Detection and validation of objects from sequential images of a camera by using homographies

Abstract

A method and a device are for identifying objects from camera images, e.g. for vehicle driver assistance systems. The method involves: capturing a series of camera images, determining a plurality of planes in space by associating adjacent corresponding features in at least two consecutive camera images with a given one of the planes, determining a relative translation vector of a plane, and identifying dynamic objects in the camera images based on the relative translation vector of an associated plane.

Claims

1. A method of detecting dynamic objects, comprising the steps: a) with a camera of a vehicle, capturing a series of images including a first image at a first time and a second image at a second time after the first time; b) determining a plurality of corresponding feature pairs, wherein each one of the corresponding feature pairs respectively consists of corresponding first and second features in the first and second images; c) establishing a plurality of spatial planes including at least one back plane extending normal to a longitudinal direction of the vehicle, at least one ground plane extending horizontally, and at least one side plane extending vertically and along the longitudinal direction of the vehicle, wherein each respective one of the spatial planes is established by associating, with the respective spatial plane, a respective plurality of adjacent feature pairs among the corresponding feature pairs; d) detecting an object in the series of images based on the adjacent feature pairs associated with a selected one of the spatial planes that have been established; e) determining a respective relative translation vector having orthogonal vector components t.sub.x, t.sub.y and t.sub.z for each respective one of the spatial planes as a measure of a motion of the respective spatial plane; and f) identifying the object as a dynamic object based on the relative translation vector of the selected spatial plane; wherein the associating of the adjacent feature pairs with the respective spatial planes comprises computing homographies and using the homographies to associate the adjacent feature pairs with the respective spatial planes; wherein the homographies describe correspondences of points or features in a respective selected image region in the first image at the first time with same or corresponding points or features in the respective selected image region in the second image at the second time; wherein each one of the homographies has only three degrees of freedom respectively relating to the orthogonal vector components t.sub.x, t.sub.y and t.sub.z; and wherein the determining of the respective relative translation vector for each one of the spatial planes involves inverting a 3×3 matrix to solve for the three degrees of freedom of the respective homography associated with the respective plane; and wherein the respective relative translation vector indicates the orientation of the selected image region.

2. The method according to claim 1, wherein the relative translation vector is defined as t/d, wherein t represents a translation of the camera of the vehicle, and d represents a distance from the selected spatial plane.

3. The method according to claim 2, wherein t/d=(t.sub.x, t.sub.y, t.sub.z).

4. The method according to claim 1, wherein the associating of the adjacent feature pairs with the respective spatial planes comprises: computing respective ones of the homographies respectively for the ground plane, the back plane and the side plane, projecting the first feature from the first image as a respective projected feature respectively onto the ground plane, the back plane and the side plane using respective applicable ones of the homographies, determining respective projection errors as respective differences between the second feature in the second image and the respective projected feature respectively for the ground plane, the back plane and the side plane, and selecting, as the spatial plane to which the adjacent feature pairs are associated, the one of the ground plane, the back plane and the side plane for which the respective projection error is the smallest among the projection errors.

5. The method according to claim 1, wherein a respective one of the homographies is computed for the back plane in accordance with: $\left[\begin{array}{c}{x}_{0}c-a\\ y_{0}c-b\\ \mathrm{.Math.}\end{array}\right]=\left[\begin{array}{ccc}-x_1& 0& x_1{x}_0\\ 0& -x_1& x_1{y}_0\\ & \mathrm{.Math.}& \end{array}\right]\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right]$ wherein a, b and c are constants, x.sub.1 and y.sub.1 are coordinates of the first feature in the first image, x.sub.0 and y.sub.0 are coordinates of the second feature in the second image corresponding to the first feature in the first image, t.sub.x, t.sub.y and t.sub.z are the orthogonal vector components of the relative translation vector represented as t/d, t represents a translation of the camera, and d represents a distance from the respective back plane.

6. The method according to claim 1, wherein a respective one of the homographies is computed for the ground plane in accordance with: $\left[\begin{array}{c}{x}_{0}c-a\\ y_{0}c-b\\ \mathrm{.Math.}\end{array}\right]=\left[\begin{array}{ccc}-y_1& 0& y_1{x}_0\\ 0& -y_1& y_1{y}_0\\ & \mathrm{.Math.}& \end{array}\right]*\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right]$ wherein a, b and c are constants, x.sub.1 and y.sub.1 are coordinates of the first feature in the first image, x.sub.0 and y.sub.0 are coordinates of the second feature in the second image corresponding to the first feature in the first image, t.sub.x, t.sub.y and t.sub.z are the orthogonal vector components of the relative translation vector represented as t/d, t represents a translation of the camera, and d represents a distance from the respective ground plane.

7. The method according to claim 1, wherein a respective one of the homographies is computed for the side plane in accordance with: $\left[\begin{array}{c}{x}_{0}c-a\\ y_{0}c-b\\ \mathrm{.Math.}\end{array}\right]=\left[\begin{array}{ccc}-1& 0& x_0\\ 0& -1& y_0\\ & \mathrm{.Math.}& \end{array}\right]\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right]$ wherein a, b and c are constants, x.sub.1 and y.sub.1 are coordinates of the first feature in the first image, x.sub.0 and y.sub.0 are coordinates of the second feature in the second image corresponding to the first feature in the first image, t.sub.x, t.sub.y and t.sub.z are the orthogonal vector components of the relative translation vector represented as t/d, t represents a translation of the camera, and d represents a distance from the respective side plane.

8. The method according to claim 1, further comprising: determining that a particular one of the spatial planes is a static plane when the motion thereof is zero as indicated by the respective relative translation vector thereof having a vector length of zero, and determining a self-rotation of the camera between the first image and the second image based on the adjacent feature pairs associated with the at least one static plane.

9. The method according to claim 8, further comprising predicting image coordinates of an epipole x.sub.s,t0 of the second image at the second time t.sub.0 according to x.sub.s,t0=H.sub.st*x.sub.s,t1, wherein H.sub.st represents a homography of the at least one static plane and x.sub.s,t1 represents image coordinates of an epipole of the first image at the first time t.sub.1.

10. The method according to claim 9, further comprising establishing a pitch rate of the camera as a tan((x.sub.s0−x.sub.s1)/f) wherein f is a focal length of the camera with respect to one pixel of the camera.

11. The method according to claim 9, further comprising establishing a yaw rate of the camera as a tan(y.sub.s0−y.sub.s1)/f) wherein f is a focal length of the camera with respect to one pixel of the camera.

12. The method according to claim 1, further comprising: determining and compensating a self-rotation of the camera, and identifying the dynamic object as an overtaking vehicle when the relative translation vector of the selected spatial plane to which the dynamic object has been associated includes the vector component t.sub.z having a negative value, wherein the vector component t.sub.z extends in the longitudinal direction of the vehicle.

13. The method according to claim 1, further comprising: determining and compensating a self-rotation of the camera, and identifying the dynamic object as an approaching potential collision object when the relative translation vector of the selected spatial plane to which the dynamic object has been associated includes the vector component t.sub.z having a positive value, wherein the vector component t.sub.z extends in the longitudinal direction of the vehicle.

14. The method according to claim 1, further comprising: determining and compensating a self-rotation of the camera, and identifying the dynamic object as an other vehicle driving in a curve when the relative translation vector of the selected spatial plane to which the dynamic object has been associated includes the vector component t.sub.x having a non-zero value, wherein the vector component t.sub.x extends horizontally and transversely to the longitudinal direction of the vehicle.

15. The method according to claim 1, further comprising labeling the dynamic object as a moving object, and tracking the dynamic object in successive images of the series of images.

16. The method according to claim 1, further comprising: assigning to the dynamic object a fixed width, a fixed height and/or a fixed length in space, and performing geometric tracking of the dynamic object.

17. The method according to claim 1, further comprising operating a driver assistance system of the vehicle in response to and dependent on the dynamic object.

18. A device for detecting dynamic objects, comprising a camera controller and evaluation electronics, wherein: the camera controller is configured to capture, with a camera of a vehicle, a series of images including a first image at a first time and a second image at a second time after the first time; and the evaluation electronics are configured: to determine a plurality of corresponding feature pairs, wherein each one of the corresponding feature pairs respectively consists of corresponding first and second features in the first and second images; to establish a plurality of spatial planes including at least one back plane extending normal to a longitudinal direction of the vehicle, at least one ground plane extending horizontally, and at least one side plane extending vertically and along the longitudinal direction of the vehicle, wherein each respective one of the spatial planes is established by associating, with the respective spatial plane, a respective plurality of adjacent feature pairs among the corresponding feature pairs; to detect an object in the series of images based on the adjacent feature pairs associated with a selected one of the spatial planes that have been established; to determine a respective relative translation vector having orthogonal vector components t.sub.x, t.sub.y and t.sub.z for each respective one of the spatial planes as a measure of a motion of the respective spatial plane; and to identify the object as a dynamic object based on the relative translation vector of the selected spatial plane; wherein the associating of the adjacent feature pairs with the respective spatial planes comprises computing homographies and using the homographies to associate the adjacent feature pairs with the respective spatial planes; wherein the homographies describe correspondences of points or features in a respective selected image region in the first image at the first time with same or corresponding points or features in the respective selected image region in the second image at the second time; wherein each one of the homographies has only three degrees of freedom respectively relating to the orthogonal vector components t.sub.x, t.sub.y and t.sub.z; and wherein the determining of the respective relative translation vector for each one of the spatial planes involves inverting a 3×3 matrix to solve for the three degrees of freedom of the respective homography associated with the respective plane; and wherein the respective relative translation vector indicates the orientation of the selected image region.

19. The device according to claim 18, wherein the camera is a single monocular camera and the images are each mono-images.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) Further features, advantages and effects of the invention are set out by the following description of preferred embodiment examples of the invention, wherein:

(2) FIG. 1 schematically shows a typical deformation of an approaching back plane;

(3) FIG. 2 schematically shows a typical deformation of an approaching ground plane;

(4) FIG. 3 schematically shows a typical deformation of a) a rapidly approaching back planes and b) a slowly approaching or more distant back plane;

(5) FIG. 4 schematically shows a subdivision of an image having two different segments into cells;

(6) FIG. 5 shows segmenting results following a third iteration step;

(7) FIG. 6 shows a plane orientation for target validation (validation of potential collision objects);

(8) FIG. 7 shows time to collision monitoring; and

(9) FIG. 8 shows a projection (or warping) of the guardrail segment at the point in time t−0 (right) onto t−1 (left).

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS OF THE INVENTION

(10) Parts corresponding to one another are, as a general rule, provided with the same reference numerals in all of the figures.

(11) FIG. 1 schematically shows a back plane which occupies the region (20, dotted line) represented by hatching, at a first point in time t−1. At a subsequent second point in time t, the clearance or spacing distance between the vehicle camera and the back plane has decreased, resulting in the deformation of the region (21, continuous line) of the back plane in the image as indicated by the arrows (d1). The region (20; 21) is scaled or increased in size as a consequence of the relative motion of the vehicle camera with respect to the back plane.

(12) FIG. 2 schematically shows a ground plane which occupies the region (30, dotted line) represented by hatching at a first point in time t−1. This could be a section of a road surface, on which the vehicle is driving. As a consequence of the egomotion of the vehicle camera, the region (as represented in the image) changes at a subsequent second point in time t, resulting in the deformation of the region (32) of the ground plane as indicated by the arrows (d2). At the second point in time t, the lines designated with 32 delimit the region of the ground plane. Here, the term “ground plane” thus denotes a delimited region on the road surface. The edge region is produced e.g. from signatures (or edge points) on the road surface, which can be tracked in the series of images.

(13) FIG. 3 illustrates the difference between a quickly approaching back plane (FIG. 3a: 20, 21; deformation d1) and a slowly approaching back plane (FIG. 3b: 20, 23; deformation d3), if at the first point in time t−1 the back plane (20) in FIG. 3a has the same clearance or spacing distance from the vehicle camera as the back plane (20) in FIG. 3b.

(14) Alternatively, FIG. 3 could represent the difference between a close back plane (FIG. 3a: 20, 21; deformation d1) and a more distant back plane (FIG. 3b: 20, 23; deformation d3), which move e.g. at the same (relative) speed, in which case the object (20, 23) represented in FIG. 3b would be larger in real space than the object (20, 21) represented in FIG. 3a.

(15) If, instead of individual correspondences, multiple adjacent correspondences are observed, objects can be segmented due to different speeds, scalings and deformation.

(16) If it is assumed that the world consists of planes, these can be described by homographs i.e. homographies and can be separated or distinguished from one another as shown below by means of their distance, speed and orientation. A homograph describes the correspondence of points on one plane between two camera positions or the correspondence of two points in two consecutive frames:

(17) $\begin{array}{cc}{x}_{t0}=H*x_{t1}\mathrm{with}{x}_{t0}=\left[\begin{array}{c}{x}_0\\ y_0\\ 1\end{array}\right],x_{t1}=\left[\begin{array}{c}{x}_1\\ y_1\\ 1\end{array}\right].& \left(1\right)\end{array}$

(18) In this case, the vector x.sub.t0 describes the 3D correspondence at the point in time t−0 of the vector x.sub.t1 at the point in time t−1. A homograph can be computed, in an image-based manner, by knowledge of four point correspondences (cf. Tutorial: Multiple View Geometry, Hartley, R. and Zisserman, A., CVPR June 1999: https://de.scribd.com/document/96810936/Hartley-Tut-4up accessed on 26 Sep. 2016). The relationships indicated at the top left (slide 21) of page 6 of the tutorial can be formulated as follows in the notation of equation 1:

(19) $\begin{array}{cc}\left[\begin{array}{ccccccccc}-x_0& -y_0& -1& 0& 0& 0& x_1{x}_0& x_1{y}_0& x_1\\ 0& 0& 0& -x_0& -y_0& -1& y_1{x}_0& y_1{y}_0& y_1\\ & & & & \mathrm{.Math.}& & & & \end{array}\right]*\left[\begin{array}{c}{h}_1\\ h_2\\ h_3\\ h_4\\ h_5\\ h_6\\ h_7\\ h_8\\ h_9\end{array}\right]=0\mathrm{where}H=\left[\begin{array}{ccc}{h}_1& h_2& h_3\\ h_4& h_5& h_6\\ h_7& h_8& h_9\end{array}\right].& \left(2\right)\end{array}$

(20) Alternatively, knowing the camera translation t, the rotation R and the distance d along the normal vector n of the plane, the homograph can be computed in accordance with equation 3. Equation 3 illustrates that, at a nonzero inverse TTC t/d, planes having a different orientation n can be modelled and that planes having an identical orientation n can be separated by means of their inverse TTC.

(21) $\begin{array}{cc}H=\left[R-\frac{t*n^{\prime }}{d}\right]& \left(3\right)\end{array}$

(22) A homograph can be theoretically decomposed into the normal vector n, the rotation matrix R and the inverse TTC t/d. Unfortunately, this decomposition is numerically extremely unstable and sensitive to measuring errors.

(23) If a scene is described by planes, it can be segmented as indicated below.

(24) FIG. 4 schematically shows a subdivision into cells (grid, gridlines). The scene is subdivided into N×M initial cells and a clear i.e. unambiguous or unique ID is assigned respectively to each point correspondence. This ID firstly indicates the affiliation to a cell. The ID can subsequently indicate the affiliation to a cluster or an object. An object (in particular a back plane) is represented hatched in the foreground. The background is represented in white.

(25) If a cell comprises only one object (cells B3, D3), a homograph or homography will describe this cell very well. If, however, a cell contains more than one object (cell C3), the homograph will not describe either of the two objects well. If the point correspondences (black dot or black cross or x) are associated with the clusters (or segment) of the adjacent cells (B3 or D3) by means of their back projection error, the black dot is associated with the segment of the cell B3 and the black cross is associated with the segment of the cell D3, because the homograph for the cell C3 does not describe either the foreground or the background well.

(26) If prior knowledge of a scene exists, the segment sizes can be adjusted to the scene in that e.g. larger regions in the close region of the vehicle or in regions having a positive classification answer can be generated. A dedicated back plane, ground plane and side plane homograph i.e. homography is computed for each segment, as shown in equations 5 to 10.

(27) The computation of the back plane, ground plane and side plane homograph i.e. homography increases the selectivity because a homography with fewer degrees of freedom can only poorly model regions which contain different planes and, consequently, corresponding points will have a higher back projection error, see FIG. 4. Therefore, the back projection error e.sub.i is a measure of how well a point x at the second point in time t−0 is described by the homography of a plane i of the corresponding point at the first point in time t−1:
e.sub.i=x.sub.t0−H.sub.ix.sub.e1.  (4)

(28) If the static installation position of the camera and camera rotation are assumed in two different views (e.g. due to knowledge of the camera calibration and due to the computation of the fundamental matrix in a monocular system or due to rotation values of a rotation rate sensor cluster), the inverse TTC t/d can be computed by means of the flux vectors compensated for by the static camera rotation, as is shown below by way of example for a ground plane n′=[0 1 0]. If the rotation is not known, it can be approximately replaced by a unit matrix.

(29) If the quotient t/d is substituted by the inverse time to collision

(30) $\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right]\hspace{1em}$
in equation 3, it follows that:

(31) $\begin{array}{cc}\left[\begin{array}{c}{x}_0\\ y_0\\ 1\end{array}\right]=\left[\begin{array}{ccc}0& t_x& 0\\ R-0& t_y& 0\\ 0& t_z& 0\end{array}\right]\left[\begin{array}{c}{x}_1\\ y_1\\ 1\end{array}\right].Math.\left[\begin{array}{c}{x}_0\\ y_0\\ 1\end{array}\right]-R\left[\begin{array}{c}{x}_1\\ y_1\\ 1\end{array}\right]=-\left[\begin{array}{ccc}0& t_x& 0\\ 0& t_y& 0\\ 0& t_z& 0\end{array}\right]\left[\begin{array}{c}{x}_1\\ y_1\\ 1\end{array}\right].& \left(5\right)\end{array}$

(32) By introducing the constants a, b, c, wherein

(33) $\left[\begin{array}{c}a\\ b\\ c\end{array}\right]:=R\left[\begin{array}{c}{x}_1\\ y_1\\ 1\end{array}\right],$
equation 5 produces the simplified form:

(34) $\begin{array}{cc}\left[\begin{array}{c}{x}_0\\ y_0\\ 1\end{array}\right]-\left[\begin{array}{c}a\\ b\\ c\end{array}\right]=-\left[\begin{array}{c}{t}_x{y}_1\\ t_y{y}_1\\ t_z{y}_1\end{array}\right].Math.\left[\begin{array}{c}{x}_0\\ y_0\\ 1\end{array}\right]=\left[\begin{array}{c}a\\ b\\ c\end{array}\right]-\left[\begin{array}{c}{t}_x{y}_1\\ t_y{y}_1\\ t_z{y}_1\end{array}\right].& \left(6\right)\end{array}$

(35) The result of standardizing the homogeneous coordinates is:
x.sub.0(c−t.sub.zy.sub.1)=a−t.sub.xy.sub.1  (7)
y.sub.0(c−t.sub.zy.sub.1)=b−t.sub.yy.sub.1  (8)

(36) For more than one measurement, an equation system of the form Mx=v with a vector x to be established, a matrix M and a vector v (see equation 9) is produced, which can be solved for at least three image correspondences as sampling points by e.g. a singular value decomposition or a least squares method:

(37) $\begin{array}{cc}\left[\begin{array}{c}{x}_{0}c-a\\ y_{0}c-b\\ \mathrm{.Math.}\end{array}\right]=\left[\begin{array}{ccc}-y_1& 0& y_1{x}_0\\ 0& -y_1& y_1{y}_0\\ & \mathrm{.Math.}& \end{array}\right]*\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right]\hspace{1em}.& \left(9\right)\end{array}$

(38) The back and side plane homographs i.e. homographies can be deduced similarly and respectively produce:

(39) $\begin{array}{cc}\left[\begin{array}{c}{x}_{0}c-a\\ y_{0}c-b\\ \mathrm{.Math.}\end{array}\right]=\left[\begin{array}{ccc}-x_1& 0& x_1{x}_0\\ 0& -x_1& x_1{y}_0\\ & \mathrm{.Math.}& \end{array}\right]\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right]& \left(10\right)\\ \mathrm{or}\mathrm{respectively}\left[\begin{array}{c}{x}_{0}c-a\\ y_{0}c-b\\ \mathrm{.Math.}\end{array}\right]=\left[\begin{array}{ccc}-1& 0& x_0\\ 0& -1& y_0\\ & \mathrm{.Math.}& \end{array}\right]\left[\begin{array}{c}{t}_x\\ t_y\\ t_z\end{array}\right].& \left(11\right)\end{array}$

(40) In order to segment larger objects consisting of multiple cells, adjacent cells can be combined in a further step, in that the back projection or reprojection errors Σx.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i or Σx.sub.t0.sup.j−H.sub.ix.sub.t1.sup.j are computed by means of sampling points (see point 1 below: RANSAC) of the adjacent segments j and i and their homographs i.e. homographies. Two adjacent clusters are combined, if Σx.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i is less than Σx.sub.t0.sup.i−H.sub.ix.sub.t1.sup.i or e.g. the back projection error standardized for the predicted flux length is below an adjustable threshold. In particular, two adjacent clusters can be combined, if Σx.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i is less than Σx.sub.t0.sup.i−H.sub.ix.sub.t1.sup.i and the two back projection errors Σx.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i and Σx.sub.t0.sup.i−H.sub.ix.sub.t1.sup.i fall below a threshold standardized for the flux length. Alternatively, back projection errors can be used as potentials in a graph and a global solution can be computed. The compactness of the clusters can, in this case, be established via the edge potentials in the graph.

(41) If the segments have been combined, the homographs i.e. homographies are computed again and the point correspondences are associated with the clusters having the smallest back projection error. f only directly neighboring clusters are observed, very compact objects can be generated. If the minimum error exceeds an adjustable threshold, new (cluster/object) IDs are assigned to the correspondences, in order to be able to identify partially concealed objects or objects having a slightly different TTC. By adjusting the threshold, the resolution of (slightly) different objects can be adjusted.

(42) The back projection errors can be provided with a bias which reduces the costs for related regions or a bias which increases the costs for an ID change, if point correspondences were to have the same ID affiliation over a longer period of time.

(43) FIG. 5 shows one example of a scene segmentation:

(44) FIG. 5a shows an image which has been captured by a vehicle camera which is arranged in the interior of the vehicle and records the surroundings lying ahead through the windshield. A three-lane road (51), e.g. a motorway, can be seen. The lanes are separated by appropriate lane markings. Vehicles are driving on all three lanes. The vehicle (53) driving ahead on the ego lane possibly conceals further vehicles driving ahead, which are located on the ego lane. A structural elevated delimitation (52) with respect to the opposite carriageway is located on the left of the three-lane road. A shoulder or breakdown lane, which is delimited to the right by a guardrail, behind which there is woodland, is located to the right of the three-lane road (51). Sign gantries (54) can be identified some distance in front of the ego vehicle, one of which spans the three-lane road (51).

(45) This scene can be segmented in a similar way to the method described by means of FIG. 4. In FIGS. 5b to 5d, cells (56) can be identified. Point correspondences (55) are represented in the cells. The association of a cell (56) with a segment is represented by means of the color of the cell frame or the point correspondences (55).

(46) FIG. 5b shows the red channel of the segmented image,

(47) FIG. 5c shows the green channel and FIG. 5d shows the blue channel. Different segments have been provided with different colors. A segment, which is green in the original, extends over the lowest five to six lines (accordingly represented in white in FIGS. 5b and 5d and without a cell border). This segment corresponds to the ground plane, that is to say the surface of the road (51) on which the ego car is driving.

(48) A further segment can be identified in the middle of the image, in the original it is pink. It therefore has high red values in FIG. 5b, weaker blue values in FIG. 5d and no green values in FIG. 5c. This segment corresponds to the back plane of the (transporter) vehicle (53) driving ahead on the ego lane.

(49) The segmenting result shown was determined without prior knowledge of the scene in only three iteration steps. This shows the enormous speediness and performance of an embodiment of the invention by temporal integration.

(50) FIG. 6 shows a determination of the orientation of planes in the scene already described in FIG. 5. For the purposes of orientation, FIG. 6a shows the surrounding situation according to FIG. 5a again.

(51) FIG. 6b shows all of the correspondences which are associated with a side plane. The correspondences at the left edge have been associated with a right side plane, which is correct because the right side of the structural delimitation (52) with respect to the opposite carriageway is located there in the image. The correspondences in the right half of the image have been associated with the left side planes, which is likewise correct, because the “left side” of the road peripheral development or planting of vegetation is located there in the image.

(52) FIG. 6c shows which correspondences are associated with a ground plane, which is correct, because the surface of the road (51) can be seen there in the image.

(53) FIG. 6d shows which correspondences are associated with a back plane. This is mostly correct. Different back planes cannot yet be sufficiently distinguished from this determination alone, e.g. that of the delivery van (53) driving ahead on the same lane from the signs of the sign gantry (54) arranged above it in the image. However, important information regarding where elevated objects occur in the surroundings of the vehicle can already be extracted from this representation.

(54) As illustrated in FIG. 7, the inverse TTC (t.sub.x, t.sub.y, t.sub.z) can be used to identify dynamic objects.

(55) FIG. 7a, in turn, shows the image of the vehicle situation (identical to FIG. 6a). The vehicle (73) driving ahead on the ego lane is a delivery van. Two vehicles (71 and 72) are driving on the left lane and two further vehicles (74 and 75) are driving on the right lane.

(56) FIG. 7b shows correspondences which, in turn, correspond to the ground plane (violet in the original) and are the only ones to have a red proportion or color component.

(57) FIG. 7c shows correspondences which are associated with moving objects. These are green in the original if they are moving away from the ego vehicle (that is to say they are driving faster), or turquoise if they are driving more slowly.

(58) FIG. 7d shows correspondences having a blue proportion or color component, that is to say those which correspond to the ground plane (cf. FIG. 7b), moving objects which are approaching the ego vehicle (cf. FIG. 7c) and those which correspond to static elevated objects, these are only represented in FIG. 7d, such as e.g. woodland to the left and right of the motorway and the sign gantries. It can be seen from FIGS. 7c and 7d jointly that the vehicle (73) is approaching on the ego lane. The same applies to the front vehicle (75) on the right lane. On the other hand, the remaining vehicles (71, 72 and 74) are moving away. Due to a lack of structure in the image, the region which corresponds to the sky in the image does not result in any correspondences (white in FIGS. 7b to 7d).

(59) If the natural rotation is considered in the correspondences prior to the computation of the homograph i.e. homography, or if the natural rotation is considered in the rotation matrix R, overtaking vehicles can be identified due to their negative t, component or swerving vehicles or vehicles driving in a curve can be identified by a nonzero lateral t.sub.x component. If the dynamic segments are predicted by means of their homographs (see “consolidation of the optical flux based on homographs” below), a dynamic map can be constructed over time.

(60) If equation 3 is observed, it can be seen that segments having an inverse TTC equal to zero describe the rotation matrix and these can be established by computing a homograph with a full degree of freedom (equation 2) from segments with t/d equal to zero. If it is assumed that the translatory components in the vicinity of the epipole cannot make themselves felt, the pitch rate and yaw rate can also be established by predicting the coordinates of the epipole (x.sub.e, y.sub.e) through the homograph of static segments and computing the a tan((x.sub.e0−x.sub.e1)/f) or a tan((y.sub.e0−y.sub.e1)/f) with the focal length f based on one pixel.

(61) If a homograph is computed with all degrees of freedom for each cluster, this can also be used to reconstruct the 3D surroundings in that, instead of the measured position x.sub.t0, the predicted position H*x.sub.t1 is used for triangulation. This not only reduces the effect of measuring errors, but also makes it possible to reconstruct objects close to the epipole.

(62) One embodiment example for consolidating the optical flux based on homographs is described below.

(63) If the segmentation is known at the point in time t−1, it can be used to both predict the objects and to generate a dense flux field. Signature-based flux methods produce signatures and cause these to be clearly associated in consecutive frames. The signatures are mostly computed from a patch (image section or image region) of a defined size. If, however, the size and form of a patch alter, it is no longer possible to find a correspondence with a fixed template (model, specimen, e.g. an image section of an image of the series of images, which corresponds to an object—for example a vehicle template—is meant). If e.g. one is approaching a back plane, the size of a patch changes. Or if one is moving over a ground plane or parallel to a side plane, both the size and the form of a patch change, see FIGS. 1 and 2). If the segmentation exists at the point in time t−1, the homographs can be computed again by means of flux vectors which have already been found, and can be used to predict the position and form of correspondences of t−1 to t−0 which have already been established.

(64) Alternatively, the current frame can be transformed at the point in time t−0 to the point in time t−1, in order to compensate for changes in scale and form.

(65) FIG. 8 illustrates such a procedure.

(66) FIG. 8a shows an image of another driving situation which has been captured by the vehicle camera at a point in time t−1. A motorway having three lanes in each direction of travel can be seen. To the left of the ego three-lane road there is a guardrail (81) as an elevated delimitation with respect to the opposite carriageway. A noise barrier (82) is located on the right of the road.

(67) FIG. 8b shows an image which was captured at the subsequent point in time t and transformed (warped) by means of the homograph of the guardrail such that the changes in the image occurring as a consequence of the motion of the vehicle and therefore of the vehicle camera between the two capturing times are compensated for in the region of the guardrail. In FIG. 8b, the forward motion of the ego vehicle results in the most obvious graduation line of the lane marking being closer to the ego vehicle than in FIG. 8a. The transformation results in the trapezoidal shifting of the image, which is illustrated by a dashed line in FIG. 8f.

(68) FIG. 8c then shows corresponding features (85), which have been determined in the region of the guardrail (81, cf. FIG. 8a) as white dots.

(69) FIG. 8d shows where these corresponding features (86) are to be expected in the next image, after said image has been transformed as described with reference to FIG. 8b.

(70) In FIGS. 8e and 8f, this state of affairs is again shown in a black and white representation, wherein the corresponding features (85) now correspond to the black dots on the guardrail (81) in the left half of the image.

(71) In order to generate a dense flux field, the current image can thus be warped onto the previous image for each segment, in order to rediscover already existing correspondences which have changed in their scale or form, or in order to establish new correspondences by means of congruent templates.

(72) If not enough flux vectors for computing a homograph again are present in a current frame, the homograph from the last frame can be approximately used to make the correspondence finding more robust to changes in form and scale.

(73) The following configuration forms or aspects are advantageous and can be provided individually or in combination: 1. The image is subdivided into N×M cells and a clear i.e. unique or unambiguous cell ID is assigned to the point correspondences of a cell. The back plane, ground plane and side plane homographs i.e. homographies (equations 9, 10 and 11) are computed by means of RANSAC from the correspondences with the same IDs, and both the homography having the lowest back projection error and the sampling points used to calculate the homography are stored. In the case of the RANSAC (RAndom SAmple Consensus) method, a minimum number of randomly selected correspondences is usually used for each iteration, in order to form a hypothesis. A value, which describes whether the corresponding feature supports the hypothesis, is subsequently computed for each corresponding feature. If the hypothesis attains sufficient support through the corresponding features, then the non-supporting corresponding features can be rejected as outliers. Otherwise, a minimum number of correspondences is selected again at random. 2. The back projection errors Σx.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i or Σx.sub.t0.sup.i−H.sub.ix.sub.t1.sup.j are computed by means of the sampling points of the adjacent homograph i.e. homography for adjacent cells i, j. If the back projection error Σx.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i is less than Σx.sub.t0.sup.i−H.sub.ix.sub.t1.sup.i or if the errors fall below a threshold standardized for the flux length, the IDs are combined and the homographies are computed again. In particular, two adjacent cells can be clustered as belonging to the same plane (or to the same segment or to the same object), if the back projection error Σ.sub.i(x.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i) is less than Σ.sub.i(x.sub.t0.sup.i−H.sub.ix.sub.t1.sup.i) and if both back projection errors Σ.sub.i(x.sub.t0.sup.i−H.sub.jx.sub.t1.sup.i) and Σ.sub.i(x.sub.t0.sup.i−H.sub.ix.sub.t1.sup.i) fall below a threshold standardized for the flux length. 3. The back projection errors x.sub.e0−H.sub.ix.sub.t1 of all of the point correspondences are computed for the adjacent segments and a point correspondence is associated with the segment having the lowest back projection error. If the minimum error exceeds a threshold, the correspondences are provided with a new object ID in order to also be able to identify smaller or partially concealed objects. 4. The homographs of the segments extracted at the point in time t−1 are computed again at the start of a new frame (t−0) by means of the image correspondences already found and the already existing segment IDs in the current frame are predicted. If not enough flux vectors are available to compute a homograph again in the current frame, the homographs from the last frame can be approximately used. 5. In order to generate a dense flux field, the current frame (t−0) is warped onto the last frame (t−1) for each segment, in order to rediscover already existing correspondences which have changed in their scale or form, or in order to establish new correspondences. 6. The back projection errors of the back plane, ground plane and side plane can be used to validate elevated targets, see FIG. 6. 7. If e.g. a disparity map exists in a vehicle stereo camera, the absolute speeds can be computed from the inverse TTC t/d, because then the absolute distances d are present for individual pixels in the disparity map. 8. If a complete homograph is computed with all degrees of freedom for each segment, the rotation matrix R can be established from segments having a TTC close to infinite (or inverse TTCs approaching zero). 9. The 3D surroundings can be reconstructed from the predicted position (Hx.sub.t1, x.sub.t1) instead of the measured position (x.sub.t0, x.sub.t1) and make it possible to also reconstruct objects at the epipole.