Methods for improving the accuracy of dimensioning-system measurements

Abstract

Methods to improve the accuracy of non-contact measurements of an object's dimensions using a dimensioning system are disclosed. The methods include a method for creating a mathematical model (i.e., error model) based on an observed correlation between errors in an estimated dimension and the characteristics of the measurement used to obtain the estimated dimension. These error models may be created for various dimensions and stored for future use. The methods also include a method for using the stored error models to reduce the error associated with a particular dimensioning-system measurement. Here an error model is used to create an estimated error. The estimated error is then removed from the estimate of the dimension to produce a final estimate of the dimension that is more accurate.

Description

FIELD OF THE INVENTION

The present invention relates to systems for determining an object's physical dimensions (i.e., dimensioning systems) and, more specifically, to methods for creating and using error models to improve the accuracy of dimensioning system measurements.

BACKGROUND

Determining an item's dimensions is often necessary as part of a logistics process (e.g., shipping, storage, etc.). Physically measuring objects, however, is time consuming and may not result in accurate measurements. For example, in addition to human error, measurement errors may result when measuring irregularly shaped objects or when combining multiple objects into a single measurement. As a result, non-contact dimensioning systems have been developed to automate, or assist with, this measurement. These dimensioning systems sense an object's shape/size in three-dimensions (3D) and then use this 3D data to compute an estimate of an object's dimensions (e.g., length, width, height, etc.).

Accurate dimensioning is highly valued. For example, regulatory certification often demands highly accurate measurements when dimensioning is used for commercial transactions (e.g., determining shipping costs). Unfortunately, there are errors in the dimensions estimated by dimensioning system. One way to reduce these errors is to (i) constrain the size/shape of measured objects and (ii) place strict requirements on the measurement setup. These constraints, however, limit the flexibility of the dimensioning system and the speed at which a measurement may be taken. Therefore, a need exists for methods to reduce the errors associated with estimated dimensions returned from a dimensioning system.

SUMMARY

Accordingly, in one aspect, the present invention embraces a method for removing errors from a dimensioning-system measurement. First, a dimensioning system is provided to perform a dimensioning-system measurement of an object in an environment. The dimensioning-system measurement results in three-dimensional (3D) data corresponding to the object/environment. Next, a particular dimension to be estimated is selected. Then, using the 3D data, an intermediate estimate of the particular dimension is created. In addition, values for predictor variables, pertaining to the aspects of the dimensioning-system measurement, are obtained. To remove errors from the intermediate estimate the method first estimates and then removes the errors.

To create an error estimate for the particular dimension, the method retrieves the particular dimension's error model, which relates the one or more predictor variables to an estimated error, from a library of error models. Then, the error estimate for the particular dimension is computed using the error model and the values obtained for the one or more predictor variables.

To remove the errors from the dimensioning-system measurement, the method subtracts the error estimate from the intermediate estimate of the particular dimension to obtain a final estimate for the particular dimension.

In an exemplary embodiment of the method, the predictor variables include variables that describe intrinsic properties of the dimensioning system, such as the dimensioning-system's acquisition parameters.

In another exemplary embodiment of the method, the predictor variables include variables that describe intrinsic properties of the object, such as the object's size, shape, and/or appearance.

In another exemplary embodiment of the method, the predictor variables include variables that describe intrinsic properties of the environment, such as the light level of the environment.

In another exemplary embodiment of the method, the predictor variable include variables that describe extrinsic aspects of the dimensioning-system measurement, such as the physical relationships between (i) the dimensioning system and the object, (ii) the dimensioning system and the environment, and/or (iii) the object and the environment.

In another exemplary embodiment of the method, the error model includes a linear equation relating the error estimate to the one or more predictor variables.

In another exemplary embodiment of the method, the error model includes a non-linear equation relating the error estimate to the one or more predictor variables.

In another exemplary embodiment of the method, (i) the 3D data includes a minimum-volume-bounding box (MVBB), and (ii) the particular dimension is the length, width, or height of the MVBB.

In another exemplary embodiment of the method, (i) the 3D data includes a minimum-volume-bounding box (MVBB) having a length, a width, and a height, and (ii) the method estimates and removes errors for each particular dimension of the MVBB (i.e., the length, the width, and the height).

In another exemplary embodiment of the method, the library of error models includes classes of error models; wherein each class corresponds to (i) a particular operating environment and/or (ii) a feature-set corresponding to the object. In this case, the method's step of retrieving an error model from the library includes selecting a class of error models from the library and retrieving an error model for a particular dimension from the selected class of error models.

In another aspect, the present invention embraces a method for creating an error model for a measured feature. First, a dimensioning system and a calibration object, having a feature with a known size, are provided. Next, measurements of the feature are gathered using the dimensioning system. Errors for the measurements (i.e., measured errors) are then calculated by comparing each measurement to the known size. In addition, predictor variables, which describe aspects of the measurements, are defined, and a mathematical model relating the predictor variables to an estimated error for the measurements is derived. The mathematical model includes predictor variables and predictor coefficients, wherein each predictor variable corresponds to a particular predictor coefficient. Next, by adjusting the predictor coefficients, the mathematical model is fit to the measured errors. The mathematical model is then refined to create the error model for the measured feature.

In an exemplary embodiment of the method, the error model is stored for future use.

In another exemplary embodiment of the method, the mathematical model is a linear combination of predictor variables and predictor coefficients or a nonlinear equation using predictor variables.

In another exemplary embodiment of the method, the predictor variables describe aspects of the measurements including intrinsic properties of the dimensioning system, the object, and/or the environment.

In another exemplary embodiment of the method, the predictor variables describe aspects of the measurements including physical relationships between (i) the dimensioning system and the object, (ii) the dimensioning system and the environment, and/or (iii) the object and the environment.

In another exemplary embodiment of the method, refining the mathematical model includes removing insignificant predictor variables and their corresponding predictor coefficients.

In another exemplary embodiment of the method, fitting the mathematical model to the errors includes a linear regression.

In another exemplary embodiment of the method, refining the mathematical model includes (i) obtaining residuals by comparing the estimated errors to the errors; (ii) creating a histogram of the residuals; and (iii) rejecting or accepting the mathematical model based on the normality of the histogram.

In another exemplary embodiment, the feature is the object's length, width, or height.

The foregoing illustrative summary, as well as other exemplary objectives and/or advantages of the invention, and the manner in which the same are accomplished, are further explained within the following detailed description and its accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 schematically depicts and exemplary dimensioning system according to an embodiment of the present invention.

FIG. 2 graphically illustrates the principle of sensing three dimensions using a spatially offset pattern projector and range camera according to an embodiment of the present invention.

FIG. 3 graphically illustrates exemplary predictor variables for a dimensioning system measurement according to an embodiment of the present invention.

FIG. 4 graphically depicts a flow diagram illustrating a method for creating an error model for a feature measured by a dimensioning system according to an embodiment of the present invention.

FIG. 5 graphically depicts a flow diagram illustrating a method for removing errors from a dimensioning-system measurement according to an embodiment of the present invention.

FIG. 6 graphically depicts the retrieval of an error model from a library of error models according to an embodiment of the present invention.

DETAILED DESCRIPTION

The present invention embraces improving the accuracy of dimensioning-system measurements through the use of mathematical models to estimate error (i.e., error models). The error models are created and then used to create an error estimate associated with a particular dimension/measurement. This error estimate may then be removed from the dimensioning system's estimate in order to improve the accuracy of the measurement. Some advantages of using error models to correct measurement errors in dimensioning are (i) improved measurement accuracy, (ii) improved measurement precision (i.e., repeatability), (iii) added flexibility (e.g., measuring a wider variety of objects), and (iv) easier/faster measurement acquisition (e.g., setup).

In general, dimensioning systems sense an object to gather 3D data corresponding to the object's shape/size, and then use this 3D data to compute the object's dimensions. In some cases, the 3D data is used to create a minimum bounding box (MVBB), which is a computer model of a box that surrounds the object (e.g., an irregularly shaped object) or a collection of objects (e.g., multiple boxes on a pallet). In these cases, the dimensioning system may return the dimensions of the MVBB.

A variety of techniques may be used to actively sense an object (e.g., structured-light, ultrasound, x-ray, etc.) and create 3D data (e.g., time-of-flight, triangulation, etc.). All of these techniques are within the scope of the present invention, however one exemplary embodiment (i.e., the triangulation of a structured-light pattern) will be described in relation to the disclosed methods.

The exemplary dimensioning system senses an object by projecting a light pattern (i.e., pattern) into a field-of-view. Objects within the field-of-view will distort the appearance of the reflected light-pattern. The dimensioning system captures an image of the reflected light-pattern and analyzes the pattern distortions in the captured image to compute the 3D data necessary for estimating the object's dimensions.

A block diagram of the dimensioning system is shown in FIG. 1. The dimensioning system 10 includes a pattern projector 1 that is configured to project a light (e.g., infrared light) pattern into a field-of-view 2. The light pattern typically comprises points of light arranged in a pattern (i.e., point cloud). The points of light may be (i) sized identically or differently and (ii) may be arranged in some order or pseudo-randomly. The pattern projector may create the light pattern using a light source (e.g., laser, LED, etc.), a pattern creator (e.g., a mask, a diffractive optical element, etc.), and one or more lenses.

The dimensioning system 10 also includes a range camera 3 configured to capture an image of the projected light pattern that is reflected from the range camera's field-of-view 4. The field-of-view of the range camera 4 and the field-of-view of the pattern projector 2 should overlap but may not necessarily have identical shapes/sizes. The range camera 3 includes one or more lenses to form a real image of the field-of-view 4 onto an image sensor. Light filtering (e.g., infrared filter) may be also be used to help detect the reflected pattern by removing stray light and/or ambient light. An image sensor (e.g., CMOS sensor, CCD sensor, etc.) is used to create a digital image of the light pattern. The range camera may also include the necessary processing (e.g. DSP, FPGA, ASIC, etc.) to obtain 3D data from the light-pattern image.

As shown in FIG. 2, the pattern projector 1 and the range camera 3 are spatially offset (e.g., stereoscopically arranged). The spatial offset 8 (i.e., the baseline) allows for changes in the range 5 of an object 6 to be detected as an image offset 7 on the range camera's image sensor. The spatial offset 8 may be adjusted to change the image offset 7 to change the resolution at which range differences 5 may be detected. In this way, image offsets in the point-cloud pattern may be converted into 3D data for objects within the dimensioning system's field-of-view.

Accurate dimensioning requires that (i) the sensing obtains sufficient, high-quality 3D data and (ii) that the dimensioning system's algorithms can convert the 3D data into precise estimates of the object's dimensions. This accuracy may be affected by many different variables. The variables (i.e., predictor variables) may be classified into two categories: intrinsic and extrinsic. FIG. 3 graphically illustrates exemplary predictor variables for a dimensioning system measurement according to an embodiment of the present invention.

Intrinsic variables describe properties related to the essential nature, constitution, or operation of a particular element in the dimensioning system measurement. A particular element may be the dimensioning system, the object, or the environment in which the object resides (i.e., the environment). Furthermore, a particular element may include subsystems within the dimension system such as the range camera 3 or the pattern projector 1.

Intrinsic variables related to the object may describe the object's shape or appearance. For example, the object may be classified by shape (e.g., a box, a cylinder, etc.) and the class of shape may be an intrinsic variable 11. The object may have sides that are not flat (e.g., curved) and the mean curvature of the object may be an intrinsic variable 12. Other object intrinsic variables include (but are not limited to) estimated height 13, estimated length 14, estimated width 15, average color (e.g., red, green, blue), and reflectivity.

Intrinsic variables related to the dimensioning system may include (but are not limited to) the base line (i.e., the spatial offset between the range camera and the pattern projector).

Intrinsic variables related to the range camera may include (but are not limited to) the focal length (i.e., of the range camera's lens), lens distortion, the optical center of the image (i.e., where the range camera's optical axis intersects with the range camera's image sensor), the orientation of the object on the image sensor, the range-camera's image height/width, and the maximum number of pattern points detected per frame.

Intrinsic variables related to the pattern projector may include (but are not limited to) the pattern density, the projected divergence angle, and the pattern type.

Intrinsic variables related to the environment may include (but are not limited to) the environment's light level and/or properties of the ground plane. The ground plane (i.e., ground) is the surface in the environment on which the object rests during the measurement. The ground typically fills much of the dimensioning system's field of view and serves as a baseline from which certain dimensions may be obtained. For example, a mathematical projection of an object surface to the ground may help determine one or more dimensions. As a result intrinsic variables related to the ground are typically defined and may include (but are not limited to) the ground's reflectivity and the area of the ground.

Extrinsic variables external factors affecting the dimensioning-system measurement. For example, extrinsic variables may describe the physical relationships between (i) the dimensioning system and the object, (ii) the dimensioning system and the environment, or (iii) the object and the environment. Extrinsic variables may also describe how an object intersects with the pattern projected by the dimensioning system. In addition, extrinsic variables may describe the position (e.g., pitch, roll, height) of the dimensioning system with respect to the ground 16. As shown in FIG. 3, other extrinsic variables include (but are not limited to) the dimensioning system's height 17, the object's center of gravity (i.e., COG) 18, the minimum distance between the dimensioning system and the object 19, and the maximum distance between the dimensioning system and the object 20.

Arbitrarily shaped objects (e.g., an object with a radius or a curvature) may be measured with a dimensioning system. While these measurements may include dimensions for curved or irregular surfaces (e.g., radius, curvature, etc.), a typical measurement includes estimating three dimensions (i.e., length, width, height) of computer-generated box that surrounds the object (i.e., the MVBB). For example, when measuring a box (e.g., a package), the edges of the MVBB coincide with the edges of the box.

The accuracy for each estimated dimension is determined by the amount of error associated with each dimension's estimate. Further, different errors may be associated with each estimated dimension. For example, there may be one error associated with the estimated length, another error associated with the estimated width, and still another error associated with the estimated height. The difference in errors may result from how each are estimated. For example, a height dimension may be estimated using a height-estimation algorithm, while a width dimension may be estimate using a width-estimation algorithm. Further, each algorithm may use different portions of the 3D data for its estimate. In many cases, however, an estimate for a particular dimension (e.g., obtained by a particular dimensioning system using a particular measurement setup) has roughly the same associated error from measurement to measurement.

Certain predictor variables, such as those described previously, may correlate well with the error associated with a measurement of a dimension. Understanding this correlation can help to accurately predict (i.e., estimate) the error associated future measurements of the dimension. This understanding is expressed as a mathematical equation (i.e., error model) that relates one or more predictor variables to an estimated error. Thus an error model may be created and then used to remove (or reduce) the error associated with a dimensioning-system measurement. Further, since different dimensions may have distinct error models, a library of error models may be created and stored in memory for future use.

A flow diagram illustrating a method for creating an error model is shown in FIG. 4. A dimensioning system is setup to measure a calibration object 25. The calibration object has a feature (or features) with a known size (e.g., a box with a known length, width, and height). The dimensioning system then gathers repeated measurements of one of the calibration object's features (e.g., the height of the box) 30. It should be noted that no changes are made to the dimensioning system setup between measurements. Each estimate of the feature (e.g., each height measurement) is then compared to the known value of the feature's size 32 in order to create multiple samples of the error (e.g., multiple samples of the error associated with measuring height).

Predictor variables used to describe aspects (e.g., intrinsic properties, physical relationships, etc.) of the measurements are defined 33. The predictor variables are used to derive a mathematical equation (e.g., mathematical model) for the estimated error 34. The mathematical model includes predictor variables and predictor variable coefficients that are assigned to each predictor variable. The mathematical model may be a linear combination of predictor variables and predictor coefficients (e.g., see FIG. 6) or may be a nonlinear equation using predictor variables.

The mathematical model is then fit to the multiple samples of the error associated with the measurements of the feature 37. Here, various fitting algorithms, such as linear regression, may be used. In addition, the fitting may require multiple iterations and refinement.

The linear regression algorithm adjusts the predictor coefficients so that the error model best matches the samples of the observed error. Here, the value of each adjusted predictor coefficient corresponds to the significance of that predictor variable's impact on the error estimate. The fitting may result with some predictor coefficients adjusted to a high absolute value and some predictor coefficients adjusted to an approximately zero value.

The linear regression algorithm may also return information regarding the error model. For example, a standard error (SE) for each coefficient may be returned. The SE helps to determine the precision of the coefficients. In addition, a p-value for each coefficient may be returned. The p-value helps to determine if the results are statistically significant.

After fitting, the error model may be refined 38 using various methods. One method includes analyzing the predictor coefficients and the information returned by the fitting algorithm. For example, insignificant predictor coefficients/variables may be removed from the error model. Another method for refinement includes comparing the estimated error (obtained using the mathematical model) to the measured errors (obtained in the multiple measurements). The result of this comparison includes a set of residual errors (i.e., residuals). A histogram of the residuals may then be created and analyzed. For example, the normality of the histogram (i.e., the correlation to a normal distribution) may determine if the error model is acceptable for use.

The final result of the one or more iterations of fitting/refining 38 is an error model 39. The error model 39 may be stored in a computer readable medium and retrieved later by a processor for computing the error associated with future measurements of the feature (e.g., height).

A flow diagram illustrating a method for using an error model to reduce the errors associated with a dimensioning measurement is shown in FIG. 5. The method begins with setting up the dimensioning system measurement 40. This setup includes positioning an object 41 in an environment (e.g., on a ground plane) 42 and then arranging the dimensioning system 26 and/or object so that the object is within the dimensioning system's field of view. The object is then sensed by the dimensioning system to obtain 3D data describing the object (and the environment) 45. A particular dimension is then selected for estimation (i.e., measurement) 46.

Here the method splits into two branches. In one branch, the 3D data is used to create an intermediate estimate of the size (e.g., length, width, height, etc.) of the selected dimension 47. In the other branch, an error model for the selected dimension is retrieved from a library of error models 48. The values for the predictor variables that are used in the retrieved error model are obtained (e.g., from the 3D data and/or from intrinsic/extrinsic information regarding the measurement) 49. Then, using the error model and the values for the predictor variables, an estimate of the error associated with the measurement is computed 50.

The error associated with the intermediate estimate of the size of the selected dimension is then reduced or removed by subtracting the error estimate from the intermediate estimate 51. What results is a final estimate of the selected dimension 52.

It should be noted that the 3D data gathered 45 allows for the measurement (i.e., estimation) of a plurality of dimensions, and while the method illustrated in FIG. 5 shows the process for obtain a final estimate of one of these dimensions, this method may be iterated in order to measure the other dimensions of the object.

As mentioned previously, the library of error models may store error models for each dimension. To expand the usability of this method, the library may also store collections (i.e., classes) of error models to suit various operating conditions. For example, a class of error models may be created to accommodate a particular operating environment and/or a feature set of an object (e.g., boxes, cylinders, etc.). This approach may improve the estimation of errors and allow for more flexibility.

FIG. 6 graphically depicts the retrieval of an error model 62 from a library of error models 61. The retrieval process begins with the selection of a particular class of error models 60. Then, based on the dimension selected for measurement, a particular error model 62 may be selected from the class of error models 60.

FIG. 6 also illustrates an exemplary error model. In this example, the error estimate for a particular dimension is equal to a linear combination of predictor variables and predictor coefficients. The predictor coefficients are derived during the model creation process (e.g., see FIG. 4) and are stored as part of the model. The predictor variable values are obtained during the measurement (e.g., see FIG. 5). The values for the predictor variables are not stored with the model and typically change from measurement to measurement.

In the specification and/or figures, typical embodiments of the invention have been disclosed. The present invention is not limited to such exemplary embodiments. The use of the term “and/or” includes any and all combinations of one or more of the associated listed items. The figures are schematic representations and so are not necessarily drawn to scale. Unless otherwise noted, specific terms have been used in a generic and descriptive sense and not for purposes of limitation.

Claims

1. A method for creating an error model for a feature measured by a dimensioning system, the method comprising: providing a dimensioning system and a calibration object, the calibration object having a feature with a known size;gathering, using the dimensioning system, measurements of the feature;

2. The method according to claim 1, further comprising storing the error model in memory for future use.

3. The method according to claim 1, wherein the mathematical model is a linear combination of predictor variables and predictor coefficients.

4. The method according to claim 1, wherein the aspects of the measurements comprise intrinsic properties of the dimensioning system, the object, and/or the object's environment.

5. The method according to claim 1, wherein the aspects of the measurements comprise (i) a physical relationship between the dimensioning system and the object, (ii) physical relationship between the dimensioning system and the object's environment, and/or (iii) a physical relationship between the object and the object's environment.

6. The method according to claim 1, wherein refining the mathematical model comprises removing insignificant predictor variables and their corresponding predictor coefficients.

7. The method according to claim 1, wherein the fitting the mathematical model comprises a linear regression.

8. The method according to claim 7, wherein refining the mathematical model comprises: obtaining residuals by comparing the estimated error to the errors;creating a histogram of the residuals; andrejecting or accepting the mathematical model based on the normality of the histogram.

9. The method according to claim 1, wherein the feature is the object's length, width, or height.

10. The method according to claim 1, wherein the mathematical model is a nonlinear equation using predictor variables.