruction, and especially the design of the volume density of the NeRF generates "ghost" on the surface of objects or even in open regions easily. It is difficult to accurately represent objects.

[0010] 3. The existing method has a poor reconstruction effect on the bottom part that is difficult to capture. By both the SDF and the NeRF, reconstruction is performed from the existing view. For the parts that are not displayed in the image and the depth map, the algorithm only guesses to form a predicted shape, which is far from the real object. SUMMARY

[0011] Aiming at the shortcomings of the prior art, the present disclosure provides a digitization method of wheel hub workpieces based on implicit three-dimensional reconstruction, which fully utilizes the morphological features of the wheel hub to better overcome the above problems. The method includes the following steps:

[0012] step 1, constructing a neural signed distance field model, including:

[0013] constructing a neural signed distance field model that characterizes a signed distance using a Multi-Layer Perceptron (MLP), which has a coordinate (x, y, z) of a spatial point and a view direction (6, o) as inputs and outputs a signed distance doutand a probability color cout of the spatial point in the view direction;

[0014] step 2, optimizing the neural signed distance field model;

[0015] adding a strong constraint to a prediction of the signed distance dout by means of a camera pose and a depth map;

[0016] step 3, extracting point cloud information;

[0017] extracting a smooth discrete surface of a wheel hub model through searching for a set with signed distances dout = 0 by the model;

[0018] step 4, acquiring a center point and a normal line;

[0019] step 5: performing point cloud rotation and completion;

[0020] performing rotation and completion on the point cloud P around the center point and the normal line acquired according to Tullkll0W. and completing a space inconsistent with spatial structure through correct spatial structure information.

[0021] In an embodiment, step 2 is as follows:

[0022] step 2-1, in the process of sampling along a ray of light, the depth map provides a strict mandatory distance constraint value dtrUe; and according to the design of the neural signed distance field model, through causing the signed distance dout output by the neutral distance field model for a point in space to be close to a true constraint value dtrue, constructing a loss Lossa:

[0023] Lossa ||dtrue"doutII2 (1)

[0024] step 2-2, for a void captured in the depth map, the void is filled using an estimated value of a color acquired by a radiation field, and an approximate volume density v, of a current point i is solved by Gaussian probability distribution on the signed distance field: / 7 2 1 out V = —C 7 ' / 5-

[0025] (2)

[0026] a density probability color Ci of the current point i in space in the view direction (0, o) is expressed as: d 2 uout c. = vc = e 2 i i out / T

[0027] (3)

[0028] a predicted color of a point on an image at a current view is acquired by accumulating the colors of the whole ray of light:

[0029] ^prediction I , J closet d 2 farthest C out 2

[0030]

[0031]

[0032] a loss Lossc is constructed: LOSSC ||Cpredicition"Ctrue|| (5) where Ctnte denotes a color of a corresponding point on a camera picture; closest denotes a closest position, which physically denotes starting position of a ray of light emitted from the camera when the camera takes pictures; farthest denotes a farthest position, which physically denotes an infinite position after the ray of light is emitted from the camera, and thus a value of closest is 0 and a value offarthest is +co; C0llt denotes a probability color of a spatial point (x, y, z) in the view direction (0, a) in the neural signed distance field model constructed in the step 1.

[0033] In an embodiment, the neural signed distance field model is optimized by Lossa and Lossc; assuming that (al, a2, a3) and (bl, b2, b3) schematically shown in a virtual space denote pixel points (x, y) in an n-th group of depth map and color map corresponding to a current point (x, y, n); first, it is determined whether there is a depth of (x, y, n) in the depth map, and if there is a depth value corresponding to (bl, b2, b3), the signed distance is optimized by using both a depth and a neural predicted color at the same time; and if there is no (al, a2, a3), the signed distance is optimized by only using a neural rendering predicted color.

[0034] In an embodiment, extracting the point cloud information in the step 3 is as follows:

[0035] since some regions are affected by image capture and noise interference, there are discontinuities in the signed distance fitted by the MLP, which neither conform to the definition of a continuous smooth surface nor conform to an actual shape of the real wheel hub; and it is necessary to conduct manual calibration and a voting method in advance to finally obtain the point cloud P.

[0036] Further, acquiring a center point and a normal line in step 4 is as follows:

[0037] step 4-1, conducting calibration using a scale-invariant feature transform (SIFT) algorithm, and acquiring a SIFT feature in the image; where the SIFT feature is a local feature of the image, and the operation is as follows:

[0038] capturing an undistorted large front picture Pa of the wheel hub, marking connection points between spokes and an outer ring of the wheel hub as well as a circle center as feature points, extracting features of the connection points by the SIFT algorithm to acquire a corresponding feature vector; in a data set used to optimize the neural signed distance field model, searching for three front pictures Pb, Pc and Pa containing the wheel hub, and acquiring plane coordinates of the connection points between the spokes and the outer ring as well as the circle center of the wheel hub in the front pictures Pb, Pc and Pa through comparison by the SIFT algorithm;

[0039] where the number of the connection points is determined according to the spokes of the wheel hub, and the number of connection points is equal to the number of spokes;

[0040] step 4-2, calculating the normal line and the center point by means of the circle center of the wheel hub and the camera pose acquired by the front pictures Pb, Pc and Pa , where the operation is as follows:

[0041] assuming that a circle center of plane of the front picture Pb is Ob, and determining a straight line Lb through Ob by means of the camera pose, where the straight line passes through Ob and is perpendicular to a plane of the front picture Pb;

[0042] similarly, acquiring straight lines Lc and La; since the straight lines Lb, Lc and La all pass through a center point O of the wheel hub in real space, and Lb, Lc and Ld are not coincident with each other, the center point O of the wheel hub is certainly a point of intersection of the straight lines Lb, Lc and Ld;

[0043] similarly, determining real spatial coordinates of key points of the front side of the wheel hub, calculating the plane representation of the front side of the wheel hub by means of the spatial coordinates of the key points through which the spokes are connected with the wheel hub, and acquiring a normal vector L of the front side of the wheel hub indirectly;

[0044] step 4-3, optimizing the normal vector and the center point of the wheel hub.

[0045] The method of optimizing the normal vector and the center point of the wheel hub uses an iterative closest point algorithm to carry out point cloud registration, and the specific method is as follows:

[0046] step 4-3-1, expanding the point cloud P extracted in step 3 after removing the ground to obtain expanded point cloud Pi, rotating Pi by a specific angle a according to the normal line to obtain point cloud P2, and denoting the rotating as Taipha;

[0047] step 4-3-2, using a greedy algorithm between the point cloud Pi and the point cloud P2, searching for closest points in distance as a corresponding point pair, calculating Euclidean distance between corresponding points, setting a threshold, eliminating corresponding point pair in which the Euclidean distance is greater than the threshold, and constructing Lose, as shown in Formula 6:

[0048] Lose =||TunknowPorigin_TaipbaTunialowPorjgjn| | (6)

[0049] where Tunknow denotes an unknown spatial change and is an object to be solved; if and only if Tunkll0W is the solution, which means that Lose is minimum, point cloud Pend acquired after the point cloud Ptrans acquired by TunknowPorigin is transformed by Talpha is just coincident with Ptrans; a true position of the center point and direction of the normal line are acquired according to Tunknow, where a physical meaning of Tullkll0W is the spatial transformation carried out on the point cloud of the wheel hub to minimize the Lose while keeping the normal line and the center point unchanged; conversely, when the transformation is known, the point cloud is fixed, and an inverse process of spatial transformation is applied to the normal line and central point; Porigin denotes the original point cloud expanded after removing the ground in step 4-3-1, which is equal to Pi numerically;

[0050] considering that a gradual process of spatial points is nonlinear, which is represented by using a special orthogonal group SO(3) of Lie algebra and a special Euclidean group SE(3), denoting the spatial transformation of the point cloud as T, and using a perturbation model of Lie algebra for nonlinear optimization, as shown in Formula 7, where AT is the left multiplication perturbation; the specific method is to update AT into T upon finding a derivative of AT equal to zero, as shown in Formula 8, and finally AT becomes an E matrix, and the value of Tunknow is optimized;

[0051] Lose ||ATTuniaiowPorigin-TalphaATTun]fliowPorigin|| (7)

[0052] Tt+i^AT, Tt (8)

[0053] where Tt+i denotes the (t+l)-th spatial transformation; Tt denotes the t-th spatial transformation.

[0054] In an embodiment, performing point cloud rotation and completion in step 5 is as follows:

[0055] performing rotation and completion on the point cloud P around the center point and the normal line acquired according to Tunknow, and completing the a space inconsistent with spatial structure through correct spatial structure information; assuming that the wheel hub has N spokes, and acquiring N pieces of spatial structure information at the same angle through spatial rotation;

[0056] superimposing N pieces of spatial structure information after subjecting to the expansion operation, into a same space, recording the vote at the same spatial point, and determining whether there is a plurality of pieces of spatial structure information at this point; and reserving points with more than half of the votes and discarding other points;

[0057] performing an erosion operation on merged point cloud until the merged point cloud after erosion operation is reserved as a layer of point cloud;

[0058] finally, the layer of point cloud being the solved point cloud of surface of the wheel hub, and thus finishing the digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction.

[0059] The present disclosure has the following beneficial effect.

[0060] 1. The present disclosure spans the fields of a Radiation Field (RF) and a Neural Directed Distance field (NDdf), fully incorporates spatial structure information constructed by a depth map into a training process of the RF in the form of a Signed Distance Field (SDF), realizes the joint training of two modal data, and enables the point cloud to be more accurate.

[0061] 2. Different from the prior visual methods in which it is difficult to train the symmetrical structure, the present disclosure actively uses the spatial features of the scene to solve the problem.

[0062] 3. The present disclosure not only uses the field of deep learning, but also excavates a mathematical three-dimensional reconstruction method and Lie algebra, which is more interpretable than a deep learning method.

[0063] 4. The present disclosure provides a method of optimizing point clouds for spatial symmetry and spatial rotation symmetry, which can be used as a standard to be extended to other scenes with spatial symmetry and spatial rotation symmetry. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] FIG. 1 is a flow chart of the present disclosure.

[0065] FIG. 2 is a diagram of a neural signed distance field model according to the present disclosure.

[0066] FIG. 3 is a diagram of a process of training a neural signed distance field model according to the present disclosure.

[0067] FIG. 4 is a schematic diagram of extracting point clouds in an implicit space.

[0068] FIG. 5 is a schematic diagram of extracting corresponding feature points using a Scale-Invariant Feature Transform (SIFT) method.

[0069] FIG. 6 is a schematic diagram of searching for a center point in an implicit space according to feature points.

[0070] FIG. 7 is a schematic diagram of optimizing a normal line and a center point using a Lie algebra method.

[0071] FIG. 8 is a flow chart of acquiring a surface of a wheel hub model after expansion, voting and erosion of the space. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] The present disclosure will be further explained with reference to the attached drawings and embodiments.

[0073] The present disclosure provides a digitization method of workpieces based on implicit three-dimensional reconstruction and neural signed distance field. Based on the neural signed distance field, the smooth surface geometry of the object can be acquired. Based on the neural radiance field, color information can be added into the training, and the training of the network can be accelerated. The model that can represent the smooth workpiece surface can be acquired quickly and well.

[0074] The most important thing is to use the rotation symmetry, which means that the graphic after rotating by a certain angle in the plane coincides with the original graphic. The wheel hub is manufactured under industrial conditions. The wheel hub also follows the three-dimensional rotation symmetry, that is, the wheel hub coincides with the original wheel hub in space after rotating in a certain direction for a certain angle.

[0075] As shown in FIG. 1, the method of the present disclosure mainly includes the following steps.

[0076] step 1, a neural signed distance field model is constructed.

[0077] A neural signed distance field model that characterizes a signed distance using a Multi-Layer Perceptron (MLP) is constructed. The neural signed distance field model is as shown in FIG. 2. A coordinate (x, y, z) of a spatial point and a view direction (0, o) are set as inputs to output a signed distance dout and a probability color cout of the point in the direction.

[0078] step 2, the neural signed distance field model is optimized.

[0079] A strong constraint is added to the prediction of the signed distance dout by means of a camera pose and a depth map. Specifically, in the process of sampling along a ray of light, the depth map provides a strict mandatory distance constraint value dtmL and according to the design of the model, the signed distance dout output by the neutral distance field model for a point in space is close to a true constraint value dtrae, so that a loss Lossa is constructed:

[0080] Lossa ||dtrue"doutII2 (1)

[0081] However, in a void captured in the depth map, the void is filled using an estimated value of a color acquired by a radiation field, and an approximate volume density Vj of a current point i is solved by Gaussian probability distribution on the signed distance field: ] ^0» / V = —S z Fj~

[0082] (2)

[0083] a density probability color Ci of the current point i in space in the view direction (0, g) is expressed as: d 2 X-> UOllt c = v c = 111— e 2 i i out rz

[0084] yj 171

[0085] a predicted color of a point on an image at the current view is acquired by accumulating the colors of the whole ray of light: farthest f farthest C c = c = Zut e 2 ^prediction \ i + t I i / T c r J closet J closet

[0086] V (4)

[0087] a loss Lossc is constructed as follow:

[0088] LOSSC = IlCpredicition^Ctniell2 (5)

[0089] where Ctnie denotes a color of a corresponding point on a camera picture; closest denotes the closest position, which physically denotes the starting position of the ray of light emitted from the camera when the camera takes pictures; farthest denotes the farthest position, which physically denotes the infinite position after the ray of light is emitted from the camera, so that the value of closest is 0 and the value offarthest is +oo; Cout denotes the probability color of the spatial point (x, y, z) in the view direction (0, a) in the neural signed distance field model constructed in step 1.

[0090] In particular, there are significant differences in color between the metal wheel hub and the environment, especially in color and lustre and lightness. The RGB color cannot represent the difference in color and lustre and lightness well. Therefore, the Hue, Saturation, Lightness (HSL) color space is used to linearly convert the RGB color into hue, saturation and lightness. The hue is calculated and normalized according to the counterclockwise angle difference between the predicted value and the real value. Moreover, the accumulated color of the whole ray of light can be constrained by means of the high consistency of hue in a metal surface. Specifically, the penalty function is used to ensure that the final predicted color is close to color of the metal surface.

[0091] The depth map network is optimized by Lossaand Lossc. As shown in FIG. 3, (1, 2, 1) and (4, 1, 1) schematically shown in the virtual space in the figure denote pixel point (x, y) in an n-th group of depth map and color map corresponding to the current point (x, y, n). First, it is determined whether there is the depth of (x, y, n) in the depth map, and if there is d41 corresponding to (4, 1, 1), the signed distance is optimized by using both a depth and a neural predicted color at the same time; and if there is no (1, 2, 1), the signed distance is optimized by only using a neural rendering predicted color.

[0092] Step 3, point cloud information is extracted.

[0093] As introduced in the background, a positive SDF distance indicates that the point is outside the object, and a negative SDF distance indicates that the point is inside the object. As shown in FIG. 4, a smooth discrete surface of a wheel hub model is extracted by searching a set with the SDF distance of 0 in the wheel hub model, that is, a set with signed distances dollt = 0.

[0094] However, since some regions are affected by image capture and noise interference, there are discontinuities in the signed distance fitted by the MLP, which neither conform to the definition of a continuous smooth surface nor conform to an actual shape of the real wheel hub. This is the noise resulted from an original data. For example, when there is a large range of reflection, it is easy for the signed distance fitted by the MLP to generate a point cloud in space, which physically means a luminous object in space that does not actually exist, so that the luminous object needs to be discarded. The point cloud P can be acquired finally by manual calibration in advance and a voting method in step 5.

[0095] Step 4, a center point and a normal line are acquired.

[0096] Step 4-1. Calibration is conducted using a Scale-Invariant Feature Transform (SIFT) algorithm, and a SIFT feature in the image is acquired. The SIFT feature is the local feature of the image, which is invariant to rotation, size scaling and lightness change, and also stable to a certain extent to view change, affine transformation and noise. The specific operation is as follows.

[0097] As shown in FIG. 5, an undistorted large front picture Pa of the wheel hub is captured. Connection points between spokes and an outer ring of the wheel hub as well as the circle center are marked as feature points. Features of the connection points are extracted by the SIFT algorithm to acquire a corresponding feature vectors. In a data set used to optimize the neural signed distance field model, three front pictures Pb, Pc and Pa containing the wheel hub are searched for. Plane coordinates of the connection point between the spoke and the outer ring as well as the circle center of the wheel hub in the front pictures Pb, Pc and Pa are acquired through comparison by the SIFT algorithm.

[0098] The number of the connection points is determined according to the spokes of the wheel hub, and the number of connection points is equal to the number of spokes.

[0099] Step 4-2. A normal line and a center point are calculated by means of the circle center of the wheel hub and the camera pose acquired by the front pictures Pb, Pc and Pa, where the specific operation is as follows.

[0100] As shown in FIG. 6, it is assumes that the circle center of the plane of the front picture Pb is Ob. a straight line Lb is determined through Ob by means of the camera pose, where the straight line passes through Ob and is perpendicular to the plane of the front picture Pb. Similarly, straight lines Lc and La are acquired. Since the straight lines Lb, Lc and La all pass through the center point O of the wheel hub in real space, and Lb, Lc and La are not coincident with each other, the center point O of the wheel hub is certainly a point of intersection of the straight lines Lb, Lc and Ld. Similarly, real spatial coordinates of key points of the front side of the wheel hub are determined. The plane representation of the front side of the wheel hub is calculated by means of the spatial coordinates of the key points through which the spokes are connected with the wheel hub. A normal vector L of the front side of the wheel hub is acquired indirectly.

[0101] Step 4-3. The normal vector and the center point of the wheel hub are optimized.

[0102] Theoretically, O and L should be the determined normal vector in a space and the spatial center point of the wheel hub, but there may be some deviation due to the error of image capture. As shown in FIG. 7, the used optimization method uses an iterative closest point algorithm to carry out point cloud registration, and the specific method is as follows.

[0103] In reality, the wheel hub is three-dimensional rotationally symmetric, that is, the wheel hub that is consistent with that before rotation can be acquired only by rotating the front side of the wheel hub by the angle (a = (360 degrees / number of spokes)) corresponding to the number of spokes. The perfect point cloud of the wheel hub is also rotationally symmetric. Therefore, by means of this feature, the spatial transformation of rough registration of point cloud registration is set as no transformation. When determining the number of spokes of the wheel hub, the number of spokes can be determined by manual input or be predicted by simple convolutional neural network.

[0104] Step 4-3-2. A greedy algorithm is used between the point cloud Pi and the point cloud P2. The closest points in distance are searched for as a corresponding point pair. The Euclidean distance between the corresponding points is calculated. A threshold is set, and the corresponding point pair in which the Euclidean distance is greater than the threshold is eliminated. Lose is constructed, as shown in Formula 6:

[0105] Lose 1'imkiiouP origin"TalphaTunknowPoriginII (6)

[0106] where Tullknow denotes an unknown spatial change and is the object to be solved; if and only if Tunknow is the solution that minimize Lose, the point cloud Pend acquired after the point cloud Ptrans acquired by TunknOwPorigin is transformed by Taipha is just coincident with Ptrans; the true position of the center point and direction of the normal line are acquired according to Tunknow, where the physical meaning of Tunknow is the spatial transformation that the point cloud of the wheel hub needs to carry out to minimize the Lose when keeping the normal line and the center point unchanged; conversely, when the transformation is known, the point cloud is fixed, and an inverse process of spatial transformation is applied to the normal line and central point; POrigin denotes the original point cloud expanded after removing the ground in step 4-3-1, which is equal to Pi numerically. Considering that a gradual process of spatial points is nonlinear, which is represented by using a special orthogonal group SO(3) of Lie algebra and a special Euclidean group SE(3), the spatial transformation of the point cloud is denoted as T, and a perturbation model of Lie algebra is used for nonlinear optimization, as shown in Formula 7, where AT is the left multiplication perturbation; the specific method is to update AT into T upon finding a derivative of AT equal to zero, as shown in Formula 8, and finally AT becomes an E matrix, and the value of Tunknow is optimized

[0107] Lose=||ATTimknowPorigin-Talpha^TTuni^QyvP originII (7)

[0108] Tt+i^AT, Tt (8)

[0109] where Tt+i denotes the (t+1 )-th spatial transformation; Tt denotes the t-th spatial transformation.

[0110] Step 5: point cloud rotation and completion is performed.

[0111] As shown in FIG. 8, rotation and completion are performed on the point cloud P around the center point and the normal line acquired according to Tunkn0W. A space inconsistent with spatial structure is completed through the correct spatial structure information. Taking a 5-spoke wheel hub as an example, 5 pieces of spatial structure information can be acquired at the same angle through spatial rotation.

[0112] A plurality of pieces of spatial structure information are superimposed into the same space after the expansion operation is performed. The vote at the same spatial point is recorded (determining whether there is a plurality of pieces of spatial structure information at this point). The points with more than half of the votes are reserved, and other points are discarded.

[0113] An erosion operation is performed on the merged point cloud until the merged point after erosion operation is reserved as a layer of point cloud. Finally, the layer of point cloud is the solved point cloud on the surface of the wheel hub. The digitization method of wheel hub workpieces based on implicit three-dimensional reconstruction is finished so far.

Claims

1. A digitization method of wheel hub workpieces based on implicit three-dimensional reconstruction, comprising:step 1, constructing a neural signed distance field model, comprising:constructing the neural signed distance field model that characterizes a signed distance using a Multi-Layer Perceptron (MLP), which has a coordinate (x, y, z) of a spatial point and a view direction (6, o) as inputs and outputs a signed distance doutand a probability color cout of the spatial point in the view direction;step 2, optimizing the neural signed distance field model, comprising:adding a strong constraint to a prediction of the signed distance dout by means of a camera pose and a depth map;step 3, extracting point cloud information, comprising:extracting a smooth discrete surface of a wheel hub model through searching for a set with signed distances dout = 0 by the model;step 4, acquiring a center point and a normal line;step 5, performing point cloud rotation and completion, comprising:completing spatial points directly represented by the neural signed distance field model through voting after point cloud rotation and expansion by means of spatial rotation of a wheel hub, and acquiring surface point cloud of the wheel hub by spatial point erosion.

2. The digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction according to claim 1, wherein the step 2 is as follows:step 2-1, in a process of sampling along a ray of light, providing, by the depth map, a strict mandatory distance constraint value dtrue; and according to a design of the neural signed distance field model, through causing the signed distance dout output by the neutral distance field model for a point in space to be close to a true constraint value dtrue, constructing a loss Lossa:Lossd=||dtrae-dout||2; (1)step 2-2, for a void captured in the depth map, filling the void using an estimated value of a color acquired by a radiation field, and solving an approximate volume density v, of a current point i by Gaussian probability distribution on the signed distance field:1V — —..................................................p 7^271expressing a density probability color q of the current point i in space in the view direction(0, o) as:Ci = ^outCoutd 2Uoute 2acquiring a predicted color of a point on an image at a current view by accumulating colors of whole ray of light:tfanlKSIprediction I , ^i I , / TJ closet J closet ~r; (4)constructing a loss Lossc:LoSSc=||Cpredicition-Ctrue||2 ; (5)wherein CtrUe denotes a color of a corresponding point on a camera picture; closest denotes a closest position, which physically denotes a starting position of a ray of light emitted from a camera when the camera takes pictures; farthest denotes a farthest position, which physically denotes an infinite position after the ray of the light is emitted from the camera, and thus a value of closest is 0 and a value offarthest is +co; Cout denotes a probability color of a spatial point (x, y, z) in the view direction (6, a) in the neural signed distance field model constructed in the step 1.

3. The digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction according to claim 2, wherein the neural signed distance field model is optimized by Lossa and Lossc;assuming that (al, a2, a3) and (bl, b2, b3) schematically shown in a virtual space denote pixel points (x, y) in an n-th group of depth map and color map corresponding to a current point (x, y, n);first, determining whether there is a depth of (x, y, n) in the depth map, and in a case that there is a depth value corresponding to (bl, b2, b3), optimizing the signed distance by using both a depth and a neural predicted color at the same time; and in a case that there is no (al, a2, a3), optimizing the signed distance by only using a neural rendering predicted color.

4. The digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction according to claim 2 or 3, wherein extracting the point cloud information in the step 3 is as follows:in view of a fact that since some regions are affected by image capture and noise interference, there are discontinuities in the signed distance fitted by the MLP, which neither conform to a definition of a continuous smooth surface nor conform to an actual shape of a real wheel hub, conducting manual calibration and a voting method in advance to finally obtain point 14cloud P.

5. The digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction according to claim 2 or 3, wherein acquiring the center point and the normal line in the step 4 is as follows:step 4-1, conducting calibration using a Scale-Invariant Feature Transform (SIFT) algorithm, and acquiring a SIFT feature in the image; wherein the SIFT feature is a local feature of the image, and operation of the step 4-1 is as follows:capturing an undistorted large front picture Pa of the wheel hub, marking connection points between spokes and an outer ring of the wheel hub as well as a circle center as feature points, extracting features of the connection points by the SIFT algorithm to acquire a corresponding feature vector; in a data set used to optimize the neural signed distance field model, searching for three front pictures Pb, Pe and Pa containing the wheel hub, and acquiring plane coordinates of the connection points between the spokes and the outer ring as well as the circle center of the wheel hub in the front pictures Pb, Pc and Pa through comparison by the SIFT algorithm;wherein a number of the connection points is determined according to the spokes of the wheel hub, and a number of connection points is equal to a number of spokes;step 4-2, calculating the normal line and the center point by means of the circle center of the wheel hub and the camera pose acquired by the front pictures Pb, Pc and Pa , wherein operation of the step 4-2 is as follows:assuming that a circle center of a plane of front picture Pb is Ob, and determining a straight line Lb through Ob by means of the camera pose, wherein the straight line passes through Ob and is perpendicular to the plane of the front picture Pb;similarly, acquiring straight lines Lc and Ld; wherein, since the straight lines Lb, Lc and Ld all pass through a center point O of the wheel hub in real space, and Lb, Lc and Ld are not coincident with each other, the center point O of the wheel hub is certainly a point of intersection of the straight lines Lb, Lc and Ld;similarly, determining real spatial coordinates of key points of front side of the wheel hub, calculating a plane representation of the front side of the wheel hub by means of spatial coordinates of the key points through which the spokes are connected with the wheel hub, and acquiring a normal vector L of the front side of the wheel hub indirectly;step 4-3, optimizing the normal vector and the center point of the wheel hub.

6. The digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction according to claim 5, wherein a method of optimizing the normal vector and the center point of the wheel hub in the step 4-3 uses an iterative closest pointalgorithm to carry out point cloud registration, which is as follows:step 4-3-1, expanding the point cloud P extracted in the step 3 after removing the ground to obtain expanded point cloud Pb rotating Pi by a specific angle a according to the normal line to obtain point cloud P2, and denoting the rotating as Taipha;step 4-3-2, using a greedy algorithm between the point cloud Pi and the point cloud P2, searching for closest points in distance as a corresponding point pair, calculating Euclidean distance between corresponding points, setting a threshold, eliminating corresponding point pair in which the Euclidean distance is greater than the threshold, and constructing Lose, as shown in Formula 6:Lose TunknowPorigin-TalphaTunknowPorigin|| (6)wherein Tunknmv denotes an unknown spatial change and is an object to be solved; if and only if Tunknow is a solution, which means that Lose is minimum, point cloud Pend acquired after point cloud Ptrans acquired by TunknOwPorigin is transformed by Taipha is just coincident with Ptrans; a true position of the center point and direction of the normal line are acquired according to Tunkn0W, wherein a physical meaning of Tunkll0W is a spatial transformation carried out on point cloud of the wheel hub to minimize the Lose while keeping the normal line and the center point unchanged; conversely, when the transformation is known, the point cloud is fixed, and an inverse process of spatial transformation is applied to the normal line and the central point; POrigin denotes original point cloud expanded after removing the ground in the step 4-3-1, which is equal to Pi numerically;considering that a gradual process of spatial points is nonlinear, which is represented by using a special orthogonal group SO(3) of Lie algebra and a special Euclidean group SE(3), denoting the spatial transformation of the point cloud as T, and using a perturbation model of Lie algebra for nonlinear optimization, as shown in Formula 7, wherein AT is a left multiplication perturbation; specific operation is to update AT into T upon finding a derivative of AT equal to zero, as shown in Formula 8, and finally AT becomes an E matrix, and the value of Tunknow is optimized;Lose |ATTu^i^owPorjgjj-i-Tai^aATTunk]10WPiingn-i | , (7)Tt+1<—AT, Tt; (8)where Tt+i denotes a (t+l)-th spatial transformation; Tt denotes a t-th spatial transformation.

7. The digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction according to claim 6, wherein performing the point cloud rotation and completion in the step 5 is as follows:performing rotation and completion on the point cloud P around the center point and thenormal line acquired according to TUnknow, and completing a space inconsistent with spatial structure through correct spatial structure information; assuming that the wheel hub has N spokes, and acquiring N pieces of spatial structure information at a same angle through spatial rotation; superimposing the N pieces of spatial structure information after subjecting to expansion operation, into a same space, recording a vote at a same spatial point, and determining whether there is a plurality of pieces of spatial structure information at this point; and reserving points with more than half of votes and discarding other points;performing an erosion operation on merged point cloud until the merged point cloud after erosion operation is reserved as a layer of point cloud;finally, the layer of point cloud being solved point cloud of surface of the wheel hub, and thus finishing the digitization method of the wheel hub workpieces based on the implicit three-dimensional reconstruction.PCT / CN2024 / 075682A. CLASSIFICATION OF SUBJECT MATTERG06T17 / 00(2006.01)iAccording to International Patent Classification (IPC) or to both national classification and IPCB. FIELDS SEARCHEDMinimum documentation searched (classification system followed by classification symbols) IPC:G06TDocumentation searched other than minimum documentation to the extent that such documents are included in the fields searchedElectronic data base consulted during the international search (name of data base and, where practicable, search terms used)CNTXT, CNKI, DWPI, ENTXT, ENTXTC, IEEE: IM IM ft®, E1H, M®, ttSM, SliM three dimensional, reconstruct, NERF, distance, point cloud, central point, rotation, depthDOCUMENTS CONSIDERED TO BE RELEVANTCategory* Citation of document, with indication, where appropriate, of the relevant passages Relevant to claim No. PX CN 116597082 A (HANGZHOU DIANZI UNIVERSITY) 15 August 2023 (2023-08-15) claims 1-7 1-7 A CN 115294275 A (ZHUHAI PROMETHEUS VISION TECHNOLOGY CO., LTD.) 04 November 2022 (2022-11-04) description, paragraphs 0052-0091 1-7 A CN 115619951 A (ZHEJIANG UNIVERSITY) 17 January 2023 (2023-01-17) entire document 1-7 A CN 116051740 A (SOUTH CHINA UNIVERSITY OF TECHNOLOGY) 02 May 2023 (2023-05-02) entire document 1-7 A US 2023130281 Al (GOOGLE LLC.) 27 April 2023 (2023-04-27) entire document 1-7| | Further documents are listed in the continuation of Box C.annex.* Special categories of cited documents: “T” later document published after the international filing date or priority “A” document defining the general state of the art which is not considered date and not in conflict with the application but cited to understand the to be of particular relevance principle or theory underlying the invention “D” document cited by the applicant in die international application “X” document of particular relevance; the claimed invention cannot be “E" earlier application orpatent but published on or after the international considered novel or cannot be considered to involve an inventive step filing date when the document is taken alone •SL” document which may throw doubts on priority claim(s) or which is “Y” document of particular relevance; the claimed invention cannot be cited to establish the publication date of another citation or other considered to involve an inventive step when the document is special reason (as specified) combined with one or more other such documents, such combination “O” document referring to an oral disclosure, use, exhibition or other being obvious to a person skilled in the art means document member of the same patent family “P” document published prior to the international filing date but later than the priority date claimed Date of the actual completion of the international search 10 April 2024 Date of mailing of the international search report 25 April 2024 Name and mailing address of the ISA / CN China National Intellectual Property Administration (ISA / CN) China No. 6, Xitucheng Road, Jimenqiao, Haidian District, Beijing 100088 Authorized officer Telephone No.INTERNATIONAL SEARCH REPORT Information on patent family membersInternational application No.PCT / CN2024 / 075682Patent document cited in search report Publication date (day / month / year) Patent family member)s) Publication date (day / month / year) CN 116597082 A 15 August 2023 None CN 115294275 A 04 November 2022 US 2024046557 Al 08 February 2024 CN 115619951 A 17 January 2023 None CN 116051740 A 02 May 2023 None US 2023130281 Al 27 April 2023 None

Images