TW201915953A - Redundant point detection method for point cloud data attachment capable of effectively reducing the data storage cost and increasing the data transmission efficiency and computation efficiency - Google Patents

Abstract

Provided is a redundant point detection method for point cloud data attachment, which is performed by a data processing device. The method comprises: receiving the first and second point cloud data with a continuous relationship from a scanner; performing down-sampling on the first and second point cloud data to generate the first and second simplified point cloud data, respectively; establishing a three-dimensional point cloud index structure of the first simplified point cloud data for searching each data point in the second simplified point cloud data so as to correspond to each data point of the first simplified point cloud data; establishing an optimized conversion matrix, such that, after the second simplified point cloud data is converted through the optimized conversion matrix, there is a minimum distance between the tangent planes of two data points corresponding to the first and second simplified point cloud data; and determining the two data points to be the redundant data points when the minimum distance is smaller than a judgment value.

Description

Redundant point detection method for point cloud data bonding

This creation is about a redundant point detection method for point cloud data bonding, especially when two points with point cloud data are combined, and the duplicate redundant point data is detected. The rest of the method.

The conventional point cloud data scanning system uses a moving scanner to continuously scan a target object. Among them, each scan can obtain a point cloud data, the point cloud data contains a plurality of data points, and each data point contains a three-dimensional The coordinate information is used to construct an image of the target. Therefore, when performing fast scanning and acquisition, multiple point cloud data can be generated in a short time. The three-dimensional coordinate information of each point cloud data enters the scanning system at a speed of tens to hundreds of strokes per second.

For example, please refer to FIG. 8A, the scanner successively generates a first point cloud data 41 and a second point cloud data 42. Please refer to FIG. 8B. The scanning system may refer to the repeating area 40 of the two point cloud data 41 and 42. The two point cloud data 41 and 42 are adhered to each other, wherein the boundary 410 of the first point cloud data 41 falls within the range of the second point cloud data 42. In contrast, the boundary 420 of the second point cloud data 42 If it falls within the range of the first point cloud data 41, the repeated area 40 is between the two boundaries 410 and 420. Please refer to FIG. 8C and FIG. 8D, the scanner then generates a third point cloud data 43, and the scanning system then pastes the first and second point cloud data 41, 42 after the third point cloud data 43 is pasted. By analogy, please refer to FIG. 8E. The scanning system may receive a fourth point cloud data 44 or sequentially receive more point cloud data in order to further fit.

As mentioned above, the scanning system is based on the overlapping areas of two adjacent point cloud data, and the two point cloud data are connected and connected in series. As a whole, after the scan of the target object is completed, the points are at different points. Between cloud data, the repeatability of data points in its repeated regions is high, which not only increases the cost of data storage, but also affects the transmission efficiency and calculation efficiency of point cloud data.

In view of this, the main purpose of this creation is to provide a redundant point detection method for point cloud data bonding, which is used to detect the repeated redundant points of the bonded point cloud data. When eliminating redundant points , Which can effectively reduce the cost of data storage, and also improve data transmission efficiency and calculation efficiency.

This method is a redundant point detection method for point cloud data bonding, which is executed on a data processing device. The data processing device is connected to a scanner to receive point cloud data. The method includes: receiving a continuous relationship First point cloud data and a second point cloud data; downsampling the first point cloud data and the second point cloud data to become a first simplified point cloud data and a second simplified point cloud data, respectively; Establishing a three-dimensional point cloud index structure of the first simplified point cloud data for searching each data point in the second simplified point cloud data to correspond to each data point of the first simplified point cloud data; establishing an optimized transformation Matrix, after the second simplified point cloud data is transformed by the optimized transformation matrix, a tangent plane between the two simplified data points corresponding to the first simplified point cloud data and the second simplified point cloud data has a Minimum distance; judging whether the minimum distance is less than a judgment value; if yes, the two data points are redundant data points; if not, returning to "establishing a three-dimensional point cloud index structure of the first simplified point cloud data, for The step of searching each of the second data each data point in the point cloud data corresponding to the first data point of the point cloud simplified "to iterate the calculation of optimum conversion matrix.

According to the detection method of this creation, duplicate redundant data points in two simplified point cloud data are detected, and the redundant data points represent data points in the overlapping area scanned by the two simplified point cloud data. Therefore, any redundant data points that simplify the point cloud data can be eliminated in the application, so that duplicate data points between different point cloud data will not be repeatedly stored and calculated, reducing the data throughput, so it can effectively reduce the data storage. Cost also improves data transmission efficiency and calculation efficiency.

Please refer to FIG. 1, a system for performing the creative detection method includes a scanner 10 and a data processing device 20. The scanner 10 can scan a target object using structured light scanning technology to obtain a point cloud data. The point cloud data includes a plurality of data points, and these data points can form a map of the target object. For example, each data point may include brightness information, depth information (range image), and three-dimensional space parameters. The three-dimensional space parameters may include the x-axis rotation index value (yaw), x-coordinates, and the angle between the x-axis ( α), y-axis rotation index (pitch), y-coordinate, angle (β) with y-axis, z-axis rotation index (roll), z-coordinate, angle (γ) with z-axis, etc. The data processing device 20 may be a computer, but is not limited thereto. The data processing device 20 is connected to the scanner 10 to receive point cloud data captured by the scanner 10.

It should be noted that the embodiment of the creative detection method is described using an intra-oral scanning system as an example, but it is not limited thereto. The creative detection method can also be applied to other types of 3D reconstruction systems. For example, This creation can be applied to 3D inspection, rapid prototyping and 3D printing in various industrial fields ...

First, the overall architecture and flow of the creative detection method will be briefly explained. First, the scanner 10 moves along a trajectory. While moving, the scanner 10 continuously scans an object, wherein each scan can generate a point Cloud data, so continuous scanning can generate multiple point cloud data, and the data processing device 20 receives the point cloud data. Please refer to Figure 2A, consider the first point cloud data scanned successively And second point cloud data Since the scanner 10 scans continuously during travel, the two point cloud data , The content of is not completely consistent, but there are parts that overlap with each other. The data processing device 20 can adjust the second point cloud data. For correct fit to the first point cloud data . To adjust the second point cloud data Position, this creation establishes the first point cloud data 3D point cloud index structure. According to the 3D point cloud index structure, the two point cloud data can be obtained initially. , Correspondence between data points. Then, after several iterations in this creation, an optimized transformation matrix is generated for the second point cloud data. Fitting to the first point cloud data through the transformation of this optimized transformation matrix . After fitting, the two point cloud data , The overlapping parts have roughly overlapped, so the data in the two point clouds can be , Duplicate redundant data points were detected at the overlap.

The following describes the technical methods of this creation in detail. Please refer to FIG. 3 for reference. The embodiment of the detection method of this creation includes the following steps:

Step S1: Scan the target

The scanner 10 may be a hand-held scanner or a fixed scanner. Taking the hand-held scanner as an example, a dentist extends the scanner 10 into a patient's mouth to perform a tooth photographing operation. When the scanner 10 moves along the scanner 10 When the patient's tooth surface is moved and scanned, several tooth point cloud data with continuous relationship can be obtained. For illustration, please refer to FIG. 2A. In this work, the scanner 10 is used to continuously shoot during travel to obtain the first point cloud data. And second point cloud data For example, because the two point cloud data , The first point cloud data is continuously scanned while the scanner 10 is moving. And the second point cloud data Corresponds to different intraoral locations.

Step S2: Down-sampling

The data processing device 20 receives the first point cloud data And the second point cloud data Among them, the data processing device 20 may adopt uniform sampling, which acquires and stores the first point cloud data according to a sampling period. And the second point cloud data Data points. For example, when the first point cloud data With N data points, the sampling period can take one data point for every n data points, then the first point cloud data After downsampling, only N / n data points are included, so the number of data points after sampling is lower than the number of data points before sampling, thereby effectively reducing the amount of data transmission and calculation.

For illustration, the point cloud data after downsampling is called simplified point cloud data, so please refer to FIG. 2B, FIG. 4 and FIG. 5. This first point cloud data Become the first simplified point cloud data after downsampling , The second point cloud data Reduced sampling becomes second simplified point cloud data .

Step S3: establishing a three-dimensional point cloud index structure

In order to improve the search efficiency of data points, the author created the first simplified point cloud data. 3D point cloud index structure for searching the second simplified point cloud data Each data point in the corresponding to the first simplified point cloud data The purpose of creating the three-dimensional point cloud index structure in this creation is to simplify the point cloud data in the first place. Quickly search for a data point p _i in the data point p _i and the second simplified point cloud data One of the data points q _{j is the} closest. In this creative embodiment, a k-dimensional tree is used to establish a three-dimensional point cloud index structure. However, using a kd tree to establish a three-dimensional point cloud index structure of point cloud data belongs to the technical field. The general knowledge is only briefly explained as follows.

First explain in a one-dimensional case, assuming the first simplified point cloud data And second simplified point cloud data The dimension of is a dimension. Through binary search, the first simplified point cloud data can be All data points p _i are sorted according to their coordinate values to establish a one-dimensional point cloud index structure. Because one dimension is taken as an example, the first point cloud data is simplified. After the sorted data points p _i are expanded on the number line, let the second simplified point cloud data The data points q _{j are} compared with the data points p _i sorted in the middle element one by one, so that the second simplified point cloud data Every data point in q _j can be searched with the first simplified point cloud data The closest data point p _{i in} .

If the first simplified point cloud data And second simplified point cloud data The dimension of is three dimensions, then the aforementioned kd tree is generalized to establish the three-dimensional point cloud index structure of this creation. First, for the first simplified point cloud data, All data points p _i are sorted according to their coordinate values to establish a three-dimensional point cloud index structure. This structure includes a root node RN as shown in FIG. 6A. The root node RN contains the first simplified point cloud data. For all data points p _i , please refer to FIG. 6B. A bounding box BB stores all data points p _i . The bounding box BB is determined by the coordinates of the data points p _i at the maximum and minimum three-dimensional dimensions of x, y, and z. Value determination, which is the first simplified point cloud data In the data range of the three-axis coordinate system, each node below the root node RN represents the data range of the point cloud covered by the node. The node records its three-dimensional bounding box BB, the indicators of the left and right child nodes, and the cutting dimension as the basis for judgment. (cut dimension) and cut value, as well as the included data points, without loss of generality, if the median m1 of the three-dimensional point cloud Z coordinate corresponding to the root node RN is found along the Z axis, please refer to FIG. 6D, Then the plane Plane1 (ie, the plane of z = m1) created by the median m1 cuts the bounding box BB space into two spaces, which correspond to the two nodes SN shown in FIG. 6C; then repeat The above step of cutting the space, that is, for the lowest node at present, for example, the two nodes SN at the bottom of FIG. 6C, respectively, find the median m2 and m3 along the Y axis, as shown in FIG. 6F, the median m2 and m3 Each has a cutting plane Plane2 (that is, a plane of y = m2) and Plane3 (that is, a plane of y = m3). These two cutting planes Plane2 and Plane3 cut the two spaces divided by the plane Plane1 into two pieces. If at this time there is only one space left after the plane Plane2 and Plane3 are cut For dimension points, cutting is stopped. At this time, the nodes corresponding to the cut spaces are the leaf nodes LN shown in FIG. 6E, that is, the leaf nodes LN contain only one data point, and are the end points of the search. If there is more than one 3D point in the space after cutting, continue to cut along the X axis. Such cutting steps will be repeated until there is only one 3D point in the space.

Step S4: Establish the initial transformation matrix

Under the Cartesian coordinate system, this creation uses the initial x-axis angle a ', the initial y-axis angle b', the initial z-axis angle g ', the initial x-axis displacement tx', the initial y-axis displacement ty ', and the initial z-axis displacement tz'. Build a 4x4 initial transformation matrix , Expressed as follows:

among them Is the initial rotation matrix, Are the initial displacement matrices, which are expressed as follows:

In this creative embodiment, the initial x-axis included angle a ', the initial y-axis included angle b', the initial z-axis included angle g ', the initial x-axis displacement tx', the initial y-axis displacement ty ', and the initial z-axis displacement tz' can be respectively 0, but not limited to this.

Step S5: Search for corresponding data points

The purpose of this step is to simplify the second point cloud data. All data points q _j in the first simplified point cloud data The closer data point p _i is found in to establish the correspondence between the two data points p _i and q _j . According to the three-dimensional point cloud index structure of step S3, the query data point q _{j is} compared with the range of the node to determine which side child node to move during the search, and the kd tree is continuously tracked until the lowest leaf node.

After completing this step for the first time, the first simplified point cloud data Each data point p _i can be preliminarily corresponded to the second simplified point cloud data Each data point q _j . Please refer to FIG. 3, and it is explained here that steps S5 to S7 are the steps of the iteration operation. Each iteration operation can generate an optimized transformation matrix. Considering that it is the kth iteration operation, and Is the previous iteration operation, and after k-1 iteration operation returns to step S5, the first simplified point cloud data Each data point p _i can correspond to the transformation through the k-1th optimal transformation matrix Second simplified point cloud data Each data point q _j . When k = 1, the k-1st optimal transformation matrix is the initial transformation matrix .

Step S6: Establish an optimized transformation matrix

After the foregoing step S5, if the first simplified point cloud data Having a data point d _i, a second point cloud data simplification Has a corresponding data point s _i , that is, the data point s _i is the k-1 optimized transformation matrix searched in step S5 Second simplified point cloud data after conversion Data points, in step S6, each data point s _i becomes a data point after passing through a transformation matrix. , The data point Establish data points and a vector d _i, wherein, when the transformation matrix is a matrix converter Optimizer Tangent plane, the point information d _i and the data point There is a minimum distance between the tangent planes of. The minimum distance can be determined by the projection of the vector on the data point d _i tangent plane normal vector n _i . The general formula of the optimized transformation matrix can be expressed as follows:

among them:

In the above formula, Is the displacement matrix, Are rotation matrices, which are expressed as follows: ; ,among them:

tx, ty, and tz are the x-axis displacement, y-axis displacement, and z-axis displacement, respectively. This creation uses a linear solution approach, assuming that the angle between the three axes approaches 0, that is, the angle between the x axes , Y-axis angle , Z-axis angle ,then , , , , , , Then the rotation matrix become , Expressed as follows:

The transformation matrix become , Expressed as follows:

Substitute this transformation into the optimized transformation matrix To make the optimized transformation matrix become , Expressed as follows:

Then the ith term can be linearly expanded, which is expressed as follows:

So, the linear expansion of all i terms can form a linear system:

among them: ,among them:

then

That is, the above formula can be made The parameter vector x with the minimum value (including six conversion parameters such as α, β, γ, t _x , t _y , t _z ) is an optimized conversion parameter, which is expressed as α _opt , β _opt , γ _opt , t _xopt , t _yopt , t _zopt , which are represented as follows:

Among these, the optimized conversion parameters α _opt , β _opt , γ _opt , t _xopt , _tyopt , t _zopt can be obtained by using a pseudo inverse operation.

Finally, according to the optimized conversion parameters α _opt , β _opt , γ _opt , t _xopt , t _yopt , t _{zopt and} applied to the aforementioned rotation matrix R _k and displacement matrix T _k , the optimized conversion matrix is obtained. This is the updated optimized transformation matrix, which is expressed as follows:

Therefore, the second simplified point cloud data of this creation Optimized transformation matrix with updates After conversion becomes Can make the first simplified point cloud data And second simplified point cloud data after conversion There is a minimum distance between the corresponding tangent planes of the two data points.

Step S7: Convergence evaluation

The foregoing steps S5 to S7 are iterative processes. Step 7 evaluates whether the iterations meet a convergence condition. If the convergence conditions are met, the two simplified point cloud data are indicated. , The updated optimized transformation matrix is now available Fitting is completed; otherwise, if the convergence conditions are not met, the updated optimized transformation matrix is used Reply and execute steps S5 to S7 to update the optimized transformation matrix again for convergence evaluation until convergence.

It should be noted that whenever returning to step S5 again, the first simplified point cloud data Each data point uses the three-dimensional point cloud index structure of step S3 to search for the second simplified point cloud data transformed by the previous optimized transformation matrix (ie, the k-1th optimized transformation matrix). Corresponding data points; and in step S6, the transformation matrix M used is the previous optimization transformation matrix, that is, the k-1th optimization transformation matrix is used to generate a new update optimization again. Transformation matrix (ie, the k-th optimized transformation matrix), and then calculate the corresponding two pieces of data between the first simplified point cloud data and the second simplified point cloud data transformed by the k-th optimized transformation matrix The minimum distance between the points' tangent planes, and so on.

The determination of the convergence condition will be described later. The result produced in step S6 is the minimum distance, assuming After the kth iteration, the minimum distance between the two data points d _i , s _i has a maximum error value e _k , which is expressed as follows:

The convergence condition is judged by the maximum error value of the kth and k-1th times. The convergence condition of this creation is , Where t _e is a preset judgment value. If iterations fail to converge for many times, it may cause a bottleneck due to too many iterations. Therefore, this creation can set an upper limit of iterations. When the data processing device 20 When it is determined that the number of iterations reaches the upper limit number of iterations, the iteration is stopped, and the group with the smallest error in the estimated multiple sets of conversions is used as the optimal conversion matrix.

Step S8: Mark redundant points

Redundant points are the two simplified point cloud data , For the data points of repeated acquisition, please refer to FIG. 2C and FIG. 7. After overlapping, they should overlap to create an overlapping area 30. The first simplified point cloud data Boundary 31 falls into this second simplified point cloud data And the second simplified point cloud data Boundary 32 falls into the first simplified point cloud data In the range, the two simplified point cloud data , The pattern features have roughly overlapped. However, because the surface of the tooth has been digitally scanned, in fact, the data points in the overlap area of the two data may not have exactly the same coordinate values. In this creation, the minimum distance between the data points corresponding to the conditions is used to evaluate whether it is redundant. Remaining points:

Among them, t _R is a preset judgment value, and the data points d _i and s _i that meet the conditions of the above formula are used as the two simplified point cloud data. , Duplicate redundant data points. Among them, since the corresponding relationship between the two data points d _i and s _i has been established in step S5, the distance between the two data points d _i and s _i can be directly calculated based on the coordinate information for judging whether the distance is lower than the judgment The value t _R.

In summary, after detecting the redundant points in this creation, the redundant points can be excluded from the application, which can effectively reduce the data storage cost and further optimize the data transmission efficiency and data operation efficiency. In addition, this creation can also effectively reduce the amount of data and increase the speed of calculation through technical methods such as downsampling, three-dimensional index structure, and linear approximation.

10‧‧‧ Scanner

20‧‧‧Data Processing Device

30‧‧‧ overlapping area

31‧‧‧ border

32‧‧‧ border

P‧‧‧First point cloud data

‧‧‧The first simplified point cloud data

Q‧‧‧Second point cloud data

‧‧‧Second simplified point cloud data

RN‧‧‧root node

BB‧‧‧Bounding Box

SN‧‧‧node

LN‧‧‧ Leaf Node

40‧‧‧ repeat area

41‧‧‧First point cloud data

410‧‧‧ border

42‧‧‧Second point cloud data

420‧‧‧ border

43‧‧‧ third point cloud data

44‧‧‧ Fourth point cloud data

FIG. 1 is a block diagram of an embodiment of a system for implementing the creative detection method. FIG. 2A to FIG. 2C are schematic diagrams of bonding the first point cloud data and the second point cloud data to each other in this creative embodiment. FIG. 3 is a schematic flowchart of an embodiment of the creative detection method. Figure 4: Schematic of the first simplified point cloud data. Figure 5: Schematic of the second simplified point cloud data. 6A ~ 6F: Schematic diagrams of the three-dimensional point cloud index structure generated by this creation. FIG. 7 is a schematic diagram of attaching the simplified point cloud data of FIG. 4 and FIG. 5 to each other. 8A-8E are schematic diagrams of conventionally connecting and combining multiple point cloud data in series.

Claims (5)

A redundant point detection method for point cloud data bonding is performed by a data processing device. The data processing device is connected to a scanner to receive point cloud data. The method includes: receiving a first One point cloud data and one second point cloud data; downsampling the first point cloud data and the second point cloud data to become a first simplified point cloud data and a second simplified point cloud data, respectively; A three-dimensional point cloud index structure of the first simplified point cloud data for searching each data point in the second simplified point cloud data to correspond to each data point of the first simplified point cloud data; establishing an optimized transformation matrix , After the second simplified point cloud data is transformed by the optimized transformation matrix, there is a minimum between the tangent planes of two data points corresponding to the first simplified point cloud data and the second simplified point cloud data. Distance; judging whether the minimum distance is less than a judgment value; if yes, the two data points are redundant data points; if not, return to "establish the three-dimensional point cloud index structure of the first simplified point cloud data for Search Each step of each data point of the second data in the point cloud data corresponding to the first data point of the point cloud simplified "to iterate the calculation of optimum conversion matrix. The redundant point detection method for point cloud data bonding as described in claim 1, the optimized transformation matrix Represented as follows: Where k is the number of iterations; R _{k is a} rotation matrix; T _{k is a} displacement matrix; α _opt , β _opt , γ _opt , t _xopt , t _yopt , t _zopt : optimize conversion parameters. According to the redundant point detection method for point cloud data bonding described in claim 2, further determine the maximum error value of the two data points corresponding to the first simplified point cloud data and the second simplified point cloud data. Whether it meets a convergence condition; if not, recalculate the optimization transformation matrix until it meets the convergence condition; the convergence condition is expressed as follows: Where e _k : the maximum error value obtained according to the k-th best conversion matrix, , D _i is the data point of the first simplified point cloud data, and s _i is the corresponding data point of the second simplified point cloud data; e _k-1 : the maximum error obtained according to the k-1th best conversion matrix Value; t _e : judgment value. According to the redundant point detection method for point cloud data bonding described in claim 3, in the step of determining whether the minimum distance is less than the judgment value, the following formula is used to determine whether the minimum distance is a redundant point: Among them, t _R is a judgment value. According to the redundant point detection method for point cloud data fitting described in claim 4, a linear solution approximation method is used to establish the optimized transformation matrix.