DE102021203927A1 - Method and device for evaluating stereo image data from a camera system based on signatures - Google Patents

Abstract

The invention relates to a method and a device for evaluating stereo image data from a camera system, for example a vehicle-mounted environment detection camera, and can be used in a system for assisted or automated driving. The method evaluating stereo image data comprises the steps: a) receiving the stereo image data; b) generating local signatures (411, 412, 413, 421, 422) in first and second images (21, 22; 31, 32; 41, 42) of the stereo image data; c) forming hash keys based on one or more signatures ( 411, 412, 413, 421, 422) as soon as the strength of the signature exceeds a predetermined threshold; Hash function such that the hashkeys of different signatures (411, 412, 413, 421, 422) are stored in the hashtable at a distance from each other, with the hashkeys of less different n signatures are stored closer together and the hashkeys of more strongly different signatures are stored further away from each other;e) performing a hashkey correspondence search by matching the hashtable of the first and second images (21, 22; 31, 32; 41, 42) in such a way that the signatures (411, 412, 413, 421, 422) from the first and second images (21, 22; 31, 32; 41, 42) are unambiguously assigned; f) outputting 3D information on the basis of the assigned signatures (411, 412, 413, 421, 422). A major advantage is efficiency: the hashing mechanism used means that the correspondence can be determined quickly.

Description

The invention relates to a method and a device for evaluating stereo image data from a camera system, for example a vehicle-mounted environment detection camera, and can be used in a system for assisted or automated driving.

Today's vehicles are equipped with camera-based driver assistance systems, which are used to recognize objects to avoid collisions and to recognize road boundaries to keep the vehicle in lane.

A mono or stereo camera-based depth estimation with different camera systems is gaining in importance. The depth estimation benefits the recognition of objects and the understanding of the current situation surrounding the vehicle.

This depth estimation is done classically on feature-based (feature-based) approaches for correspondence search in rectified images or with neural networks, which calculate the depth from image pairs with trained features. Based on an offset of image features (features) in two images that were captured from different camera positions, the spatial distance to the real object that corresponds to this image feature can be calculated or at least estimated.

In current and future systems, due to safety-critical applications, especially in the area of depth estimation, it is assumed that a combination of classic and CNN-based depth estimation, or at least one classic or CNN-based depth estimation, is implemented. CNN is short for Convolutional Neural Network.

With stereo cameras, not only can a 2-dimensional image of the environment be recorded, but also - due to the recording of the environment from two different positions by the two offset camera modules of the stereo camera - the distance to recognized patterns (or image features) can be determined. In this way, the 3D geometry of detected objects can be reconstructed. An established method is based on the "Semi Global Matching" (SGM) procedure.
H. Hirschmüller, "Accurate and efficient stereo processing by semiglobal matching and mutual information," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2005, pp.807-814 and
H. Hirschmüller, S. Gehrig: "Stereo Matching in the Presence of Sub-Pixel Calibration Errors", International Conference on Vision and Pattern Recognition, CVPR 2009, show SGM methods.

In order to simplify the correspondence search, precise rectification of the camera images is required. With perfect rectification, the search for correspondence can be carried out on image lines - i.e. in one dimension. However, this requires precise calibration of the two camera modules. If the correspondence search has to be carried out in the image area (i.e. two-dimensionally), the computing requirements increase significantly.

It is an object of the invention to provide solutions for an improved evaluation of stereo image data that enables a reliable 3D reconstruction of the camera environment.

A starting point for the solution is to use an abstracted and compressed representation of the two images from the stereo camera

• Das beschriebene System besteht aus einer Anzahl von zufälligen, unabhängig voneinander generierten Processing Kernels. Die Kernel bilden Datensegmente auf parallelen Datenströmen als kompakte Signaturen in einen hochdimensionalen Raum ab. Dieses Vorgehen ist in WO 2007/049282 A1 näher beschrieben (siehe unten).
• Das System gemäß WO2009/026433A1 verwendet Multimedia-/Video-Daten als Input und streamt diese durch die Verarbeitungs-Kernel. Die Kernel generieren kompakte Signaturen sobald ein bestimmtes Inhalts-Segment gefunden wird. Die Signaturen werden auf herkömmliche Weise (in „a conventional way“) in einer Datenbank/Tabelle der Größe N gespeichert.
• Ein echtzeitfähiges Matching zwischen den Signaturen auf den Videodaten und den Signaturen in der Datenbank ist dann mit Berechnungsaufwand <= log(N) möglich Für das Video Signatur-Matching werden L Kernel verwendet. Jeder Kernel erzeugt auf den Bilddaten zwei Signaturen eine „Signature S“ und eine „Robust Signature RS“. Die Kernelfunktionen sind als einfache Look-up-Tables(gewichtete Summen) ausgelegt: V_i = SUMME_j(w_ij*k_j).
• Durch die Summation sind die Kernelfunktionen robust gegenüber „einfachem Rauschen“ (Gauss, Scratch, ..). Dagegen sind die Kernel (aus Laufzeitgründen) explizit nicht scale-, crop-, shift-, und rotations-invariant.
• Der Kernel-Output wird mit Schwellwerten Th_i beaufschlagt. Je ein hoher Schwellwert für die Generierung der robusten Signaturen „RS“ und ein niedriger Schwellwert für die Generierung „schwacher“ Signaturen „S“.
• Zur Abbildung von scale, distortion, crop, shift und rotation wird die Anzahl der Kernel um die entsprechende Anzahl Rotations-Stufen, Shift-Stufen u. Scale-Stufen erhöht.

The publication shows an example of an abstracted and compressed representation of image data WO 2009/026433 A1 :

• The system described consists of a number of random, independently generated processing kernels. The kernels map data segments on parallel data streams as compact signatures in a high-dimensional space. This approach is in WO 2007/049282 A1 described in more detail (see below).
• The system according to WO2009/026433A1 takes multimedia/video data as input and streams it through the processing kernels. The kernels generate compact signatures as soon as a specific content segment is found. The signatures are stored in a conventional way in a database/table of size N.
• A real-time matching between the signatures on the video data and the signatures in the database is then possible with calculation effort <= log(N) For video signature matching who uses the L kernel. Each kernel creates two signatures on the image data, a "Signature S" and a "Robust Signature RS". The kernel functions are designed as simple look-up tables (weighted sums): V_i = SUM_j(w_ij*k_j).
• Due to the summation, the kernel functions are robust against "simple noise" (Gauss, scratch, ..). In contrast, the kernels (due to runtime reasons) are explicitly not scale, crop, shift, and rotation invariant.
• Threshold values Th_i are applied to the kernel output. A high threshold for the generation of the robust signatures "RS" and a low threshold for the generation of "weak" signatures "S".
• To map scale, distortion, crop, shift and rotation, the number of kernels is increased by the corresponding number of rotation levels, shift levels and scale levels.

• Die Signaturen bestehen aus binären Einträgen für

Feature-Index [0001000000] z.B. Feature Nr.4 bei 10 Features Feature Scale/Rotation [1000000000] Feature Position [0000100000] falls das Feature bei Pixel 5
gefunden wurde
Diese Einträge werden aneinandergehängt und ergeben die Signatur:
[000100000010000000000000100000] für ein Einzelfeature.
Objekte bestehen aus mehreren Features, die überlagert werden. Die Signaturen können in realen Bildern mit mehr als 10 Pixeln sehr groß werden.The underlying release WO 2007/049282 A2 shows something like the following in an example:

• The signatures consist of binary entries for

feature index [0001000000] eg Feature No.4 with 10 features Feature Scale/Rotation [1000000000] feature position [0000100000] if the feature on Pixel 5
was found
These entries are concatenated and result in the signature:
[000100000010000000000000100000] for a single feature.
Objects are made up of multiple features that are overlaid. The signatures can become very large in real images with more than 10 pixels.

However, the "1" is not very common. Most binary entries are "0". Thus, for example, the binary representation is very well suited for fast hash matching algorithms.

One aspect of the invention relates to a computationally efficient method for determining correspondence in stereo image data of a camera system.

A further aspect of the invention relates to a device which is configured in such a way to carry out such a computationally efficient method.

1. a) Empfangen der Stereobilddaten;
2. b) Generieren von lokalen Signaturen in ersten und zweiten Bildern der Stereobilddaten;
3. c) Bilden von Hashkeys auf der Grundlage von einer oder mehreren Signaturen sobald eine Stärke der Signatur einen vorgegebenen Schwellwert überschreitet;
4. d) Speichern der Hashkeys in einer Hashtable für das erste und das zweite Bild mittels einer Hashfunktion derart, dass die Hashkeys von unterschiedlichen Signaturen in der Hashtable voneinander entfernt abgelegt werden, wobei die Hashkeys von weniger unterschiedlichen Signaturen näher beieinander abgelegt werden und die Hashkeys von stärker unterschiedlichen Signaturen weiter entfernt voneinander abgelegt werden;
5. e) Durchführen einer Hashkey-Korrespondenzsuche durch Matchen der Hashtable des ersten und des zweiten Bildes derart, dass eine eindeutige Zuordnung der Signaturen aus erstem und zweitem Bild resultiert;
6. f) Ausgabe der 3D-Informationen auf der Grundlage der zugeordneten Signaturen.

A method of evaluating stereo image data includes the steps:

1. a) receiving the stereo image data;
2. b) generating local signatures in first and second images of the stereo image data;
3. c) forming hash keys based on one or more signatures as soon as a strength of the signature exceeds a predetermined threshold value;
4. d) Storing the hashkeys in a hashtable for the first and the second image using a hash function such that the hashkeys from different signatures are stored away from each other in the hashtable, with the hashkeys from less different signatures being stored closer together and the hashkeys from stronger different signatures are placed further apart;
5. e) performing a hash key correspondence search by matching the hash table of the first and second images in such a way that the signatures from the first and second images are unambiguously assigned;
6. f) Output of the 3D information based on the assigned signatures.

The stereo image data can, for example, have been captured by a stereo camera with two camera modules at the same time (or approximately simultaneously), with the two camera modules being aligned in such a way that they image an overlapping area of the camera environment. Correspondences between the same image objects or structures in the image are established in this overlapping area.

Alternatively, the stereo image data can have been captured by a moving mono camera as individual images captured one after the other.

A stereo camera can be arranged within a housing, so that both camera modules or optronics are mechanically rigidly connected to one another. Alternatively, the camera modules can be used as in DE 102016217450 A1 described as individual cameras attached to different areas in or on a device or vehicle. The two camera modules can have identical components (optics and image recording sensor).
However, different components are also conceivable, such as, for example WO 2017/028848 A1 indicates. The stereo camera can be made up of a wide-angle and a telephoto camera module.
A stereo evaluation is basically possible in the "overlap area" of the image areas of two camera modules.
A stereo camera that is fixed in or on a vehicle (vehicle-mounted stereo camera) can be arranged, for example, behind the windshield inside the vehicle and can include two camera modules that can capture and image the area of the vehicle environment in front of the vehicle through the windshield.

As an alternative to a stereo camera with two separate camera modules, a moving mono camera can also supply stereo image pairs or stereo image data. Individual images recorded in chronological succession have different perspectives due to the movement of the mono camera. This also enables a "motion stereo" evaluation and thus a depth estimation or (partial) 3D reconstruction of the surroundings captured by the mono camera. A mono camera is actively moved, e.g. by driving a short distance or cornering a vehicle in or on which the mono camera is attached.

In one embodiment, the spatial constellation of the signatures/hashkeys, ie the spatial (2D) position in the image, is taken into account for an unambiguous assignment of the signatures (or the pixels).

According to one exemplary embodiment, the signatures are sorted along a stereo-epipolar line direction, ie along the lines that are also used as the search direction in the classic SGM method. An epipolar line goes through a pixel and the epipole in the respective image plane. The line connecting the projection centers of an object in the two cameras of a stereo camera pierces the respective image plane in its epipole.

In one embodiment, the matching is carried out using the hash tables in such a way that the hash keys that are most similar to one another are assigned to one another in the event that no complete match between two hash keys is found.

According to one embodiment, a metric can be used to determine the hash keys that are most similar to each other.

In one embodiment, the Hamming distance (also Hamming distance) is used as a metric. The Hamming distance is a measure of the distinctness of character strings.

According to one embodiment, as part of the hashkey correspondence search, a neighborhood comparison is used to sort out supposed correspondences (i.e. complete or extensive matches of a signature from the first image with a signature from the second image) that do not allow a clear assignment of the neighboring (side-by-side) signatures. In other words, this means the condition that a signature in the first image can only be located next to a signature in the second image that is adjacent (i.e. in the same image area) in order to be considered valid correspondence.

In one embodiment, the first and the second image have been captured simultaneously by two camera modules of a stereo camera. Any point in time that lies before the point in time at which a camera module captures the following image can be regarded as “simultaneously”.

In an alternative embodiment, the first and the second image have been captured one after the other by a moving mono camera.

According to an embodiment with a moving mono camera, the optical flow for a signature can be determined by matching the hash tables for the first and the second image.

A further aspect relates to the use of the method for determining the stereo epipolar lines that can be used as part of an auto-calibration of the camera. For this purpose, the direction of local epipolar lines is determined from the difference in the disparities in the x and y directions of neighboring pixels. The calibration of the stereo camera can thus be carried out in a small image area by the family of curves from local epipolar lines.

• - eine Empfangseinheit, konfiguriert zum Empfangen von der Stereobilddaten;
• - eine Stereobilddatenverarbeitungseinheit, konfiguriert zum;
  • b) Generieren von lokalen Signaturen in ersten und zweiten Bildern der Stereobilddaten;
  • c) Bilden von Hashkeys auf der Grundlage von einer oder mehreren Signaturen sobald eine Stärke der Signatur einen vorgegebenen Schwellwert überschreitet;
  • d) Speichern der Hashkeys in einer Hashtable für das erste und das zweite Bild mittels einer Hashfunktion derart, dass die Hashkeys von unterschiedlichen Signaturen in der Hashtable voneinander entfernt abgelegt werden, wobei die Hashkeys von weniger unterschiedlichen Signaturen näher beieinander abgelegt werden und die Hashkeys von stärker unterschiedlichen Signaturen weiter entfernt voneinander abgelegt werden;
  • e) Durchführen einer Hashkey-Korrespondenzsuche durch Matchen der Hashtable des ersten und des zweiten Bildes derart, dass eine eindeutige Zuordnung der Signaturen aus erstem und zweitem Bild resultiert; und
• - eine Ausgabeeinheit, konfiguriert zur Ausgabe der 3D-Informationen auf der Grundlage der zugeordneten Signaturen.

A device for evaluating stereo image data is included

• - a receiving unit configured to receive the stereo image data;
• - a stereo image data processing unit configured to;
  • b) generating local signatures in first and second images of the stereo image data;
  • c) forming hash keys based on one or more signatures as soon as a strength of the signature exceeds a predetermined threshold value;
  • d) Storing the hashkeys in a hashtable for the first and the second image using a hash function such that the hashkeys from different signatures are stored away from each other in the hashtable, with the hashkeys from less different signatures being stored closer together and the hashkeys from stronger different signatures are placed further apart;
  • e) performing a hash key correspondence search by matching the hash table of the first and second images in such a way that the signatures from the first and second images are unambiguously assigned; and
• - an output unit configured to output the 3D information based on the associated signatures.

The device can in particular have a microcontroller or processor, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) and The like include more and software for performing the corresponding method steps.

The invention also relates to a computer program element which, when a device or individual units of the device is or are programmed with it, instructs the device to carry out a method for evaluating image data from the stereo camera.

The invention further relates to a computer-readable storage medium on which such a program element is stored.

The present invention can thus be implemented in digital electronic circuitry, computer hardware, firmware or software.

A significant advantage of the method and the device is its efficiency. Thanks to the hashing mechanism used, the correspondences can be determined faster than with classic, correlation-based methods that have to systematically scan the image space in two dimensions of a local neighborhood.

Exemplary embodiments and figures are explained in more detail below.

• 1 schematisch eine Erfassung von Stereobilddaten von Objekten durch eine Stereokamera,
• 2 zwei von der Stereokamera zeitgleich erfasste Bilder,
• 3 ein erstes und ein zweites Bild einer Stereokamera mit Korrespondenzen,
• 4 ein Beispiel einer Korrespondenzsuche,
• 5 die Generierung von Signaturen und das Befüllen der Hashtable für ein Bild, und
• 6 schematisch das Matchen von Signaturen aus zwei Bildern durch das Vergleichen von Einträgen in Hashtables.

Show it:

• 1 schematically a recording of stereo image data of objects by a stereo camera,
• 2 two images captured simultaneously by the stereo camera,
• 3 a first and a second image of a stereo camera with correspondence,
• 4 an example of a correspondence search,
• 5 generating signatures and populating the hashtable for an image, and
• 6 Schematically the matching of signatures from two images by comparing entries in hashtables.

As an exemplary embodiment, stereo image data from a stereo camera is assumed in the following description of the figures.
When used for a mono camera, images taken at different times and in different places are correlated with one another. Here, the distance traveled between the two recordings is required as additional information, which can be provided, for example, via the vehicle's odometry. Alternatively, an auto-calibration as for a stereo camera system is also possible if sufficient correspondence between the two recordings can be established.

1 shows schematically the detection of objects 3, 4 by two camera modules 11, 12 of a stereo camera in a plan view.
The left stereo camera module 11 is a base width b away from the right stereo camera module 12. The optical axis of the left stereo camera module 11 is represented schematically by the straight line a1, the optical axis of the right stereo camera module 12 by the straight line a2. The detection or viewing range of the left-hand stereo camera module 11 is symbolized by the straight lines s1 shown as dotted lines, and that of the right-hand stereo camera module 12 by the dotted straight lines s2. The distance of the circular object 3 to the stereo camera 11, 12 (perpendicular to the line that indicates the base width b) is z.

2 shows schematically the two images that the stereo camera modules 11, 12 in the situation described (cf. 1 ) capture.
The two objects 3, 4 are different in the two images 21, 22 due to the different position of the two camera modules in the horizontal direction. In the first image 21 (shown on the left) of the left stereo camera module 11, both objects 3, 4 appear further to the right. The horizontal image distance of the circular object 3 in the first image 21 measured from the left edge of the image is dL.

In the second image 22 (shown on the right) of the right stereo camera module, both objects 3, 4 appear further to the left than in the first image 21. The horizontal image distance of the circular object 3 in the second image 22 measured from the left edge of the image is dR.

Anhand von 1 und 2 wird ersichtlich, dass die Disparität d von der Entfernung z und der Basisbreite b abhängt.The displacement of the circular object 3 between the first and second images 21, 22 is the disparity d, the following applies: $$\text{i.e}=\left|\text{dL}-\text{dr}\right|.$$

Based on 1 and 2 it can be seen that the disparity d depends on the distance z and the base width b.

Stereo algorithms determine correspondences to the same objects 3, 4. The disparity d is determined from the displacements dL and dR in which the object 3 appears in the images. The depth z of the object 3 can be inferred from the disparity d.

3 shows schematically a first 31 and a second 32 image of a stereo camera that is installed in a vehicle and captures the surroundings in front of the vehicle. Correspondences 33 between pixels or features of the first and second images are sought in the overlapping area.
In order to simplify the search for correspondence, a precise rectification of the camera images 31, 32 is required. With perfect rectification, the search for correspondences 33 can be carried out on image lines - that is, in one dimension. However, this requires a precise calibration of the two camera modules 11, 12. If the correspondence search has to be carried out in an image area (that is to say two-dimensionally), the computing requirements increase sharply.

4 illustrates an embodiment of the correspondence search. A first partial area 41 is defined in the first image 31 and a partial area 42 corresponding thereto is defined in the second image 32 . These partial areas 41, 42 lie in the overlapping area of the viewing areas of the first and second cameras 11, 12 and are in 4 shown enlarged below. In the first subarea 41, three signatures 411, 412, 413 are generated. Correspondences of these signatures 411, 412, 413 are searched for in the second partial area 42. The result is the positions of corresponding signatures 411, 412, 413 in the first image 31 (or in the first partial area 41) and the positions (of the found correspondences) of signatures 421, 422 in the second image 32. If no correspondence 33 becomes a local Signature 413 of the first tens sub-area 41 is found in the second sub-area 42, an "ndef" is generated at the position of the signature in the sub-area, which means that no correspondence could be found ("not defined")

5 schematically explains the generation of signatures and the filling of a hash table for the first image 31 or the first partial area 41. Instead of conventional correlations in the two-dimensional image, the correspondence search is generated here using signatures and hash tables.

First, local signatures in the first image 31 are generated. For illustration is in 5 the first partial area 41 of the first image 31 is shown. However, the signatures can be generated for the entire first image 31 or the entire overlapping area of the first image 31 (with the second image 32). The signatures are a combination of local features in the image structure. A signature can also correspond to a clearly defined feature. The features are, for example, points, corners, edges or two-dimensional textures, which are calculated with suitable operators.
One or more signatures form a hashkey as soon as the strength of the signature exceeds a threshold.

The hash key is entered into a hash table using a hash function. Attributes such as signature type, feature (Feature#) and position (Pos x,y) are saved. The signatures can be entered in the hashtable as a digital value under an integer index, taking into account the position in the image. The Hashtable is designed very large, so that the allocation of entries is very small (English: "sparse representation"). The hash function ensures that different signatures are stored far apart and similar signatures are close together.

6 schematically explains the matching of signatures from two images 31, 32 (or partial areas 41, 42) by comparing entries in hash tables.

First, signatures and hash keys are also generated for the second image and can be entered in a second hash table, for example. Alternatively, suitable entries can be searched for directly for each signature generated in the second image 32 in the hash table of the first image using a “matcher” (correspondence searcher).
In the case of two generated hash tables, correspondence 33 for the hash keys from the first image 31 can be efficiently determined by querying the hash table with the hash keys from the second image 32 . In many cases, perfect matches are not made by the matcher. The most similar correspondences 33 are then used. A suitable metric (eg Hamming distance) is used for this purpose.
The hash key correspondence search produces a large number of possible matches (ie possible correspondences 33) between positions in the first and second image 31, 32. The Hamming distance of the digital signatures can be used as the distance measure.

For a clear assignment of the pixels, it is helpful to consider the spatial constellation of the signature/keys. For this purpose, sorting the signatures along the stereo-epipolar line direction is expedient.

A neighborhood comparison sorts out matches that do not allow a clear assignment of the neighboring (neighboring) signatures. At the end of the neighborhood comparison, all of the signatures between the two stereo images 31, 32 are unambiguously assigned. The sorting of the signatures along the stereo epipolar lines is thus achieved.

The hashkeys contain, among other things, the positions of the features in the individual images. For matching hash keys, the matcher can determine the associated disparity via the positions and from this, using the known camera parameters, the distance d or the distance z to the feature can be determined. Here, the similarity between 2 hash keys is determined using a similarity measure, such as the Hamming distance.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• WO 2009/026433 A1 [0010]
• WO 2007/049282 A1 [0010]
• WO 2007/049282 A2 [0011]
• DE 102016217450 A1 [0018]
• WO 2017/028848 A1 [0018]

Claims (12)

Method for evaluating stereo image data, comprising the steps: a) receiving the stereo image data; b) generating local signatures (411, 412, 413, 421, 422) in first and second images (21, 22; 31, 32; 41, 42) of the stereo image data; c) Forming hash keys based on one or more signatures (411, 412, 413, 421, 422) as soon as a strength of the signature exceeds a predetermined threshold value; d) Saving the hash keys in a hash table for the first and the second image (21, 22; 31, 32; 41, 42) using a hash function in such a way that the hash keys have different signatures (411, 412, 413, 421, 422) placed farther apart in the hashtable, with hashkeys from less disparate signatures (411, 412, 413, 421, 422) placed closer together and hashkeys from more disparate signatures (411, 412, 413, 421, 422) further apart be discarded from each other; e) performing a hash key correspondence search by matching the hash table of the first and second images in such a way that the signatures (411, 412, 413, 421, 422) from the first and second images (21, 22; 31, 32; 31, 32; 41, 42) results; f) outputting 3D information based on the associated signatures (411, 412, 413, 421, 422). procedure after claim 1 , the spatial constellation (Pos x, y) of the signatures (411, 412, 413, 421, 422) being taken into account for a clear assignment of the signatures (411, 412, 413, 421, 422). procedure after claim 2 , where the signatures (411, 412, 413, 421, 422) are sorted along a stereo-epipolar line direction. Method according to one of the preceding claims, wherein the hashtables are matched in such a way that the hashkeys that are most similar to one another are assigned to one another in the event that no complete match between two hashkeys is found. procedure after claim 4 , where a metric is used to determine the most similar hashkeys. procedure after claim 5 , where the Hamming distance is used as the metric. Method according to one of the preceding claims, wherein in the context of the hash key correspondence search, supposed correspondence (33) is sorted out by means of a neighborhood comparison for which there are no adjacent signatures (411, 412, 413, 421, 422) in the first and second image (21, 22 ; 31, 32; 41, 42) can be found. Method according to one of the preceding claims, wherein the first and the second image (21, 22; 31, 32; 41, 42) have been captured simultaneously by two camera modules (11, 12) of a stereo camera. Procedure according to one of Claims 1 until 7 , wherein the first and the second image (21, 22; 31, 32; 41, 42) have been captured in chronological succession by a moving mono camera. procedure after claim 9 , wherein the optical flow for a signature (411, 412, 413, 421, 422) is determined by matching the hash tables for the first and the second image (21, 22; 31, 32; 41, 42). Use of a method according to one of the preceding claims for determining the stereo epipolar lines as part of a calibration of the camera. Device for evaluating stereo image data, comprising - a receiving unit, configured to receive the stereo image data; - a stereo image data processing unit configured to; b) generating local signatures (411, 412, 413, 421, 422) in first and second images (21, 22; 31, 32; 41, 42) of the stereo image data; c) Forming hash keys based on one or more signatures (411, 412, 413, 421, 422) as soon as a strength of the signature exceeds a predetermined threshold value; d) Saving the hash keys in a hash table for the first and the second image (21, 22; 31, 32; 41, 42) using a hash function such that the hash keys have different signatures (411, 412, 413, 421, 422) are placed farther apart in the hashtable, with hashkeys from less dissimilar signatures (411, 412, 413, 421, 422) being placed closer together and hashkeys from more dissimilar signatures (411, 412, 413, 421, 422) be placed farther apart; e) performing a hash key correspondence search by matching the hash table of the first and second images in such a way that the signatures (411, 412, 413, 421, 422) from the first and second images (21, 22; 31, 32; 31, 32; 41, 42) results; and - an output unit configured to output 3D information based on the associated signatures (411, 412, 413, 421, 422).

Images