WO2021093946A1

WO2021093946A1 - A computer assisted method for determining training images for an image recognition algorithm from a video sequence

Info

Publication number: WO2021093946A1
Application number: PCT/EP2019/081154
Authority: WO
Inventors: Kay Talmi; Matthias Kirchner
Original assignee: Car.Software Estonia As
Priority date: 2019-11-13
Filing date: 2019-11-13
Publication date: 2021-05-20

Abstract

In summary, the invention relates to a method for computer-implemented recognition and/or recognition of at least one object (Obj_i, i=1,_…,n) in an image (P_i, I=1,_…,n). Preferably, image sections (ROI) are determined by a user (U) and/or an image-recognition algorithm (KI). The determined image sections (ROI) are preferably collected and/or displayed in a gallery (G). The gallery (G) shows the image sections (ROI), whereby, the respective image section (ROI) is a section of the respective image (P_i), where an object (Obj_i) is shown or at least assumed. Verification of the appearance of an object (Obj_i) in the respective image section (ROI) is performed by a user (U) and/or an image-recognition algorithm (KI), by using the representation of the gallery (G). Image sections (ROI), not comprising the assumed object (Obj_i) are detected and preferably eliminated. The position and/or the size of the respective image section (ROI) and optionally other attributes (A1, A2, A3) of the respective object (Obj_i) are provided by the output information (D). The output information (D) is preferably a text file which can be assigned to the at least one image (P_i).

Description

A COMPUTER ASSISTED METHOD FOR DETERMINING TRAINING IMAGES FOR AN IMAGE RECOGNITION ALGORITHM FROM A VIDEO SEQUENCE

Description

The invention relates to a method for computer-implemented recognition or verification of at least one object, a computer program, a computer-readable memory unit and a driver-assistance system.

Algorithms based on artificial algorithms are widespread in modern applications.

In order to train algorithms based on artificial intelligence, labelling tools play an im portant rule. Labelling tools are used to define the position and/or the size of a rele vant object in an image or a video. In other words, labelling tools are used to classify the objects in an image or a video. The algorithms are trained by using the classified objects.

For example, cars, street signs, and/or pedestrians are classified in images or in a video using labelling tools.

Classification means that the position and/or the size of an object in an image is de fined. In general, classification means that a user looks at an image and define an ob ject by using a frame. Frames are also often called bounding boxes. The position and/or the size of the frame is assigned to the respective image.

A disadvantage of the present labelling tools is the recognition of the position and/or the size of the objects are easy to achieve.

Some labelling tools itself using algorithms, based on artificial intelligence. Due to a low certainty of the results of these labelling tools, verification is often necessary. In most cases, a verification by a human is necessary. Verification by a user is often very time consuming and therefore an expensive task.

US 6,046,740 shows a method for detecting an object in an image.

Also, US 2018/0300576 A1 describes a procedure in the field of labelling tools. It is an object of the invention to present a method for better labelling results.

It is also an object of the invention, to achieve a better driving assistance system, es pecially for vehicles.

To achieve a first aspect of the invention, a method according to claim 1 relates to the above raised issue.

A computer program performing the method achieves also at least one aspect of the invention. Also, a computer-readable storage medium, containing this computer pro gram relates to these aspects.

Furthermore, a driver assistance system according to claim 15 relates to the aspects of the invention.

Advantageous designs and further developments of the inventions are presented by the dependent claims.

The method for computer-implemented recognition or verification of at least one object in an image, preferably an image of a video, comprises at least the following steps:

- Selection of at least one image,

- selection of an image section of the image, whereby the image section is chosen by the assumption that the image section shows the respective object,

- display the at least one image section of the at least one image in a gallery,

- verification of the existence of the respective object in the respective image section by a user and/or by an image recognition algorithm,

- Assignment of the at least one object or a position of the image section (ROI) to the respective image and/or the video.

Preferably, the images or the video with the verified objects are made available to an other computer-implemented procedure, in particular for training an algorithm based on artificial intelligence.

Preferably, the method provides output information, wherein the output information comprising an identification of the respective image, the respective image and/or the identification of the image on the video. The output information can be provided in the form of a text file, comprising the identification of the image, the position and optionally the size of the image section or the respective object. Furthermore, the output infor mation preferably comprises at least one attribute of the respective object.

Preferably, the position and the size of the image section can be visualized in the im age by a frame. The frame, also called bounding box, is located around the verified object or the assumed object.

Optionally, the output information can have assignments of objects or image sections to other objects or other image sections. For example, these assignments could relate to similar objects (e.g. pedestrians) in an image.

Preferably, the video and/or the at least one image is stored in a first memory area of the memory (RAM) of a computer.

Preferably, the respective image areas are stored in a second memory area of the memory (RAM) of the computer.

Preferably, the output information is collected during performance of the method in a memory area.

A memory area is preferably understood as a memory area on a hard disc, on a main memory of the computer and/or on a graphic-card memory.

During the performance of the method, the CPU and/or the GPU of the computer are in data exchange with the memory areas.

The designation of the image in the video refers in particular to an image number of the respective image in the video. Alternatively, or additionally, the identification of the respective image can also include a time stamp indicating the time at which the respective image is displayed in the vid eo.

An object can be a vehicle, a pedestrian, an animal, a cyclist, a tree or a traffic sign.

Alternatively, depending on the application the method, the object can also be a fruit (agriculture), a pathogenic area of tissue (medicine), passage or a piece of luggage (transportation) or a cloud (meteorology).

If objects are detected manually, the objects in every 10^th or 100^th image, more par ticularly every 15^th to every 30^th image, of the video can be selected.

Preferably, the size of the image section relates to the size of the object, shown in the image. Preferably, the position of the image section relates to the position of the object or the assumed object.

The selection of the respective image section is preferably performed based on a pre defined pattern and/or a pattern recurring on the image. The pattern preferably corre spondents to the optical appearance, e.g. an edge, a shape or a color-change, of the respective object.

Preferably, the selection of the image section is performed manually by a user. There fore, the image is presented to the user. The user marks the respective objects by put ting a frame around the object. The frame preferably defines the image section.

Preferably the user is at least supported by the image-recognition algorithm. Prefera bly, the user only marks the object. The frame size and/or position of the frame is then preferably generated by the image-recognition algorithm. The gallery preferably shows a collection of image sections. One advantage of the gallery is, that e.g. a user could detect image sections without the assumed object (the so-called “false positives”) very quickly.

The method described here is preferably used to verify objects already detected in an image or a video.

Preferably, it is checked during verification, if the assumed object is shown in the re spective image section.

Image sections, not showing the assumed object are referred here as “false positive” image sections.

By performing the step of verification, so called “false positive” image sections could be recognized and optionally eliminated.

Preferably the elimination is performed by a user and/or an image-recognition algo rithm. Preferably, the verification is performed by selecting the false positive image sections.

An image section is preferably an area in the respective image. Preferably, the image section is presented by a frame shown in the image. The image section preferably contains the object as assumed.

Preferably, there could be some pre-verification. During pre-verification, statistical tendencies of the location of objects can be considered. For example, pedestrians on the street are less likely than pedestrians at the edge of the street or on a sideway.

A user and/or the image-recognition algorithm can preferably check several image sections very fast due to collected presentation of the image sections in the gallery. The gallery is preferably displayed to a user for verification by using a graphical dis play as a human machine interface (HMI), e.g. a computer screen, a web-interface, or a Virtual-Reality device.

Preferably, image sections showing or assuming similar objects are arranged adjacent to another. The gallery preferably shows similar objects from different images close to each other.

If an image section without an object is selected as a “false-positive”-case by a user or an image-recognition algorithm, this image section would preferably deleted from the gallery. Preferably, the image section is also deleted from the respective memory area and/or from the output information.

For example, a “false-positive” image section shows may show a shadow of an object.

After verification of the image sections, the verified image sections and/or the images with the marked image sections are provided in form of the output information.

One advantage of the invention is, that the user or the image-recognition algorithm can perform a labelling with very low rate of false-positive image sections. Further more, the combination of the similar image sections in the gallery leads to an easier verification. Therefore, the use of less sophisticated image-recognition algorithms is possible.

The invention could be used for traffic guidance, autonomous driving, support for driv er-assistance systems.

Preferably, at least a part of the images and/or at least a part of the image sections is used for training an algorithm based on artificial intelligence, in particularly the image recognition algorithm. The algorithm could be based on:

- a support vector machine,

- a decision tree - based algorithm, or

- an artificial neural network in particular, based on a deep learning approach.

One advantage of the before-mentioned used of the output information using for train ing algorithms, based on artificial intelligence would lead to high learning rates due to the low amount of image section without the respective object.

Preferably, the image recognition algorithm, which is intended for the recognition of the objects and/or the respective image sections, can also be trained by the output information as given by the method described here.

Preferably, the images are displayed to the user via a n human machine interface, in particular a computer screen.

Preferably, the human machine interface shows the gallery. The user can choose the at least one image section qualified as “false positive” and remove these image sec tions by choosing. Furthermore, besides the gallery, the screen shows the respective image. The user is preferably able to see the image sections as well as the respective image for comparison.

Due to the presentation of the gallery to the user, the user can see verify the objects in the image sections and/or see how well the image-recognition algorithm works.

Preferably, the image section is selected by a user or an image recognition algorithm.

The selection could be achieved by positioning a frame around the respective object. The position and the size of the respective frame preferably defines the position and/or the size of the image segment. It is one advantage of the image-recognition algorithm that the image section can be selected in a simple and fast way. Due to the verification, a false positive image sec tion is generally eliminated, even an image-recognition algorithm with a high error rate could be suitable for this kind of verification.

Preferably, the user verifies the image sections selected by the image-recognition al gorithm randomly.

Especially due to the verification, the quality of the image-recognition algorithm could be evaluated easily.

Preferably, the image sections in the gallery are arranged such that:

The image sections which are assumed to show matching objects in different images are arranged in a following order

And/or

Different objects from an image are arranged in a following order.

Preferably, matching image sections show the probably same object in different imag es.

Preferably, the gallery shows similar types of objects in a group.

Preferably, image sections of one image are arranged in columns.

Preferably, the gallery shows image sections showing the same object arranged in rows.

The image section of the same object in different images can classified by similar po sition of the respective image section. Furthermore, changing positions of objects in tends to be continuous. Therefore, according to the positions of at least two image sections, the position and/or the size of the image section of the next image can be calculated and therefore selected.

Preferably, the image section is displayed as a frame in the image.

The frame is preferably a rectangle. According to the type of the object, the frame can be shown in a respective color. Furthermore, the frame preferably matches the size of the object as represented in the image.

Frames are also called as “bounding boxes”.

By using visible frames in the images around the objects, the user is able to see if every object is selected in the respective image.

Preferably, the user and/or the image recognition algorithm select the image sections from the gallery not showing the assumed object.

Preferably, the user detects the at least one image section, which does not show the assumed object. In order to verify the respective object, shown in the gallery, a mouse-click on the respective image section preferably eliminates the respective im age section. So, false-positive image sections could easily be eliminated

Preferably, the elimination of the image section also deletes the image section in the respective memory area.

By selecting and eliminating these image sections, defined as false-positive image sections, a verification of the labelling procedure becomes easier in comparison to the state of the art. Preferably, at least one attribute being assigned to the respective object verified in the respective image section, whereby in particular the at least one attribute is being as signed while the display of the respective image section in the gallery.

Possible attributes are the type of a vehicle, the view of the vehicle, e.g. from rear, from the front, from the left/right side; an animal, e.g. an elephant, a dog or a cat; a traffic sign, e.g. a stop-sign or a speed-limit sign, or a one-way-street sign.

Preferably, the user choses the respective attribute during the presentation of the re spective image section. Preferably, possible attributes could be chosen from a menu. The menu preferably opens, if the respective image section is chosen. After the chose-procedure of at least one attribute, preferably the menu closes until the next image section is chosen.

The combination of verification and the assignment of an attribute to the respective object in the respective image section preferably reduces the time for the labelling process even further.

Preferably a position of the image section showing a first object in a first image is de termined, wherein the image section in a subsequent image is arranged at the same position and wherein the user and/or the image recognition algorithm adapts the posi tion of the image section to the position of the respective object in the subsequent im age.

In most cases, images of a video are rather similar if the images are shown in a short period of time. Therefore, if an object is shown at a first position in a first image, the position of the object would be in a rather similar position in a subsequent image.

Preferably, the user or the image-recognition algorithm can just adjust the position and/or the size of the image section to the size and/or position of the respective object in the subsequent image. Especially the arrangement and the adaption of the position and/or the size of the re spective image section is preferably easier than the detection of the respective object.

Preferably, the at least one verified object, the respective image section and/or the respective attributes are assigned to another, wherein the assignment optionally is saved in a file.

The assignment of the respective attribute is preferably done by assigning the memory area of the respective attribute to the memory area of the image section showing the respective object and/or the respective image.

Preferably, the file contains the output information as described before. The assign ment of the respective attribute may be assigned to the position of the image section of the respective image.

Preferably, the output information is provided in form of a text file, e.g. in ASCII format.

Preferably, one column comprises the image identification, the position and/or size of the at least one image section, and optionally the at least one attribute of the respec tive image section or object.

The combination of the above-mentioned parameters in one file is easily assignable to the respective image and/or the video.

Preferably, the image recognition algorithm is based on artificial intelligence.

Preferably, the image recognition algorithm could be trained by using the images, the image sections and/or the output information provided by the method as described here. Preferably, the image recognition algorithm is carried out on a first computer / first CPU and trained by using second computer / second CPU, whereby CPU stands for Central Processing Unit.

By using an image recognition algorithm based on artificial intelligence, the improve ment of the object recognition can be used to improve the performance of the method described herein.

Preferably, the recognized and verified image sections are used to train an algorithm based on artificial intelligence in order to use the algorithm for autonomous driving and/or supporting a driver assistance system.

Algorithms based on artificial intelligence are often applied in autonomous driving or many driver-assistance systems. In order to achieve suitable results, these algorithms should be trained by verified objects in image sections.

In addition, the images or videos most driver-assisted systems provide could prefera bly provide to the method as outlined above.

One aspect of the invention is a computer program for executing on a computer, wherein the computer comprises a human resource interface, a memory and a pro cessor, e.g. a CPU and/or a GPU, wherein the computer program is adapted to per form a method as outlined above.

Preferably, the computer is a decentralized server, whereby the HMI is preferably connected via a computer network like an intranet or the Internet.

Preferably, also the memory is connected via this computer network.

The connection of the HMI, where the user performs his or her tasks, for example veri fy the object in the respective image section, could preferably take place far away of the decentralized server. Another aspect of the invention is a computer-readable data storage medium, in par ticular an optical, magnetically or electronic unit, comprising the computer program as outlined above.

The computer-readable storage data storage medium comprises the computer pro gram in an object code. Alternatively, the data storage medium is connected to the internet, whereby a decentralized server is able to load the computer program and/or run the computer program.

A Driver-assistance system for a vehicle is another aspect of the invention. The driver- assistance system comprises an algorithm, based on artificial intelligence, wherein the algorithm is trained by using the recognized and/or verified objects provided by the method as outlined above.

Preferably, the driver-assistance system comprises an artificial neural network. Pref erably, the artificial neural network is trained by using objects verified by the method outlined above.

In summary, the invention relates to a method for computer-implemented recognition and/or recognition of at least one object in an image. Preferably, image sections are determined by a user and/or an image-recognition algorithm. The determined image sections are preferably collected and/or displayed in a gallery. The gallery shows the image sections, whereby, the respective image section is a section of the respective image, where an object is shown or at least assumed. Verification of the appearance of an object in the respective image section is performed by a user and/or an image- recognition algorithm, by using the representation of the gallery. Image sections, not comprising the assumed object are detected and preferably eliminated. The position and/or the size of the respective image section and optionally other attributes of the respective object are provided by/to the output information. The output information is preferably a text file which can be assigned to the at least one image. In the following, the invention is described and explained in more detail by reference to a figure. The figures show only aspects of the invention, therefore, the figures and/or the description do not limit the scope of the invention in any way.

It is shown in:

FIG 1 an exemplary representation of the method,

FIG 2 a further exemplary representation of the method,

FIG 3 images with image sections, showing an object,

FIG 4 an exemplary representation of the gallery,

FIG 5 a representation of the memory areas for objects, image sections and/or images;

FIG 6 an exemplary surface of the computer program, and

FIG 7 an exemplary computer-readable storage medium.

FIG 1 shows an exemplary representation of the method. In a first step, at least one image P_i (i=1,2,...n) is presented. The respective image P_i may be a screenshot from a video V. The images P_i shows preferably objects ObjJ (i=1,2,...,n), whereby the respective Object ObjJ is located or assumed in an image section ROI. The imag es PJ are provided for the method, whereby the method is performed by a computer program CP. The computer program CP preferably comprises an interface for the intake of the im ages P_i or the video V. After providing the image to the computer program CP, ob jects ObjJ are detected by a user U or an image-recognition algorithm Kl. The detec tion of the objects ObjJ leads to image sections ROI, whereby the respective image section ROI is assumed to show one object ObjJ. The respective image section ROI is preferably verified, if the respective image section ROI shows the assumed object ObjJ. The verification VER is also described above.

The verification VER may be performed by a presentation of a collection the respec tive image sections ROI by using the gallery G. A user U and/or an image-recognition algorithm Kl preferably select the image sections ROI not showing the assumed object ObjJ.

The images PJ, the size and/or the position of the respective image section ROI are preferably provided as output information D. The output information D can also be a text file, comprises a reference of the respective image PJ, the position and/or size of the image section ROI and optionally the type of the respective object ObjJ.

Optionally, the output information D can also be assigned to the video V or to the im ages PJ.

FIG 2 shows a further exemplary representation of the method. The images PJ are provided to the computer program CP. The computer program CP provide a gallery G, whereby the gallery G comprises the image sections ROI. The respective image sec tion ROI show a section of the image PJ, wherein the respective an object ObjJ is assumed. A user U or an image-recognition algorithm Kl detects these image sections ROI, which do not show the assumed object. In FIG 2, the objects are presented as shaded rectangles inside the respective image section ROI.

Image sections ROI, which do show the assumed object ObjJ are selected by the us er U and/or the image-recognition algorithm Kl. Image sections ROI, not showing the assumed object ObjJ are also called as “false positive” FP.

Preferably, the user selects these false positive FP type image sections ROI. The false-positive FP are preferably eliminated from the output information D.

Preferably, the gallery G shows the same object Obj_1 in subsequent images ar ranged in columns.

Preferably, the gallery G shows the at least one image-section ROI of the respective image P_i arranged in a row.

After the user U and/or the image-recognition algorithm Kl selected and optionally eliminated all false positive FP image sections ROI from the gallery G, the computer program CP combines the remaining image sections ROI and provide the output in formation D, preferably in a text file.

FIG 3 shows images P_i (1=1, ...,n) with image sections ROI, showing an object ObjJ (i=1 ). The first image P_1 shows the object P_1 at a first position. The user U or an image-recognition algorithm Kl recognized the object Obj_1 and defined an image section ROI, comprising the area of the image P_1, where the object P_1 is shown.

A subsequent image P_2 also shows the object Obj_1 at a slightly changed position.

In order to scan the subsequent image P_2 for objects ObjJ, the user U and/or the image-recognition algorithm Kl starts by setting the image section ROI at the same position as at the first image P_1.

The recognition of the object Obj_1 is more effective because the image section ROI preferably must change its position slightly. The minor change of the position of the image section ROI in order to recognize the object Obj_1 preferably is more effective. Also for further subsequent images P2,...,P_n, the image section ROI can be placed at the same position, whereby the image-recognition algorithm Kl and/or the user U easily can adapt the position of the image section ROI to match the object Obj_1 or the assumed object Obj_1.

Preferably, the position of the respective image section ROI also can be changed con secutively with the subsequent images P_i. Preferably, the position of the respective image section ROI can be repositioned in a continuous manner. The reposition is preferably performed by the image-recognition algorithm Kl.

FIG 4 shows an exemplary representation of the gallery G. The figure shows three image sections ROI, whereby the image section ROI in the lower row is “false positive” image section. The upper row shows two image sections ROI, where the two image sections each show the assumed object ObjJ (i=1). By using a menu M, a user U and/or an image-recognition algorithm Kl is able to assign at least one attribute A1,

A2, A3 to the respective object ObjJ. The assignment of the attribute A1 , A2, A3 is preferably performed by clicking on the respective attribute in the menu M with a mouse cursor MZ (see e.g. FIG 6).

FIG 5 shows a representation of the memory areas MEM1 , MEM 2 for objects ObjJ, image sections ROI and/or images PJ. The memory MEM is preferably a computer memory, e.g. RAM as main memory or memory assigned to the GPU (Graphic Pro cessor Unit).

The images PJ (i=1 ,...,n) are preferably stored in a first memory area MEM1. Also, the image sections ROI and/or the size and/or position of the respective image sec tions RIO are preferably stored in a second memory MEM2 area. The second memory area MEM2 preferably stored the gallery G in terms of the collection of image sections ROI. Each false-positive FP image section ROI, which is detected by the verification procedure of the method described above, can also be deleted from the memory. Preferably, only the position and optionally the size of the respective image area is deleted from the memory MEM. Preferably, there are further memory areas for the output information D and/or for the attributes A1 , A2, A3 and their assignment to the respective image sections ROI or the objects ObjJ. Due to the verification VER de scribed above, less memory MEM is used by the method described above.

The memory MEM is preferably connected to the CPU and/or the GPU. Preferably, the output information D is collected during the method in the respective memory area. After the recognition and/or verification of the objects ObjJ or image sections ROI in their position and/or their size, the output information D is provided as a file. The file is preferably assigned to the video V and/or to the respective image PJ.

FIG 6 shows an exemplary surface of the computer program CP. The surface is pref erably displayed to the user using a human machine interface HMI. The human ma chine interface HMI can be a computer screen or a virtual-reality device. The surface comprises the gallery G and a menu M, assigned to preferably one image segment ROI shown in the gallery G. The menu M is used for assigning at least one attribute A1 , A2, A3 to the respective object ObjJ. The assignment is preferably achieved by clicking with the mouse cursor MZ on the respective assignment.

Preferably, the surface also shows the respective image PJ (here i=1) in order to veri fy the object ObjJ (here i=1) and or the respective image section ROI. In the image PJ, the respective image section ROI can be defined by a frame F. The frame F pref erably highlines the at least assumed object ObjJ in the respective image PJ.

FIG 7 shows an exemplary computer-readable data storage medium USB. In the fig ure, an USB-Stick is presented. The computer-readable storage medium can also be a hard-disk, an optical storage medium like a CD-ROM, a DVD-ROM, or a magnetic tape.

Claims

A method for computer-implemented recognition or verification of at least one object, a computer program and a driver assistance system Claims

1. Method for computer-implemented recognition and/or verification of at least one object (ObjJ) in an image (P_i), in particular an image (P_i) of a video (V), comprising at least the following steps:

- Selection of at least one image (P_i),

- selection of an image section (ROI) of the image (P_i), whereby the image section (ROI) is chosen by the assumption that the image section (ROI) shows the respective object (ObjJ),

- display the at least one image section (ROI) of the at least one image (PJ) in a gal lery (G),

- verification of the existence of the respective object (ObjJ) in the respective image section (ROI) by a user (U) and/or by an image recognition algorithm (Kl),

- Assignment of the at least one object (ObjJ) or a position of the image section (ROI) to the respective image (PJ) and/or the video (V).

2. Method according to claim 1 , wherein at least a part of the images (PJ) and/or at least a part of the image sections (ROI) is used for training an algorithm based on arti ficial intelligence, in particularly the image recognition algorithm (Kl).

3. Method according to claim 1 or 2, wherein the images (PJ) are displayed to the user via a human machine interface (HMI), in particular a computer screen.

4. Method according to one of the preceding claims, wherein the image section (ROI) is selected by a user (U) or an image recognition algorithm (Kl).

5. Method according to one of the preceding claims, wherein the image sections (ROI) in the gallery (G) are arranged such that:

The image sections (ROI) which are assumed to show matching objects (ObjJ) in dif ferent images (P_i) are arranged in a subsequent order and/or different objects (ObjJ) from an image (PJ) are arranged in a subsequent order.

6. Method according to one of the preceding claims, wherein the image section (ROI) is displayed as a frame (F) in the image (PJ).

7. Method according to one of the preceding claims, wherein the user (U) and/or the image recognition algorithm (Kl) select the image sections (ROI) not showing the as sumed object (ObjJ) from the gallery (G).

8. Method according to one of the preceding claims, wherein at least one attribute (A1 , A2, A3) being assigned to the respective object (ObjJ) verified in the respective image section (ROI), whereby in particular the at least one attribute (A1 , A2, A3) is being as signed while the display of the respective image section (ROI) in the gallery (G).

9. Method according to one of the preceding claims, wherein a position of the image section (ROI) showing a first object (ObjJ) in a first image (PJ) is determined, where in the image section (ROI) in a subsequent image (PJ) is arranged at the same posi tion and wherein the user (U) and/or the image recognition algorithm (Kl) adapts the position of the image section (ROI) to the position of the object (ObjJ) in the subse quent image (PJ).

10. Method according to at least one of the preceding claims, wherein the at least one verified object (ObjJ), the respective image section (ROI) and/or the respective attrib utes (A) are assigned to another, wherein the assignment optionally is saved in a file.

11. Method according to at least one of the preceding claims, wherein the image recognition algorithm (Kl) is based on artificial intelligence.

12. Use of the method according to one of the preceding claims, to train an algorithm based on artificial intelligence in order to use the algorithm for autonomous driving and/or supporting a driver assistance system.

13. Computer program (CP) for executing on a computer, wherein the computer com prises a human resource interface, a memory (MEM and a processor (CPU, GPU), wherein the computer program (CP) is adapted to perform a method according to one of the claims 1 to 11.

14. Computer-readable data storage medium (USB), particular an optical, magnetical ly or electronic unit, comprising the computer program (CP) according to claim 13.

15. Driver-assistance system for a vehicle, wherein an algorithm, based on artificial intelligence, supports the function of the system, whereby the algorithm is trained by using the recognized and/or verified objects provided by the method according to one of the claims 1 to 11.