EP4222039A1

EP4222039A1 - Optical railway detection

Info

Publication number: EP4222039A1
Application number: EP21769400.9A
Authority: EP
Inventors: Benoît Bleuzé; Benjamin Hartmann; San-yu HUANG; Christoph REINBOTHE; Albi Sema; Irina Vidal Migallon
Original assignee: Siemens Mobility GmbH
Current assignee: Siemens Mobility GmbH
Priority date: 2020-12-11
Filing date: 2021-08-26
Publication date: 2023-08-09
Also published as: WO2022122196A1; US20240034372A1; DE102020215754A1

Abstract

The invention relates to a method (100) for creating a training data set for optical railway detection with integrated obstacle detection, said method comprising the following steps: - providing (101) first images (401) of railways (403) for rail vehicles (400), each first image (401) comprising a representation of a railway (403); - providing (103) second images (407) of objects (409), each second image (407) comprising a representation of at least one object (409); - combining (105) the first and second images (401, 407); and - generating (107) third images (411) consisting of the combined first and second images (401, 407), each third image (411) comprising a representation of a railway (403) with at least one object (409), and a number of the third images (411) forming the training data set for the optical railway detection with integrated obstacle detection.

Description

description

Optical track detection

The invention relates to a method for creating a training data record for optical rail route detection with integrated obstacle detection. In addition, the invention relates to a method for training an optical rail route detection with integrated obstacle detection. Furthermore, the invention relates to a method for identifying a rail route with integrated obstacle detection by a rail vehicle and a computing unit for carrying out the method for identifying a rail route with integrated obstacle detection.

An optical track detection with integrated obstacle detection can be done by means of artificial neural networks, for which purpose a large number of digital images of different track sections with obstacles located on the track sections are required in order to successfully train the neural network. However, the occurrence of obstacles on railways is a comparatively rare event, of which there are consequently only a small number of usable images. There are currently no data sets of image recordings of railways with objects arranged on the railways that would enable training of neural networks for optical railway detection with integrated obstacle detection.

The invention is based on the object of providing a method for creating a training data set for optical rail path detection with integrated obstacle detection, which circumvents these restrictions. A further object of the invention is to provide a training method for optical rail path detection with integrated obstacle detection, a method for optical rail path detection with integrated obstacle detection and a computing unit for carrying out the method for optical rail path detection with integrated obstacle detection, each based on the training data set . This object is achieved by a method for creating a training data record for optical rail track identification and a method for training optical rail track identification, a method for determining the position of a rail vehicle and a computing unit for carrying out the method for determining position. Advantageous configurations are specified in the dependent claims.

According to one aspect of the invention, a method for creating a training data set for an optical railway path detection with integrated obstacle detection is provided, the method comprising the following steps:

providing first recorded images of rail routes for rail vehicles, each first recorded image comprising a representation of a rail route;

- Provision of second image recordings of objects, each second image recording comprising a representation of at least one object;

- combining the first and second image recordings; and

- Generation of third images from the combined first and second images, with each third image comprising a representation of a railway line with at least one object, and wherein a plurality of the third image captures form the training data set for the optical railway detection with integrated obstacle detection.

In this way, the technical advantage can be achieved that an improved method for creating a training data set for an optical railway path detection with integrated obstacle detection can be provided, the training data set being based on images that each include at least one railway path and at least one object. For this purpose, first recorded images of rail routes and second recorded images of objects are provided and combined with one another. Third images are then generated from the combined first and second images of the railways and the objects, with each third image at least one railway and at least one Object includes . A plurality of the third image recordings thus forms the training set for the optical rail route detection with integrated obstacle detection.

The method according to the invention thus allows a cost-effective generation of a training data set with railways and objects, for which no time-consuming and cost-intensive creation of images of real situations in which objects are arranged on railways have to be carried out.

Within the meaning of the application, a rail route is a rail route for rail vehicles and comprises at least two parallel rails.

According to one embodiment, the first image data includes images taken by an RGB camera of real rail routes.

As a result, the technical advantage can be achieved that an improved training data record for an optical rail path detection with integrated obstacle detection can be provided, which is based on images from RGB cameras of real rail paths. An optical railway path detection with integrated obstacle detection that is trained on such a training data set can thus preferably be operated in a rail vehicle that is equipped with RGB cameras for object detection. As a result, the simplest and most precise possible optical rail route detection with integrated obstacle detection can be provided. Such a trained optical track detection with integrated obstacle detection can be operated exclusively on RGB images and does not require any further environment sensor data for track and obstacle detection.

According to one embodiment, the first image data includes images from a LiDAR sensor and/or a stereo camera of real rail routes. In this way, the technical advantage can be achieved that by using LiDAR sensors and/or stereo cameras based on the first image recordings, a distance determination can be carried out by the correspondingly trained optical railway path detection with integrated obstacle detection. Such a trained railway track detection with integrated obstacle detection can thus be operated on images from RGB cameras, LiDAR sensors and/or stereo cameras.

According to one embodiment, the second image data includes images from an RGB camera and/or a LiDAR sensor and/or a stereo camera of real objects and/or virtually generated images of real and/or virtual objects.

As a result, the technical advantage can be achieved that a wide range of uses of the method for creating a training data set for optical rail route detection with integrated obstacle detection can be provided. In this case, the second recorded images can be recorded images from RGB cameras, from LiDAR sensors or from stereo cameras. Such a trained optical track detection with integrated obstacle detection can thus be used based on RGB image data, LiDAR image data or stereo camera image data. In addition or as an alternative, the objects can be generated virtually. In this way it can be achieved that the objects can be adapted accordingly to the later training method. In addition, any objects can be generated for which there are no image recordings of real objects.

According to one embodiment, combining the first and second image data includes:

- Randomized arranging an object of the second image recordings in the first image recording of the railway, wherein the randomized arranging a randomized positioning of the obj ects relative to the railway of the first image recording and / or a randomized orientation of the obj ects relative to the railway line of the first image recording and/or a randomized size assignment of the object relative to the railway line of the first image recording.

As a result, the technical advantage can be achieved that improved training of the optical rail path detection with integrated obstacle detection and in particular improved ob ect detection or Obstacle detection can be achieved. The randomized arrangement of the objects of the second image recordings in the first image recordings makes it possible to achieve an improved training performance of the optical rail route recognition to be trained with integrated obstacle recognition.

According to a second aspect of the invention, a method for training an optical railway path detection with integrated obstacle detection is provided, the method comprising:

- Reading in by at least one neural network of third image recordings with railway tracks and objects that were generated with the method of the preceding embodiments;

- Recognition by the at least one neural network of rail routes and/or obstacles in the third image recordings, with the neural network only recognizing objects positioned on the rail routes as obstacles, and with the at least one neural network learning recognition parameters .

As a result, the technical advantage of improved training of rail detection with integrated obstacle detection can be provided, which includes the technical advantages of the method according to the invention for creating a training data record for optical rail path detection with integrated obstacle detection.

According to one embodiment, recognizing the railway tracks and the obstacles includes a segmentation of railway tracks and objects arranged on the railway tracks in the third th image recordings, wherein the segmentation of railway tracks includes a contrasting between railway tracks and image background, and wherein the segmentation of the obj ects includes a contrasting between objects arranged on the railway tracks and image background.

In this way, the technical advantage can be achieved that the most precise possible rail route detection with integrated obstacle detection can be provided. Each pixel of the third image recordings can be assigned either to the image background, a railway track or an object via the segmentation. This allows a precise railway or Object recognition can be achieved.

According to one embodiment, the segmentation includes a semantic segmentation and/or a semantic instance segmentation of the rail routes and objects.

As a result, the technical advantage can be achieved that precise detection of rail routes and objects is made possible. About the semantic segmentation or the semantic instance segmentation, a differentiation between railway tracks and objects arranged on the railway tracks can be achieved.

Furthermore, a differentiation between different objects can be achieved via the semantic instance segmentation.

According to one embodiment, the neural network is designed as a convolutional neural network, in particular as an encoder-decoder convolutional neural network, with the neural network being trained to recognize railway tracks and obstacles.

As a result, the technical advantage can be achieved that a precise and reliable optical railway path detection with integrated obstacle detection can be provided. About that as a convolutional neural network, in particular designed as an encoder-decoder convolutional neural network, simultaneous detection of rail paths and obstacles can be provided. Due to the fact that only a neural network is used, the simplest possible optical rail route detection with integrated obstacle detection can be provided.

According to one embodiment, the optical railway path detection with integrated obstacle detection comprises two neural networks, the two neural networks each being designed as a convolutional neural network, and one neural network being trained to detect railway paths, and the respective other neural network is trained to recognize obstacles.

As a result, the technical advantage can be achieved that a simplified training process of the optical railway path detection with integrated obstacle detection is made possible. The training process can be simplified by providing two neural networks, one of which is trained in each case for detecting a railway track and one for detecting an obstacle.

According to one embodiment, the optical rail path detection with integrated obstacle detection includes a distance module, with the method also including:

- Determination by the distance module of distance from obj ects to a reference point in the third image recordings, with distance determination parameters being learned by the distance module.

As a result, the technical advantage can be achieved that an improved optical railway path detection with integrated obstacle detection can be provided, by means of which a distance determination of objects to a reference point is made possible. The reference point can be given by the positioning of the rail vehicle on the rail track. According to a third aspect of the invention, a method for rail route detection with integrated obstacle detection is provided by a rail vehicle with the following steps:

- Reading in an image recorded by a camera, the camera being arranged in a front area of the rail vehicle;

- Carrying out an optical railway path detection with integrated obstacle detection, which was trained according to the method according to one of the preceding embodiments;

- Detecting a rail track on which the rail vehicle is moving and an object arranged on the rail track as an obstacle by means of optical rail detection using the detection parameters.

As a result, the technical advantage of improved rail detection with integrated obstacle detection can be provided, which includes the technical advantages of the inventive method for creating a training data set for optical rail path detection with integrated obstacle detection and the technical advantages of the inventive method for training optical rail path detection with integrated obstacle detection .

According to a fourth aspect, a computing unit is provided which includes at least one artificial neural network and is set up to be trained according to the method for training according to the preceding embodiments and to execute the method of the preceding embodiment after the training has been carried out.

The properties, features and advantages of this invention described above and the manner in which they are achieved will become clearer and more clearly understood from the explanations of the following, highly simplified, schematic representations of preferred exemplary embodiments. Here each show in a schematic representation

1 shows a flow chart of a method for creating a

Training data set for an optical track detection with integrated obstacle detection according to one embodiment;

2 shows a graphic representation of the method for creating a training data set for optical rail path detection with integrated obstacle detection according to one embodiment;

3 shows a flow chart of a method for training an optical railway path detection with integrated obstacle detection according to one embodiment;

4 shows a graphic representation of the method for training an optical railway path detection with integrated obstacle detection according to one embodiment;

FIG. 5 shows a flow chart of a method for optical railway path detection with integrated obstacle detection according to one embodiment; and

6 shows a rail vehicle with a computing unit.

FIG. 1 shows a flow chart of a method 100 for creating a training data set for an optical rail route detection with integrated obstacle detection according to one embodiment.

The description of the method 100 for creating a training data record for an optical railway path detection with integrated obstacle detection according to the embodiment in FIG. 1 is made with reference to FIG.

In a first method step 101 , first images 401 of rail routes 403 are provided. The first recorded images 401 can be real images of an RGB camera, for example, which were recorded from real railway tracks 403 . Such a real image of a railway 403 is shown in FIG. Alternatively, the first images 401 of tracks 403 include images from LiDAR sensors or stereo cameras. In order to provide the first recorded images 401 , corresponding recordings of real rail routes 403 can be created. Alternatively, already existing data sets of rail routes 403 can be used for this purpose.

In a subsequent method step 103, second image recordings 407 of objects 409 are provided. The second recorded images 407 can include RGB recorded images of real objects 409 , LiDAR sensor recordings of real objects 409 or stereo camera recordings of real objects 409 . The recordings can be generated explicitly to create the training data set. Alternatively, already existing datasets of recordings of railway tracks can be used. Alternatively or additionally, the second image recordings 407 can include virtually generated recordings of artificially generated virtual objects 409 .

The obj ects 409 can be any objects or represent people or animals. Alternatively, objects 409 can only be geometric shapes or represent three-dimensional geometric bodies.

In a further method step 105 the first and second recorded images 401 , 407 are combined.

In the embodiment shown, the combining of the first and second recorded images 401 , 407 includes a method step 109 in which the objects 409 of the second recorded images 407 are arranged randomly in the first recorded images 401 . The randomized arrangement of the objects 409 can in this case include a randomized positioning of the objects 409 in the first recorded images 401 by the objects 409 of the second recorded images 407 being arranged at any positions within the first recorded images 401 . The randomized arrangement can also include a randomized orientation of the objects 409, in which the objects 409 are arranged in any orientation relative to the railway tracks 403 of the first images 401 are arranged. In addition, the randomized arrangement can include the objects 409 of any size being integrated into the first recorded images 401 .

Subsequently, in a method step 107 , third images 411 are generated based on the combined first and second images 401 , 407 , each third image 411 comprising at least one railway 403 and at least one object 409 .

The third images 411 can in this case include a plurality of railway tracks 403 and a plurality of objects 409 . The objects 409 can be arranged in any size, position and orientation relative to the railway tracks 403 in the third images 411 .

FIG. 2 shows a graphic representation of the method 100 for creating a training data set for an optical railway path detection with integrated obstacle detection according to one embodiment.

FIG. 2 shows a graphic representation of the creation of the training data set according to the method 100 in the embodiment in FIG.

Initially, first images 401 of rail routes 403 are provided. These can be real recordings, for example from RGB cameras, of real railway tracks 403 .

Furthermore, second image recordings 407 of objects 409 are provided, which can include both real objects 409 and synthetically generated virtual objects.

The first and second recorded images 401 , 407 are then combined and third recorded images 411 based on the first and second recorded images 401 , 407 are generated accordingly. To generate the third recorded images 411, the objects 409 of the second recorded images 407 are arranged in a randomized manner in the first recorded images 401 . As shown in FIG. 2, for this purpose the various objects 409 are arranged in any desired position, orientation and size relative to the tracks 403 of the third images 411. A realistic arrangement of the objects 409 with a realistic positioning, orientation or Size in relation to the real railway tracks 403 is not absolutely necessary in this case.

A plurality of third image recordings 411 generated in this way finally forms the training data record for the rail route recognition to be trained with integrated object recognition.

FIG. 3 shows a flow chart of a method 200 for training an optical rail path detection with integrated obstacle detection according to one embodiment.

The method 200 for training the optical railway path detection with integrated obstacle detection in the embodiment in FIG. 3 is described with reference to the description of FIG.

In a first method step 201 , third image recordings 411 with railway tracks 403 and objects 409 are read in by at least one neural network 501 . In this case, the third recorded images 411 were generated using the method 100 for creating a training data set for an optical railway path detection with integrated obstacle detection according to the method steps described above.

Subsequently, in a subsequent method step 203 , rail routes 403 and/or obstacles 413 are recognized in the third image recordings 411 by the at least one neural network. In this case, the neural network 501 only recognizes objects 409 that are positioned on the railway tracks 403 in the third image recordings 411 as obstacles 413 . Objects 409 that do not cover or cover a corresponding railway track 403 in the third image recordings 411 . cross, are not recognized as obstacles by the neural network 501 .

In this case, the neural network 501 can be trained to recognize objects 409 within the third image recordings 411 and to distinguish between objects 409 that are arranged on corresponding railway tracks 403 and objects 409 that are not arranged on railway tracks 403 .

According to one embodiment, the neural network 501 can be designed as a convolutional neural network and in particular as an encoder-decoder convolutional neural network. The neural network 501 can be designed in particular to include both railway tracks 403 and objects 409 or to recognize obstacles 413 .

Alternatively, the optical rail path detection with integrated obstacle detection can include two neural networks 501, which are each designed as convolutional neural networks, for example. In this case, one neural network 501 can be set up in each case to recognize rail routes 403, while the other neural network in each case can be set up to recognize objects 409 or to recognize obstacles 413 .

According to one embodiment, the optical railway path detection with integrated obstacle detection can also include a distance module that is set up to determine a distance of the objects 409 in the third image recordings 411 from a predetermined reference point. The predetermined reference point can be set here, for example, by the positioning of the camera, which was used to record the first images 401, relative to the railway tracks 403 or Obj ects 409 be given.

In a further method step 205, the distance module recognizes distances of the objects 409 in the third image recordings 411 from the reference point P, for example the RGB camera. By recognizing the railway tracks 403 and . Obstacles 413 in method step 203 are learned by the neural network 501 recognition parameters and thus a training process of the neural network 501 is carried out. Analogously, by determining the distances of the objects 409 by the distance module in method step 205, corresponding distance determination parameters are learned, as a result of which the distance module is trained accordingly.

The training process described can be carried out according to supervised learning or unsupervised learning, for example. In particular, the neural network 501 or the plurality of neural networks 501 are trained according to a backpropagation process that is customary in the prior art.

According to one embodiment, the detection of the railway tracks 403 and/or the obstacles 413 or of the objects 409 in method step 203, a segmentation of rail routes 403 and objects 409 or Include obstacles 413 in the third images 411 . The segmentation can, for example, be a semantic segmentation or be a semantic instance segmentation. For the segmentation, it is determined for each pixel of the third recorded image 411 whether the respective pixel belongs to a railroad track 403 , to an obstacle 413 or to the image background of the third recorded image 411 . The segmentation carried out in this way can thus be used to achieve a contrast between railroad tracks 403 , obstacles 413 and the image background of the third image recording 411 . The segmentation of the third recorded images 411 can be carried out according to segmentation processes known from the prior art.

FIG. 4 shows a graphic representation of the method 200 for training an optical railway path detection with integrated obstacle detection according to one embodiment. In FIG. 4, the training process according to the method 200 in the embodiment shown in FIG. 3 is illustrated in a graphical representation.

To train the neural network 501 or of the neural networks 501 , third image recordings 411 with railway tracks 403 and objects 409 are read in by the neural networks 501 .

During the training process, recognition parameters are learned by the neural networks 501 for recognizing the railway tracks 403 and the objects 409 arranged on the railway tracks 403 by the neural networks 501 . In the course of the training process, the neural networks 501 segment the tracks 403, obstacles 413 or of the image background of the third image recordings 411 is carried out.

FIG. 4 also shows a segmented image recording 412 in which two segmented railway tracks 404 and three segmented obstacles 414 each and a segmented image background 416 of the third image recording 411 shown in FIG. 4 are shown. The segmentation shown here enables a distinction between segmented railroad 404 and segmented obstacles 414 , so that a clear distinction between railroad and objects 409 arranged on the railroad 403 and hereby a clear recognition of obstacles 413 given by objects 409 arranged on the railroad 403 the correspondingly trained neural networks 501 is made possible.

As shown in FIG. 4, the objects 409 in the third image recording 411 which are not arranged on the railway 403 are not recognized as obstacles 413.

A detection of an arrangement of objects 409 on railways 403 can be determined here, for example, by overlapping the outline of an object 409 with the outline of a railway 403 . Alternatively, perspective criteria can be taken into account through which both for example a vertical distance of the object 409 to the railroad track 403 can be determined. As a result, objects 409, such as vegetation covering the railway 403 or birds flying over the railway 403, which are arranged on the railway 403 solely because of the perspective representation in the image recording, can be identified as non-hazardous objects and thus as obstacles 413 be excluded. This enables a more precise classification of obstacles 413 and the avoidance of false positive recognition results in which objects 409 are classified as obstacles 413 but are not arranged on the railway 403 and thus do not actually represent an obstacle to the rail vehicle.

FIG. 5 shows a flow chart of a method 300 for optical rail path detection with integrated obstacle detection according to one embodiment.

The embodiment of the method 300 for rail route detection with integrated obstacle detection by a rail vehicle 400 is described with reference to FIG.

To identify railway tracks 403 or . Obstacles 413 by a rail vehicle 400 are first read in a method step 301 images of a camera 405 of the rail vehicle 400 by the optical track detection with integrated obstacle detection and in particular the neural network 501 of the optical track detection with integrated obstacle detection. The camera 405 can be an RGB camera, for example, which is arranged in a front area 406 of the rail vehicle 400 and can see an area in front of the rail vehicle 400 in the direction of travel, in which at least the railway 403 , on which the rail vehicle 400 is moving, can be seen is .

The optical track detection with integrated obstacle detection and in particular the corresponding neural network 501 were trained according to the method 200, a corresponding training data record, which was generated according to the method 100 , was used for this purpose.

Based on the image recordings read in method step 301 , in a subsequent method step 303 an optical rail route detection with integrated obstacle detection is carried out.

Based on this, in a method step 305 the railway 403 on which the rail vehicle 400 is moving is recognized and, if necessary, an object 409 arranged on the railway 403 is identified as an obstacle 413 . The recognition of the railway line 403 or . of the obstacle 413 by the optical railway path detection with integrated obstacle detection and in particular by the correspondingly trained neural network 501 can be carried out according to the explanations for the method 200 by a corresponding segmentation of the image recordings.

If a corresponding obstacle 413 is detected, a corresponding warning signal can be emitted or a braking process of the rail vehicle 400 can be initiated.

FIG. 6 shows a rail vehicle with a computing unit.

FIG. 6 shows a rail vehicle 400 which is arranged on a rail track 403 . The rail vehicle 400 includes a computing unit 500 on which a neural network 501 is set up, which is trained in accordance with the method 200 for rail route detection and obstacle detection. The processing unit 500 can be designed as a software module or as a hardware component with a corresponding software module. In a front area 406, the rail vehicle 400 comprises a camera 405 which can see an area which can see at least the rail track 403 on which the rail vehicle 400 is being moved. The camera 405 can be an RGB camera, for example, and is connected to the processing unit 500 in terms of data technology. Using the images recorded by the camera 405, the neural network 501 can use the method 300 to calculate a railway or Obstacle detection can be carried out.

FIG. 6 also shows an object 409 which is arranged on the track 403 adjacent to the front area 406 of the rail vehicle 400 and can be viewed by the camera 405 .

According to one embodiment, the rail vehicle 400 can also have a LiDAR sensor or include an additional stereo camera, which can also be arranged in the front area 406 and are not shown in FIG. Based on the data from the LiDAR sensor or The stereo camera can be used to determine the distance between the object 409 and the rail vehicle 400 using the optical rail path detection with integrated obstacle detection.

As an alternative to this, a distance module (not shown in FIG. 6) for the optical railway path detection with integrated obstacle detection can also be set up on the computing unit 500, which is designed to carry out a distance determination of the object 409 relative to the rail vehicle 400 based on the images recorded by the camera 405.

Although the invention has been illustrated and described in detail by the preferred exemplary embodiment, the invention is not restricted by the disclosed examples and other variations can be derived therefrom by a person skilled in the art without departing from the protective scope of the invention.

Claims

patent claims

1. Method (100) for creating a training data set for an optical railway path detection with integrated obstacle detection, with the following steps:

- Providing (101) first images (401) of railways (403) for rail vehicles (400), each first image (401) comprising a representation of a railway (403);

- Providing (103) of second images (407) of objects (409), wherein each second image (407) comprises a representation of at least one object (409);

- Combining (105) the first and second image recordings (401, 407); and

- Generating (107) third images (411) from the combined first and second images (401, 407), each third image (411) comprising a representation of a railway (403) with at least one object (409), and wherein a The majority of the third images (411) form the training data record for the optical rail route detection with integrated obstacle detection.

2. The method (100) according to claim 1, wherein the first images (401) include images of an RGB camera of real rail routes (403).

3. The method (100) according to claim 2, wherein the first images (401) include images of a LiDAR sensor and/or a stereo camera of real rail routes (403).

4. The method (100) according to any one of the preceding claims, wherein the second images (407) images of an RGB camera and / or a LiDAR sensor and / or a stereo camera of real objects (409) and / or virtually generated images of real and/or virtual objects (409).

5. The method (100) according to any one of the preceding claims, wherein the combining (105) of the first and second image data comprises:

- Randomly arranging (109) an object (409) of the second image recordings (407) in the first image recording (401) of the railway (403), the randomized arranging a randomized positioning of the object (409) relative to the railway (403) of the first Image recording (401) and/or a randomized orientation of the object (409) relative to the railway (403) of the first image recording (401) and/or a randomized size assignment of the object (409) relative to the railway (403) of the first image recording (401) includes.

6. A method (200) for training optical rail path recognition with integrated obstacle recognition, comprising:

- Reading (201) by at least one neural network (501) of third images (411) with railway tracks (403) and objects (409) that were generated with the method (100) of claims 1 to 5;

- Recognition (203) by the at least one neural network of railway tracks (403) and/or obstacles (413) in the third image recordings (411), whereby the neural network (501) detects objects (409 ) are recognized as obstacles (413), and the at least one neural network (501) learns recognition parameters.

7. The method (200) according to claim 6, wherein the detection of the railways (403) and the obstacles (413) involves a segmentation of railways (403) and objects (409) arranged on the railways (403) in the third images (411) includes, wherein the segmentation of railways (403) comprises a contrast between railways (403) and image background, and wherein the segmentation of the objects (409) comprises a contrast between on the railways (403) arranged objects (409) and image background.

8. The method (200) according to claim 7, wherein the segmentation comprises a semantic segmentation and/or a semantic instance segmentation of the rail routes (403) and objects (409).

9. The method (200) according to any one of claims 6 to 8, wherein the neural network (501) is designed as a convolutional neural network, in particular as an encoder-decoder convolutional neural network, and wherein the neural network (501) is trained , tracks (403) and obstacles (413) to recognize.

10. The method (200) according to any one of claims 6 to 8, wherein the optical railway path detection with integrated obstacle detection comprises two neural networks (501), wherein the two neural networks (501) are each designed as a convolutional neural network, and wherein a neural Network (501) is trained to recognize railway tracks (403), and the respective other neural network (501) is trained to recognize obstacles (413).

11. The method (200) according to any one of claims 6 to 10, wherein the optical rail path detection with integrated obstacle detection comprises a distance module, and wherein the method further comprises:

- Determination (205) by the distance module of distance of objects (409) to a reference point in the third image recordings (411), wherein distance determination parameters are learned by the distance module.

12. Method (300) for rail path detection with integrated obstacle detection by a rail vehicle (400) with the following steps:

- Reading in (301) an image recorded by a camera (405), the camera (405) being arranged in a front region (406) of the rail vehicle (400);

- Carrying out (303) an optical rail path detection with integrated obstacle detection, which according to the method 22

(200) has been trained according to any one of the preceding claims 6 to 11;

- Detecting (305) a railroad (403) on which the rail vehicle (400) is moving and an object (409) arranged on the railroad as an obstacle (413) by means of optical rail recognition using the recognition parameters.

13. Arithmetic unit (500) comprising an artificial neural network (501), wherein the arithmetic unit (500) is set up to be trained according to the method (200) for training according to claims 6 to 11 and after carrying out the training, the method (300) according to claim 12.