JP2002521699A - Calibration method and device for non-contact distance sensor - Google Patents

Abstract

(57) [Summary] A method of calibrating a non-contact type distance sensor includes a step of obtaining a distance image of an inspection object using a distance sensor. The distance image is registered together with the reference data of the object. Thereby, registration data is obtained. A reference image can be obtained from computer-aided drawing (CAD) data. A normal deviation between the registered data and the reference data is calculated. Noise is removed from the normal deviation. A plurality of bias vectors and a covariance matrix are estimated based on normal deviations stored in a memory as a look-up table with general correction factors for the sensor. Next, the object image obtained by the sensor is corrected based on the look-up table.

Description

DETAILED DESCRIPTION OF THE INVENTION

REFERENCE TO RELATED APPLICATIONS This application is a provisional application 60/094, entitled "Precise and Accurate Surface Reconstruction Using Deductive Models and Concentration Data," filed Jul. 28, 1998, which is incorporated by reference.
, 444 and provisional application 60/094442, entitled "Searching for Shape Deformation of Wings or Generalized Cylinders," filed July 28, 1998.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates generally to digital image processing technology, and more particularly to a method for calibrating a non-contact distance sensor used to obtain a digital image of an object for the purpose of inspecting an object for defects. And equipment.

Currently, in order to detect a defect in a manufactured object such as a blade (for example, a turbine blade), an automatic inspection is performed using digital data and an image of the object. Wings with cast vanes, such as those used in aircraft engines, are inspected for deformation parameters such as platform orientation, profile cross-section, stacking axis curvature, thickness and chord length at a given cross-section. One way to determine digital data representing these parameters is to use a coordinate measuring instrument (commonly known as "CMM" = coordinate measurement machines). In the CMM, a probe in contact with the object surface is translated / rotated with respect to a sample point on the object surface.

As another means for obtaining a digital notation of these parameters, there is a method using a non-contact type distance sensor. A non-contact full-surface sensor scans the outer surface of an opaque object using a laser or white light. These sensors can scan parts of them quickly to obtain large amounts of data. Examples of non-contact sensors include sensors based on laser diffraction gratings and stereo triangulation, and based on a single laser scan and rotation of a portion thereof. Other non-contact sensors are:
It is based on phase moiré and white light.

[0004] However, raw data points obtained by non-contact sensors are very inaccurate. Therefore, if the raw data is used to reconstruct the surface of this object for the purpose of detecting the deformation or defect of the object, the noisy normal deviation ( normal deviations). Since all-field sensors are relatively new, little effort is needed to calibrate them or find their properties to improve accuracy, but much of the effort is wasted.

SUMMARY OF THE INVENTION [0005] A calibration method for a non-contact distance sensor according to a preferred embodiment of the present invention includes a step of scanning an object to be inspected using a distance sensor to obtain a distance image of the object.
This distance image is registered in association with the reference image of the object, and thus registered data is obtained. Then, the normal deviation between the registered data and the reference image is calculated. Remove noise from its normal deviation. A plurality of bias vectors and a covariance matrix are estimated and used to generate a look-up table with geometric correction factors. The distance image of the object is corrected based on the look-up table.

DETAILED DESCRIPTION OF EMBODIMENTS In the present specification, calibration of a sensor refers to identifying a bias error included in a measurement value obtained by the sensor, and correcting an error of a value obtained thereafter. It is defined as a process that enables correction.

Generally, in the method of the present invention, the sensor is calibrated in the following successive steps. Using the sensor, a range image of the object to be inspected is obtained. The sensor scans the object, and at that time, translates and rotates the object according to the work space. This results in range data points. In the present embodiment, these data points are arranged in a data set format, and this data set is called a scanned image. The data set of the scanned image is registered in association with one or more sets of reference data. The reference data forms a geometric ideal model of the object, and is referred to as a reference image in the present embodiment. The scan image and the reference image are registered in association with each other so as to be expressed in the same coordinate system. Next, a normal deviation of the scan image from the registered image is calculated. A random noise describing the accuracy and a bias error (= bias error, hereinafter referred to as “bias error”) describing the accuracy are estimated from the normal deviation and stored in a look-up table. This lookup table is used for three-dimensional geometric correction and multi-view integration of a plurality of images obtained by the sensor.

One embodiment of the method according to the invention is shown in FIG. Step 4 of method 40
In 2, a scanned image of the inspection target object is obtained by moving the object on the moving table in parallel and rotating the object on the rotating table in a plurality of directions related to the three-dimensional working space of the object. This process is accomplished using typical scanning and digital imaging systems and techniques.

In step 44, a reference image having data points corresponding to an object having an ideal shape and the scanned image of an actual object are registered. One way to obtain the reference image required when registering the scanned image with the reference image is to polish and flatten a flat object. Another method is to measure the object using a more accurate sensor such as a CMM and store the measurement result as reference data. Another method is to obtain a reference image using computer-aided drawing (CAD) data or other engineering techniques or design specifications.

As shown in FIG. 4, in one embodiment of the present invention, a robust-closest-patch (RCP) algorithm is used for the registration process. The RCP algorithm was filed on April 30, 1999, is assigned to the assignee of the present invention, and is a co-pending application entitled "Image Registration Method and Apparatus," which is incorporated herein by reference.
(SN 09 / 303,241). In the RCP algorithm, the initial rigid body posture (init
Given a rial rigid pose), each scanned image data point is represented by step 62 in FIG.
As shown in step 64, a match is made with neighboring data points on the reference image.
In other words, in the RCP algorithm, a match between the reference image data point and the scan image data is determined based on the current posture estimation, and the quality of the posture estimation is enhanced based on the matching judgment.

In one embodiment of the present invention, the pose problem is solved by singular value decomposition (SVD) of a linear system represented by six parameters. In one embodiment of the invention, rather than using quaternions or orthonormal matrices, a linearly symmetric equation of the rotation constraint using the Rodrigues formula is used, which is the pose In one embodiment of the invention, a rigid M pose estimator is used to estimate rigid body pose parameters and the standard deviation of the error, which are included in the data. Ensure robustness against large errors.

A three-dimensional rigid transformation for optimally aligning the match determination results is determined in step 66 and applied to step 68 in FIG. The above process uses the most recently obtained estimated attitude,
Repeat until convergence is obtained. The registration step 44 in FIG. 1 solves the problem of obtaining the direction of translation and the axis of rotation necessary to represent the reference image by the coordinates of the scanned image.

In step 46 of FIG. 1, the deviation of the normal of the scanned image data point from the normal of the registered image data point is obtained. In one embodiment according to the present invention, by using an approximate normal distance between a model data point (referred to as a patch) and a data surface point, a local normal of a plane is obtained from noise-containing data. It avoids the need to estimate curvature. To perform step 46, one embodiment of the present invention calculates the raw error. And the raw error, relative to the direction D _k given point [q _ij] (
Hereinafter, the sentence of the translation, for convenience, represents the vector x in [x], the raw error e _ij along the normal [n _k] in represents denoted by the up arrow) to x in the formula is ,

[0014]

(Equation 1)

The following relationship is satisfied.

[0016] Random noise is removed from the deviation related to the normal by a Gaussian filter or an average filter. In the embodiment using the averaging filter of the present invention, the projection _bijk of the bias vector [ _bij ] at each point is

[0017]

(Equation 2)

Where the projection b _ijk is

[0019]

(Equation 3)

This is a case where an average filter is used. Here, M _i is the small area surrounding the data points.

The averaging filter described by equation (3) above is used to determine the projection _bijk of the bias vector [ _bij ] at each data point, as shown in step 50.

FIG. 2 shows a portion of the step 50 of estimating the bias vector. That is, a bias vector 56 is defined for each of a series of surfaces 58 along the z-axis from a surface 59 closest to the sensor to a surface farthest from the sensor. A three-dimensional geometric correction table is formed by obtaining projections of bias vectors at a plurality of lattice points p provided at regular (equal) intervals in the three-dimensional space of the object. To do this, the projection of the bias vector at each point q _ij is sampled (d
own-sampled) and interpolated. In one embodiment of the present invention, Gaussian interpolation is used to interpolate the projection of the bias vector.

Returning to FIG. 1, in step 50, the bias vector is estimated as a function of the location and orientation of the surface based on projections of the bias vector along multiple normals. Note that all of these projections come from the same grid point. The plurality of steps described above include, for each direction D _k , the normal [n _k ] of the plane at each grid point p
And the projection of the bias vector [b _pk ].

In one embodiment of the present invention, a three-dimensional working volume (working volume)
The following relational expression system for each lattice point p in e) will be solved.

[0025]

(Equation 4)

Here, N is a matrix having a size of k × 3 composed of a plurality of normal vectors,

[0027]

(Equation 5)

And [b_p] Is the bias vector at this grid point, and [v _p ]

[0029]

(Equation 6)

Thus, it is a k × 1 vector composed of a plurality of bias projections from the grid point p.

In one embodiment of the invention, if the rank of the matrix N is 3, the above relational system is solved by applying a representative singular value decomposition (SVD) algorithm. As a result, a smooth field bias vector is obtained.

FIG. 3 shows three components of X, Y, and Z, which are bias vector components of a three-dimensional lattice.

In step 52 of FIG. 1, a covariance matrix is estimated. This covariance matrix models the random noise and gives the accuracy of the distance sensor, and is based on the data points corrected for the bias described above. The noise model can be simplified by obtaining and analyzing main eigenvalues and eigenvectors of the covariance matrix. To accomplish this, the variance in the normal direction [ _nk ] at grid point p is

[0034]

(Equation 7)

Is obtained by the following relational expression. Here, e _jk is the error calculated in the point j, M _{p k} is the number of points taken over a small area surrounding the lattice point p.

[0036] In one embodiment of the present invention, the relationship between the s ² _pk and covariance matrix S _p at point p is

[0037]

(Equation 8)

Is described by here,

[0039]

(Equation 9)

Is as follows.

Further, assuming that [n _k ] = (n _kz , n _ky , n _kx ), the vector [n ′ _k ]

[0042]

(Equation 10)

And [s _p ] is

[0044]

[Equation 11]

Then, in order to explain one of the techniques for estimating the six coefficients of the covariance matrix _Sp , Equation 3 can be rewritten as follows using the vector [n ′ _k ].

[0046]

(Equation 12)

[0047] Then, s _p that contains each element section of the covariance matrix S _p can be estimated by forming the following matrix Kx6.

[0048]

(Equation 13)

Because the kx1 vector is

[0050]

[Equation 14]

Then, the constraint on s is:

[0052]

(Equation 15)

Is obtained by

In one embodiment according to the present invention, the above equation is solved by a representative least squares method. In another embodiment of the present invention, as long as the rank of [t _p ] is at least 6, the SV
A method using D or a pseudo inverse function is used.

In step 54, the calibration made by the process according to the present invention is verified using an independent data source for known features of the object. For example, horny artifacts can be used to independently verify sensor calibration. Further, using a plurality of spherical artifacts having different radii, it is possible to clarify the effects of the curvature and the noise model of the sensor according to three-dimensional geometric correction.

Those skilled in the art will appreciate that method 40 can be used to perform a calibration check and recalibration process on-site to extend the operating life of the sensor. In one embodiment according to the present invention, the calibration check is performed at least every 8 hours, and the re-calibration process is performed at least weekly. The present invention can be realized by a computer process and a device that executes the process. The invention can also be implemented by instructions including computer program code embodied in a tangible medium, such as a floppy disk, CD-ROM, hard drive, or other computer-readable recording medium. Here, when the computer program code is loaded and executed on the computer, the computer becomes an apparatus for implementing the present invention. Also, the present invention may be implemented, for example, by a computer program code stored in a recording medium, loaded and executed on a computer, or transferred via a transmission medium such as electric wiring, twisted cable, optical fiber, electromagnetic radiation, or the like. Can be embodied. Here, when the computer program code is loaded and executed on the computer, the computer becomes an apparatus for executing the present invention. When executed on a general-purpose microprocessor, the computer program code segments constitute a microprocessor that creates specific logic circuits.

Although the preferred embodiment has been shown and described, various modifications or substitutions may be made within the spirit and scope of the invention. Thus, while the invention has been described by way of example, it should be understood that it is not so limited.

[Brief description of the drawings]

Here, reference is made to the drawings. Note that the same elements are denoted by the same reference numerals throughout the drawings.

FIG. 1 is a flowchart of a method of calibrating a non-contact type distance sensor according to the present invention.

FIG. 2 is a perspective view of a plurality of projected images of a bias vector located along a Z axis.

FIG. 3 is a perspective view of the x, y, and z components of the bias vector of FIG. 2;

FIG. 4 is a flowchart according to a method of registering a distance image of an object together with a reference image of the object.

──────────────────────────────────────────────────続 き Continuation of front page (81) Designated country EP (AT, BE, CH, CY, DE, DK, ES, FI, FR, GB, GR, IE, IT, LU, MC, NL, PT, SE (72) Inventor Stuart, Charles Vernon United States, New York, 12065, Clifton Park, Jason Lane, 4 (72) Inventor Babuna, Kisher United States, New York, 12180, Troy, 11T Street, 79 (72) Inventor Abdesmed, Rila, United States, New York, 12309, Schenectady, Rosa Road, 2150 Yuaru, 53 F-term (reference) 2F065 AA06 BB05 FF04 JJ02 JJ25 KK01 MM02 PP12 QQ31 RR08 5B057 AA01 BA02 CA13 CA16 CB13 CB16 CE02 CE05 CH07 CH08 CH09 DA07 DB03 DC33

Claims (12)

[Claims] Acquiring a scan image of an object to be inspected using a sensor; registering the scan image of the object with a reference image of the object to obtain registration data; Calculating a deviation between the reference data and the reference data; estimating a random noise and a bias error based on the deviation; and correcting a scan image of the object according to the estimated random noise and the bias error. And a step of calibrating the non-contact type distance sensor. 2. The method of claim 1 including storing the random noise and bias error estimates in a memory. 3. The method of claim 1, wherein estimating random noise and bias error comprises estimating a bias vector and estimating a covariance matrix based on the deviation. the method of. 4. The method of claim 1, wherein the reference image comprises computer-aided drawing data for the object. 5. The method according to claim 1, wherein the registration step includes a robust-closest-patch matching (
The method according to claim 1, wherein the method is performed by an RCP method. 6. The method of claim 1, wherein calculating the deviation is performed by calculating the deviation along a surface normal. 7. The method of claim 1, wherein estimating random noise comprises removing random noise by applying Gaussian smoothing to normal deviations. 8. A storage medium encoded with machine readable computer program code having instructions for causing a computer to perform a predetermined method for reconstructing an object, the predetermined method comprising using a sensor. Obtaining a scan image of the object to be inspected by registering the scan image of the object together with a reference image of the object to obtain registration data; and obtaining a deviation between the registration data and the reference data. Calculating, estimating random noise and bias error based on the deviation, and correcting the scanned image of the object according to the estimated random noise and bias error. Storage media. 9. The computer-aided drawing (CAD) of the object is provided.
9. The storage medium according to claim 8, wherein the storage medium is an expression. 10. The storage medium of claim 8, wherein the step of registering data points is performed based on a robust-closest-patch matching technique. 11. The storage medium of claim 8, wherein the step of calculating a deviation is performed by calculating a deviation along a surface normal. 12. The storage medium according to claim 8, wherein the step of removing random noise includes a step of applying Gaussian smoothing processing to a deviation of a normal line.