DE102005043070B4 - Method for high-precision three-dimensional measurement and / or reconstruction of objects with the aid of digital image recordings, for example for image evaluation of traffic routes - Google Patents

Abstract

Method for high-precision three-dimensional measurement and / or reconstruction of objects with the aid of digital image recordings, for example for image evaluation of traffic routes, where at least two digital images of the object are inserted into the object view with a camera to be calibrated once prior to the first use with respect to the parameters Reference element of known geometry from different camera positions are recorded, in which in each case the reference element is determined in the evaluation of the recorded images of the object for its three-dimensional measurement or reconstruction, in which in each of the images taken or the pixels of the spatial points to be evaluated, and then taking into account the determined three-dimensional camera positions of the recorded images, the space point or points to be evaluated are themselves determined and in the case of the reconstruction of the obj the 3D coordinates of the spatial point (s) to be evaluated or, in the case of the measurement, the distance (s) determined between the spatial points to be evaluated, characterized in that - to capture the at least two digital images of the object, a planar quadrangle as the reference element in the object view is added - that in the evaluation of the recorded digital images, the image coordinates of the vertices of each imaged planar quadrilateral are determined, - that are determined from each recorded digital image, the positions of the vertices of the planar quadrilateral in each of the camera viewpoint of the image acquisition specific 3D coordinate representation in that the 3D coordinates of these vertices are produced Euclidean relative to the respective camera center, and that the three-dimensional camera positon from these in each case from the camera viewpoint of the image recordings of specific 3D coordinates of the corner points of the planar quadrilateral an optimized transformation is calculated using these 3D coordinates of the vertices for every two images, which converts 3D point correspondences between the cameras by translation and rotation only, using a representation of the transformation with dual unit quaternions, so that through ...

Description

The invention relates to a method for highly accurate three-dimensional measurement and / or reconstruction of objects using digital image recordings with minimal effort, for example for image analysis of traffic routes. Frequently, scenarios, not only in the case of accidents, geometric acquisitions, etc., must be used to record images from which one or more individual objects or pixels are determined in their position (reconstruction) and / or in their relative position relative to one another (measurement) As exact as possible coordinate detection is of crucial importance for the purpose of the image analysis.

The applicability of the invention extends to reconstruction and measurement of any mapped room points in outdoor as well as indoor areas.

State of the art

There are methods for the three-dimensional reconstruction of surfaces that use the projection of light patterns on the object to be reconstructed (for example DE 197 38 480 C1 ). These methods require complex and expensive equipment, are cumbersome in operation and evaluation and generally allow only an inaccurate evaluation due to disturbances of the light pattern as well as the detection of the same in real environments. An outdoor application is therefore extremely limited.

Furthermore, methods are known (eg. US 6,348,918 B1 ), which in the 3D reconstruction from digital image recordings presuppose the complete projection matrices of the cameras as known and include them in the calculation. Thus, the exact determination of the complete camera matrices possibly requires a very high procedural effort (in the patent was also not disclosed how the given matrices come about) and methods that will not be discussed in the specific case also. In addition, estimation-based iterative methods are used in the specific case, which do not allow statements about termination, accuracy of calculation, error rate and computation time.

Methods are also known (e.g. EP 1 098 268 A2 ), in which a camera movable on a device in only certain directions is used to obtain several shots and information about the camera positions. These require in addition to the camera an additional apparatus, which also has to work very accurately and their parameters must be determined very accurately, which in particular creates a high technical-economic and installation effort. In addition, this method of 3D measurement / reconstruction at larger distances (such as in outdoor areas) between the object to be measured and the camera is very inaccurate, since the disparities of corresponding pixels are too small and thus very susceptible to interference.

Furthermore, some methods (for example WO 99/36884 A1 ) 3D measurement only very limited. In the concrete example, the track to be measured must even be collinear to a track of known length, which is not practical for general surveying tasks.

Other methods (eg BZ Zhang: A Flexible New Technique for Camera Calibration - IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22 (11), 2000, pp. 1330-1334) use iterative nonlinear optimization to calculate the external Camera parameters and thus deliver results without statements about their accuracy or quality.

The US 6,621,921, B1 discloses an image processing method for measuring three-dimensional objects. The object of unknown dimensions is mapped together with a known pattern in different directions. On the basis of known spatial relations of selected points of the pattern, the unknown dimensions of the object can be determined by means of various coordinate systems which are converted into each other. In this case, separate coordinate systems are calculated for a left and a right camera of the image processing device in each case based on a reference coordinate system, whereby at least two calculation processes are required.

From the DE 197 46 942 A1 A method is known for determining correction factors for video signal processing based measuring systems. Based on the pattern structure proposed there, a coordinate system is established. Recorded and related to this coordinate system measurement points are compared with a video camera and determined correction factors, the specific process is not the subject of the disclosure.

task

The invention is based on the object to provide a universally applicable and fast non-pro-babilistic method for 3D measurement and / or reconstruction of objects, which with low economic and procedural effort, especially easy to implement image acquisition and evaluation conditions without any special requirements for the objects, reliable and robust against disturbances in the image acquisition largely robust (robust) data high and in this respect also assessable accuracy provides.

According to the invention, a special planar quadrilateral, for example a square, is inserted as a reference element with known geometry into the scene to be detected by camera shots, which is recorded together with the object to be measured or reconstructed from different angles (camera positions) for the purpose of three-dimensional evaluation ,

In the evaluation of these individual digital images, the vertices of the said planar quadrilateral are determined as image coordinates as well as in the specific camera coordinate point of view of the specific 3D coordinate representation. Taking into account at least two images of the object, the 3D camera positions of the image recordings are determined as the basis for the computationally known evaluation of the three-dimensional measurement or reconstruction of the object, taking into account the 3D coordinates considered by the respective camera viewpoint of the individual image acquisition.

With the proposed method, it is possible to realize a metric 3D measurement with absolutely minimal interaction at the place of use and in the postprocessing on the computer and thereby to provide mathematically stable and error-robust results that can be evaluated with regard to their accuracy. This is achieved by a semantically appropriate separation of the problem (monocular reconstruction-Euclidean transformation reconstruction). As a result of this breakdown obtained mathematically and the process costs each best possible controllable sub-problems, this can solve directly and stably (each solutions in terms of least squares) and thus dominates the entire process, which is thus highly accurate and insensitive to interference. Required input parameters, such as camera parameters and coordinates of the quadrilateral, need only be determined once in the laboratory (calibration) if not known by the manufacturer. The interaction at the measurement or reconstruction site is limited to the placement of the said reference element (planar quadrilateral), so that in addition to the operating effort (handling) and the technical complexity is minimal. The method can be used universally, both outdoors and indoors, and does not impose any conditions on the object to be measured or reconstructed.

embodiment

The invention will be explained in more detail below with reference to an embodiment shown in the drawing.

The figure shows the image recording of a scene to be evaluated, in particular an object to be measured or reconstructed in it 1 each from different camera positions 2 . 3 (Image captures the object 1 from different angles). In the overlap area of the beam paths 4 respectively. 5 shown fields of view of the two images to be recorded is a three-dimensional evaluation possible.

In these through the beam paths 4 and 5 limited common scene area of the images is inventively a planar quadrilateral 6 placed as a reference element with known geometry, which together with the object 1 from every camera position 2 . 3 is recorded by a camera, not shown for reasons of clarity. In the example shown, the reference element is realized by a planar square. This is inserted in a relative position in the said scene area, that the camera for the later exact coordinate determination to be detected in the evaluated digital images vertices of the same in the image recording from the camera positions 2 . 3 each as steep as possible on the planar quadrangle 6 is directed.

After recording the scene area including the reference element in at least two digital image recordings, the same is done on the computer. For this purpose, the image points of the corner points of the planar quadrilateral are in the image recordings 6 and the pixels of the points to be reconstructed or measured by manual selection or semi-automated by means of colored markings of the corner points of the planar quadrilateral 6 determined. With the image coordinates of the vertices of the planar quadrilateral 6 and its known geometry as inputs are the monocular reconstruction (see K. Voss, R. Neubauer, M. Schubert: Monocular reconstruction for Robot Vision - Verlag Shaker, Aachen, 1995, p.76 ff.) The 3D coordinates of these vertices related to the camera center, Euclidian produced. Thus, the exact 3D vertex coordinates of the planar quadrilateral are obtained for each image acquisition 6 in relation to the respective camera center, by solving only three uncomplicated systems of linear equations. With these image-related 3D coordinates of the vertices of the planar quadrilateral 6 For every two images a transformation can be calculated, which converts the 3D point correspondences between the cameras only by rotation and translation. Since interference always occurs in practice, this transformation is done using the method of Walker and Shao (Walker, L. Shao, Mass., Volz: Estimating 3D location Parameters using dual number quaternions - Computer Vision, Graphics, and Image Processing: Image Under-standing , Vol. 54, 1991, pp. 358-367) as the best solution in terms of least squares. In this case, a representation of the transformation with dual unit quaternions is used, which results in that for the optimal solution only the eigenvector for the largest eigenvalue of a certain matrix has to be determined, ie in the solution of an eigenvalue problem which is fully controllable. If one has determined this transformation (rotation and translation) between the three-dimensional point correspondences of two exposures or cameras, these are at the same time the parameters for rotation and translation between the cameras, ie their external parameters. Thus, without limiting generality, the center of the world coordinate system is placed in the projection center of the first camera and, using the mentioned method, it calculates the rotation and translation of the second camera (or with several cameras of each other) with respect to the first camera. This calculation determines the external parameters and thus the complete image matrices of the cameras. This can be calculated using the pixels of the spatial point (s) to be measured or reconstructed using standard methods (for example R. Hartley, A. Zisserman: Multiple View Geometry in Computer Vision, Second Edition - Cambridge University Press, Cambridge, 2002, p. 312 et seq .) the position of the respective points in the 3D space can be calculated by triangulation known per se, for which only the singular value decomposition of a matrix has to be performed, in turn the best solution in terms of least squares (but here for the 3D coordinates ) to obtain.

The applied mathematical methods (solution of linear equation systems, solution of an eigenvalue problem of a matrix, singular value decomposition of a matrix) are fully controllable, directly solvable and stable. It is, as described, each calculated the best solution in terms of least squares.

If the 3D coordinates of the one or more spatial points are determined, the Euclidean distance can be calculated in a simple manner when measuring between two spatial points, or the 3D coordinates can be obtained in the case of a reconstruction.

LIST OF REFERENCE NUMBERS

1

object

2, 3

camera position

4, 5

beam path

6

planar quadrilateral

Claims (5)

Method for high-precision three-dimensional measurement and / or reconstruction of objects with the aid of digital image recordings, for example for image evaluation of traffic routes, where at least two digital images of the object are inserted with a camera to be calibrated once before the first use with the simultaneous image of one inserted into the object view Reference element of known geometry from different camera positions are recorded, in which in each case the reference element is determined in the evaluation of the recorded images of the object for its three-dimensional measurement or reconstruction, in which in each of the images taken or the pixels of the spatial points to be evaluated, and then taking into account the determined three-dimensional camera positions of the recorded images, the space point or points to be evaluated are themselves determined and in the case of the reconstruction of the obj the 3D coordinates of the spatial point (s) to be evaluated or, in the case of the measurement, the distance (s) determined between the spatial points to be evaluated, characterized in that - to capture the at least two digital images of the object, a planar quadrangle as the reference element in the object view is added - that in the evaluation of the recorded digital images, the image coordinates of the vertices of each imaged planar quadrilateral are determined, - that are determined from each recorded digital image, the positions of the vertices of the planar quadrilateral in each of the camera viewpoint of the image acquisition specific 3D coordinate representation in that the 3D coordinates of these vertices are produced Euclidean relative to the respective camera center, and that the three-dimensional camera positon from these in each case from the camera viewpoint of the image recordings of specific 3D coordinates of the corner points of the planar quadrilateral an optimized transformation is calculated using these 3D coordinates of the vertices for every two images, which converts 3D point correspondences between the cameras by translation and rotation only, using a representation of the transformation with dual unit quaternions, so that the transformation determines the complete imaging matrices of the cameras with the external parameters of the cameras. A method according to claim 1, characterized in that for the automatic and exact subpixel accurate in the digital images, determination of the vertices of the planar quadrilateral whose corners are preferably colored or marked as brightness or gray level value. A method according to claim 1, characterized in that a square, for example with 1 m edge length, is provided as a planar quadrilateral. A method according to claim 1, characterized in that the position of the planar quadrangle to be inserted into the object view and / or the camera positions for the image recordings of the object are selected so that the camera for exact coordinate determination of the to be evaluated in the digital images to be detected vertices of the planar quadrilateral is aimed as steeply as possible on the planar quadrilateral when taking pictures. A method according to claim 1, characterized in that the camera positions for receiving the at least two digital images of the object in the position of their directed on the object optical axis each differ by an angle of at least 30 °.

Images