DE102017222600B3 - Optical position determination of a vehicle by means of a convolutional autencoder and / or a deep neural network - Google Patents

Abstract

The invention relates to a method for optically determining a position of a vehicle in which, while the vehicle is traveling with at least one exterior camera provided on the vehicle, images of surroundings of the vehicle are continuously optically detected in order to obtain the position of the vehicle in real time based on the captured images determine.

Description

The invention relates to a method for optically determining a position of a vehicle in which, while the vehicle is traveling with at least one exterior camera provided on the vehicle, images of surroundings of the vehicle are continuously optically detected in order to obtain the position of the vehicle in real time based on the captured images determine.

Recent efforts are going to develop vehicles that independently, d. H. Navigate from a starting point to a specific destination without the assistance of a driver (autonomous driving). A self-propelled vehicle must therefore dodge on the one hand while driving possible obstacles on the route and on the other hand continuously determine its position and operate his steering according to a leading to the particular destination travel route.

A first problem of a self-propelled vehicle is therefore reliable detection of a possible obstacle, such as a pedestrian, a preceding vehicle, a traffic island or a construction site mark, as well as a precise determination of the distance of the detected obstacle from the vehicle to the obstacle by steering and / or brakes. For this purpose, a combination of a plurality of sensors mounted on the vehicle for detecting different aspects and / or regions of an environment of the vehicle is generally necessary. However, the different data generated by each of these sensors must be properly combined to reliably detect and accurately detect both a kind of obstacle and a distance of the obstacle from the vehicle.

The WO 2016/100814 A1 discloses such a method for merging data from different sensors. In the process, the three-dimensional data of a LiDAR (Light Detection And Ranging) and the two-dimensional data of an outdoor camera are combined by means of a convolutional autencoder. In the merged data interesting areas and in the interesting areas in turn interesting points are evaluated by means of a deep neural network or a convolutional autencoder and marked each obstacle detected with an obstacle enclosing contour, ie framed. Relative alignment errors of the LiDAR and the outdoor camera can be compensated by training a deep neural network or a convolutional auto-encoder. However, this method does not allow to determine a position of the vehicle.

The following references address other aspects of sensors or exterior cameras of a vehicle or the processing of the images or video streams generated by the sensors or external cameras.

The WO 2016/100814 A1 discloses a method for automatically aligning multiple multimodal sensors of a vehicle. In the method, alignment errors of the sensors are detected on the basis of edges detected in respective sensor data by means of a neural network and corrected by means of actuators assigned to the sensors.

The US 2011/0267452 A1 discloses a method of compressing a video stream captured by an exterior camera of a vehicle during a ride. Each frame of the video stream is divided into several areas and each area is compressed with a relevance-dependent individual compression rate.

The WO 2017/156243 A1 discloses a method of recognizing objects in a digital two and one-half dimensional image comprising both color and depth information by means of a neural network.

The DE 10 2009 018 073 A1 discloses a method for automatically updating a geographic database for a navigation system based on an image taken by an exterior camera of a vehicle.

A second problem is the determination of a current position of the vehicle. The achievable with satellite localization systems precision in determining a position, such as the precision of the widely used in vehicles GPS system (Global Positioning System), but is completely insufficient for self-driving vehicles. On the one hand, it does not even allow a vehicle to be reliably supported on an even wide lane, much less on a relatively narrow lane on a multi-lane carriageway. On the other hand, receiving satellite-based position data due to shadowing, such as in deep canyons or in tunnels, may be impossible or severely hampered by adverse weather conditions, further reducing precision.

One possible solution to this problem is a continuous optical detection of an environment of the vehicle by means of external cameras mounted on the vehicle and algorithmic Evaluate the detected environment. In this way, the relatively inaccurate GPS position data can be specified locally.

So revealed the US 2017/0015317 A1 a method for determining a position of a vehicle on a roadway, wherein during driving on the roadway, images of a surface of the roadway in an environment of the vehicle are optically detected by outdoor cameras provided on the vehicle. An adaptive pattern recognition algorithm that compensates for static distortion due to an outside camera perspective as well as dynamic distortions due to pitching and rolling motion of the vehicle detects lane markers present on the captured images used to hold the vehicle in the lane. For training the pattern recognition algorithm, localized reference images of the surface of the roadway can be used, which are assigned to a current position of the vehicle by means of satellite-supported position data received by the vehicle. However, this method compulsorily requires lane markings on the road and, as a result, can not be used on lanes without lane markings.

Alternatively, the US 9,623,905 B2 a method for determining a position of a vehicle on a roadway, in which images of surroundings of the vehicle are continuously optically detected by exterior cameras provided on the vehicle while traveling on the roadway. Each captured image is analyzed by a deep neural network to detect typical landmarks visible from the lane, such as prominent buildings, bridges or traffic signs. A recognized landmark is aligned with a list of known and precisely located landmarks to identify it and to determine the position of the vehicle relative to the identified landmark. However, this method presupposes compellingly well-known and precisely located landmarks along the roadway and fails accordingly on lane sections with no visible landmarks extending, for example, through rural, woody or desert-like regions.

The WO 2017/053407 A1 discloses a method for determining a position of a vehicle. In the method, an image captured by an outdoor camera image of a road surface is encoded by means of a transformation into a template. The position of the vehicle is determined on the basis of position data of a template of a template database which corresponds to the coded template.

In "Robust Intensity-Based Localization Method for Autonomous Driving on Snow-Wet Road Surface" M. Aldibaja et al, IEEE Transactions on Industrial Informatics 13 (5), pp. 2369-2378, Oct. 2017 ) discloses a method for determining a lateral position of a vehicle on a wet or snowy roadway in which a compacted image is calculated by accumulating a plurality of images of a LiDAR sensor of a vehicle. An edge profile of the compressed image is compared with an edge profile of a localized image stored in an image database.

In "Absolute geo-localization thanks to hidden Markov Model and Exemplified metric learning" ( C. Le Barz et al., Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops 2015, pp. 9-17 ) describes a method in which a position of a vehicle is determined by comparing a sequence of images captured by an exterior camera of the vehicle with a sequence of images stored in an image database corresponding to roughly known positions of the vehicle by means of a local metric learned during a training phase ,

In "High-accurate vehicle localization using digital maps to coherency images" ( N. Mattern, R. Schubert, G. Wanielik, IEEE Intelligent Vehicles Symposium, San Diego, Calif., USA, June 21-24, 2010, p. 462-469 ) describes a method for determining a position of a vehicle. From information contained in a digital map, an image of an environment of the vehicle is calculated, which is compared with an image of the environment captured by an exterior camera of the vehicle.

The invention is therefore based on the object to provide an improved method for optically detecting a position of a vehicle, which eliminates the disadvantages described and has a great robustness. Moreover, it is an object of the invention to provide a system for carrying out the method provided.

• - Herstellen einer Mobilfunk-Verbindung zwischen dem Fahrzeug und einem Backend-Server mittels eines in dem Fahrzeug vorgesehenen Mobilfunkmoduls;
• - Empfangen mindestens eines Softwaremoduls zum Kodieren/Dekodieren optischer Daten von dem Backend-Server über die hergestellte Mobilfunk-Verbindung und Installieren des mindestens einen empfangenen Softwaremoduls in einem Steuergerät des Fahrzeugs;
• - Bestimmen einer Fahrtstrecke für eine Fahrt des Fahrzeugs;
• - Empfangen, von dem Backend-Server über die hergestellte Mobilfunk-Verbindung, mindestens eines Lokalisierungsstroms für jedes installierte Softwaremodul, welcher der bestimmten Fahrtstrecke zugeordnet ist, und Speichern jedes Lokalisierungsstroms in dem Steuergerät;
• - Bilden eines Positionsstroms zu jedem installierten Softwaremodul durch mehrfaches, insbesondere periodisches Ausführen der folgenden Schritte während der Fahrt des Fahrzeugs auf der bestimmten Fahrtstrecke:
  • - Erfassen eines Bildes einer Umgebung des Fahrzeugs durch eine an dem Fahrzeug vorgesehene Außenkamera;
  • - Berechnen jeweils eines zu dem erfassten Bild korrespondierenden kodierten Positionswerts mittels jedes installierten Softwaremoduls; und
  • - Speichern des berechneten kodierten Positionswerts durch das Steuergerät; und
• - Ermitteln der Position des Fahrzeugs durch fortlaufendes, insbesondere periodisches abstraktes Vergleichen des gebildeten Positionsstroms mit einem Abschnitt des empfangenen mindestens einen Lokalisierungsstroms.

An object of the present invention is a method for optically detecting a position of a vehicle. The method comprises the following steps:

• - Establishing a mobile connection between the vehicle and a back-end server by means provided in the vehicle mobile module;
• Receiving at least one software module for encoding / decoding optical data from the backend server via the established mobile radio connection and installing the at least one received software module in a controller of the vehicle;
• Determining a travel distance for a drive of the vehicle;
• Receiving, from the backend server via the established cellular connection, at least one location stream for each installed software module associated with the particular route and storing each location stream in the controller;
• Forming a position current to each installed software module by executing the following steps several times, in particular periodically, while the vehicle is traveling on the particular route:
  • - detecting an image of an environment of the vehicle by an outdoor camera provided on the vehicle;
  • Calculating a respective coded position value corresponding to the acquired image by means of each installed software module; and
  • - storing the calculated encoded position value by the controller; and
• - Determining the position of the vehicle by continuously, in particular periodically abstract comparing the position current formed with a portion of the received at least one localization current.

The position of the vehicle is determined according to the invention exclusively by abstract comparing coded and compressed optical image data. For this purpose, a continuously formed in the vehicle position current is continuously compared with a received from a back-end server and corresponding to the particular route localization current to determine that point of the localization current, which corresponds to the current position current of the vehicle. In other words, it is not important in the method according to the invention to identify objects in the position current or the localization current, ie. H. to give meaning to the position current or localization current. Rather, the comparison is made on a very high coded abstraction level. In this way, the position of the vehicle on the route can be determined very precisely and reliably, which is associated with a great robustness of the process. It is understood that instead of a cellular connection, any other wireless connection may be used without departing from the scope of the invention.

In a preferred embodiment, a correlation function is calculated for comparing the generated position current with the portion of the at least one localization current. For example, correlation functions are used in statistics and are particularly useful for finding significant matches between different sets of data. Therefore, by calculating a correlation function, the location within the localization current that corresponds to the current location in the position stream can be determined particularly reliably and easily.

In a further embodiment, satellite-based position data, in particular GPS data, are received by means of a satellite receiver provided in the vehicle, and the section of the at least one localization current is determined by means of the received position data. GPS data can reduce the overhead of comparing by providing a rough prediction of the relevant portion of the localization stream. This avoids unnecessary and time-consuming comparisons with irrelevant portions of the localization stream.

In one embodiment, from the determined position of the vehicle and at least one localization current, a length of a route section traveled by the vehicle on the route and / or a distance of the vehicle to another vehicle on the route are calculated. This assumes that the localization stream contains indications that allow a calculation of a longitudinal position. For example, a longitudinal position within the localization stream can be computed if each encoded and compressed image is assigned a time and a current velocity. Then, the longitudinal position is obtained by forming a temporal integral of the respective actual speeds since the starting point of the ride.

• - Herstellen einer Mobilfunk-Verbindung zwischen dem Messfahrzeug und einem Backend-Server mittels eines in dem Messfahrzeug vorgesehenen Mobilfunkmoduls;
• - Empfangen mindestens eines Softwaremoduls zum Kodieren/Dekodieren optischer Daten von dem Backend-Server über die hergestellte Mobilfunk-Verbindung und Installieren des mindestens einen empfangenen Softwaremoduls in einem Steuergerät des Messfahrzeugs;
• - Bestimmen einer Fahrtstrecke für eine Fahrt des Messfahrzeugs;
• - Bilden mindestens eines Lokalisierungsstroms zu jedem installierten Softwaremodul durch fortlaufendes, insbesondere periodisches Ausführen der folgenden Schritte während der Fahrt des Messfahrzeugs auf der bestimmten Fahrtstrecke:
  • - Erfassen eines Bildes einer Umgebung des Messfahrzeugs durch eine an dem Messfahrzeug vorgesehene Außenkamera;
  • - Berechnen jeweils eines zu dem erfassten Bild korrespondierenden kodierten Positionswerts mittels jedes installierten Softwaremoduls;
  • - Erfassen eines Geschwindigkeitswerts des Messfahrzeugs mittels eines in dem Messfahrzeug vorgesehenen Geschwind igkeitssensors;
  • - Erfassen eines Zeitpunkts mittels eines in dem Messfahrzeug vorgesehenen Zeitmessers;
  • - Bilden eines Lokalisierungswerts durch Verknüpfen des berechneten Positionswerts mit dem erfassten Geschwindigkeitswert und dem erfassten Zeitpunkt; und
  • - Speichern des gebildeten Lokalisierungswerts durch das Steuergerät; und
• - Senden jedes gebildeten Lokalisierungsstroms an den Backend-Server über die hergestellte Mobilfunkverbindung.

The invention also provides a method for generating at least one localization current which can be used in the above-described method during a journey of a measuring vehicle. The method comprises the following steps:

• - Establishing a mobile connection between the measuring vehicle and a back-end server by means provided in the measuring vehicle mobile phone module;
• Receiving at least one software module for encoding / decoding optical data from the back-end server via the established mobile radio connection and installing the at least one received software module in a control unit of the measuring vehicle;
• Determining a route for a journey of the measuring vehicle;
• - Forming at least one localization stream to each installed software module by continuously, in particular periodically, performing the following steps while the measurement vehicle is traveling on the particular route:
  • Detecting an image of an environment of the measuring vehicle by an external camera provided on the measuring vehicle;
  • Calculating a respective coded position value corresponding to the acquired image by means of each installed software module;
  • Detecting a speed value of the measuring vehicle by means of a speed sensor provided in the measuring vehicle;
  • Detecting a time by means of a timer provided in the measuring vehicle;
  • Forming a localization value by associating the calculated position value with the detected velocity value and the detected time; and
  • - storing the formed localization value by the controller; and
• Sending each formed localization stream to the backend server via the established cellular connection.

The method carried out in the measuring vehicle differs from the previously described method, which is carried out by a vehicle using a localization current, in that an actual speed and a traveling time of the measuring vehicle are additionally detected and integrated into the localization current during the journey. From these quantities, it is possible to calculate their longitudinal distance from the starting point of the journey for each location of the localization current.

In one embodiment, at least two localization streams are formed into an installed software module by simultaneously capturing at least two images of the surroundings of the vehicle by at least two external cameras, in particular exactly two images, by a stereoscopic camera. As a result, on the one hand, the scope of the training material for each installed software module can be increased. On the other hand, the software module can be trained to process similar, but perspectively divergent images, thereby achieving greater robustness of the method.

In a further embodiment, satellite-based position data, in particular GPS data and preferably DGPS data (Differential Global Positioning System), are received by means of a satellite receiver arranged in the measuring vehicle, and the localization value is formed by additionally combining the calculated position value with the received satellite-based position data. By linking the localization streams with satellite-based position data, the first relative, i. H. localization streams related to certain routes are given absolute terrestrial position information. In this way, satellite-based position data can be used in vehicles that can not receive satellite-based position data themselves. In addition, the precision of the position determination is further increased.

• - Empfangen des mindestens einen Lokalisierungsstroms von einem Messfahrzeug durch den Backend-Server;
• - Speichern des empfangenen Lokalisierungsstroms;
• - Bereitstellen des gespeicherten Lokalisierungsstroms zum Empfangen durch ein Fahrzeug;
• - Trainieren des mindestens einen Softwaremoduls mittels des gespeicherten Lokalisierungsstroms; und
• - Bereitstellen des mindestens einen trainierten Softwaremoduls zum Empfangen durch ein Fahrzeug.

Another object of the invention is a method for providing at least one software module for encoding / decoding optical data on a back-end server and at least one localization stream for use in one of the two methods described above. The method comprises the following steps:

• Receiving the at least one localization stream from a measurement vehicle by the backend server;
• - storing the received localization stream;
• Providing the stored localization stream for reception by a vehicle;
• - Training the at least one software module by means of the stored localization stream; and
• - Providing the at least one trained software module for receiving by a vehicle.

The main purpose of the back-end server is to train the software modules used and to provide them and localization streams for retrieval. The software modules and localization streams provided on a backend server can therefore also be used conveniently by a large number of vehicles. Basically, the vehicles only need a mobile radio connection to the backend server when starting a journey in order to possibly update a software module stored in the vehicle and to load a localization stream corresponding to the particular route from the backend server.

In a preferred embodiment of each of the methods described above, a convolutional auto-encoder is used as a first software module for encoding / decoding optical data. A Convolutional Autoencoder represents as a deep neural network implemented special form of an auto-coder (codec, codec / decoder) and serves to transform higher-dimensional input data by simple or multiple folding into low-dimension output data. Accordingly, a convolutional auto-encoder is used to one-dimensionally (vectorically) encode and compress extensive two-dimensional (matrix-like) image data. From the coded and compressed output data, the output data can be recovered by means of the convolutional autencoder with a specific reconstruction error. By comparing the output data with the input data, a convolutional autoencoder can be trained unsupervised by means of back propagation to reduce the reconstruction error, especially for relevant areas or sections of the input data. A convolutional auto-encoder thus allows to preserve the essence of input data needed for a particular use, despite significant compression in the output data, so that by means of a convolutional auto-encoder a video stream, ie a picture sequence is compressed to a position stream or localization stream of small size, which is good is manageable and also suitable for wireless transmission, such as via a mobile network.

• - Identifizieren eines Objekts, das mindestens einer vorgegebenen Objektklasse, die insbesondere Schilder, Pfosten oder Gebäude umfasst, zugeordnet ist, in dem erfassten Bild mittels des tiefen neuronalen Netzes;
• - Ermitteln einer Pixelzahl innerhalb einer minimalen, das identifizierte Objekt einschließenden, insbesondere rechteckigen Kontur durch das Steuergerät; und
• - Verwenden der ermittelten Pixelzahl als kodierter Positionswert.

In addition, it is advantageous to use a deep neural network as a second software module for coding / decoding optical data, with which the following steps are carried out for calculating a coded position value corresponding to the captured image:

• Identifying an object associated with at least one predetermined class of object, including in particular signs, posts or buildings, in the captured image by means of the deep neural network;
• - Determining a number of pixels within a minimum, the identified object enclosing, in particular rectangular contour by the control unit; and
• - Using the determined number of pixels as a coded position value.

Such a deep neural network also makes it possible to transform large higher-dimensional input data into lower-dimensioned output data and is suitable for object recognition and semantic segmentation. It is trained by means of annotated objects, which are thus assigned to certain object classes. As a result, a deep neural network can recognize objects trained in image data and associate the respective object classes with them. In other words, not only the existence but also the type of a relevant object in the image data can be determined by means of such a deep neural network. For example, recognized objects may be enclosed by outlines, such as a rectangular frame, and the varying number of pixels enclosed by the contours, depending on the position of the vehicle, may be used as the localization stream.

A subject of the invention is furthermore a system for optically determining a position of a vehicle, comprising a vehicle which comprises at least one outdoor camera, a control unit and a mobile radio module and configured to carry out the first method described above, a measuring vehicle which has at least one external camera, a controller and a cellular module and configured to execute the second method described above, a backend server configured to perform the third method described above, and more particularly to a convolutional auto-encoder for optical data encoding / decoding forming a first software module a deep neural network for encoding / decoding optical data constituting a second software module. Thus, the method according to the invention can be carried out with a relatively simple system consisting of a vehicle, a measuring vehicle and a back-end server and does not require an ability to use satellite-supported position data.

• 1 in einem schematischen Ablaufdiagramm eine Ausführungsform des erfindungsgemäßen Verfahrens;
• 2 in einer schematischen Darstellung ein von einer in Vorwärtsrichtung weisenden Außenkamera erfasstes Bild einer Umgebung des Fahrzeugs;
• 3 in einem Funktionsgraph einen in einer weiteren Ausführungsform des erfindungsgemäßen Verfahrens gebildeten Positionsstrom;
• 4 den in 3 dargestellten Funktionsgraph mit darin eingetragenen Positionen zweier Fahrzeuge.

The invention is illustrated schematically with reference to several embodiments in the drawings and will be further described with reference to the drawings. It shows:

• 1 in a schematic flow diagram, an embodiment of the method according to the invention;
• 2 in a schematic representation, an image of an environment of the vehicle detected by a forward-facing outdoor camera;
• 3 in a functional graph, a position current formed in a further embodiment of the method according to the invention;
• 4 the in 3 illustrated function graph with registered therein positions of two vehicles.

1 shows in a schematic flow diagram an embodiment of the method according to the invention. In the method for optically determining a position of a vehicle, a mobile radio connection is first established between the vehicle and a back-end server by means of a mobile radio module provided in the vehicle. The vehicle receives a software module from the backend server 11 for encoding / decoding optical data in the form of a convolutional autoencoder and installs it in a control unit.

After determining a route 13 for a ride of the vehicle, the vehicle receives over the established mobile connection from the back-end server to the installed software module 11 matching localization current 14 which of the particular route 13 is assigned and stores the received localization stream 14 in the control unit. To form one to the installed software module 11 corresponding position current 12 be while driving the vehicle on the specific route 13 Periodically, the following steps are performed: An external camera provided on the vehicle becomes an image 10 . 20 an environment of the vehicle detected. By means of the installed software module 11 a coded position value corresponding to the detected image is calculated and the calculated coded position value is stored by the control device. This creates the position current 12 during the drive of the vehicle incrementally from the stored coded position values. The position current 12 is periodic with a portion 15 of the received localization stream 14 by means of a comparator 16 abstract compared to determine the position of the vehicle. The comparator 16 calculated for abstract comparing the position current 12 with a portion 15 of the received localization stream 14 a correlation function. At this time, the localization current section 15 used for the comparison becomes 14 determined by satellite-based GPS data received by means of a satellite receiver provided in the vehicle.

The location stream received and used by the vehicle 14 is during a journey of a measuring vehicle on the same route 13 generated. In this case, essentially the method already described above is carried out. In contrast to this but no position current 12 but a localization stream 14 educated. The localization current 14 includes besides those with the installed software module 11 as described above, additionally calculated speed values and times. The speed values and times are respectively detected by a speed sensor or timer provided in the measuring vehicle and linked with the position values to localization values and then stored by the control unit. The localization current formed in this way 14 is sent to the backend server via the established mobile connection.

Two localization streams can also be synchronized in the measuring vehicle 14 to the installed software module 11 be formed. This will be two pictures 10 . 20 the surroundings of the vehicle by two external cameras, which may be formed as respective stereoscopic cameras detected simultaneously.

In addition, the localization current 14 DGPS data include. Such a localization stream 14 is formed by additionally combining calculated coded position values with DGPS data received by means of a satellite receiver arranged in the measuring vehicle.

The backend server provides the software module 11 and the localization current formed by the measuring vehicle 14 ready for use by the vehicle. To do this, the backend server receives the localization stream 14 from the measuring vehicle and stores it to provide for receiving by the vehicle. Furthermore, the backend server uses the received localization stream 14 to the software module 11 to train. The trained software module 14 is also provided for receiving by the vehicle.

2 shows a schematic representation of a captured from a forward facing outdoor camera two-dimensional image 20 an environment of the vehicle. In the picture 20 are another vehicle 21 , Building 22 , A street 23 with road markings 24 , the sky 25 and parts of a landscape 26 recognizable.

3 shows a position current in a function graph 30 which was formed in a further embodiment of the method according to the invention. To form the position current 30 becomes every captured image 20 a corresponding coded position value calculated as follows. First, in the captured image 20 through the controller by means of a deep neural network objects, like a sky 25 , Building 22 or parts of a landscape 26 , which are assigned to predetermined object classes identified. Other object classes may include signs or posts that are permanently located at fixed positions along the roadway. Then every recognized object 22 . 25 . 26 enclosed by a minimal rectangular contour, ie framed and the number of pixels 36 . 37 . 38 determined within the contour. The determined number of pixels 36 . 37 . 38 finally becomes a coded position value for forming the position current 30 used. The position current 30 contains pixel numbers 36 . 37 . 38 of objects 22 . 25 . 26 taking pictures captured during a ride 20 were identified. In the function graph are the pixel counts 36 Of the sky 25 , the pixel counts 37 of buildings 22 , the pixel counts 38 of parts of a landscape 26 and the pixel counts of other detected objects corresponding to over a route 13 applied. It can be seen from the function graph that the pixel numbers calculated in each case 36 . 37 . 38 independent of each other and depending on the location on the route 13 vary. Thus, the position current includes 30 highly individual sections that are readily accessible to an abstract comparison.

4 also shows the in 3 shown function graph, in addition to the position 33 of the vehicle and the position 34 registered in another vehicle. The on the from a starting point 31 to a destination 32 leading route 13 related longitudinal distance 35 between the vehicle and the other vehicle can be calculated by a simple difference. In addition, for each of the vehicle and the other vehicle from the starting point 31 the distance covered and as far as the destination 32 calculate remaining distance.

A significant advantage of the method according to the invention is that compared to the precision of GPS significantly better precision in determining a position of a vehicle is achieved. This applies in particular to the longitudinal position of the vehicle relative to the route. Another advantage is that the method according to the invention is based solely on detecting an environment of the vehicle, i. H. that only local information is used to determine the position of the vehicle. A reception of satellite-based position data can even be dispensed with, as a result of which the position of the vehicle can be reliably and precisely determined even in deep urban canyons or tunnels. In addition, no conditions are attached to the environment of the vehicle, d. H. For example, neither markings of lanes or landmarks are required. A third advantage arises from the fact that the use of a convolutional autencoder or a deep neural network causes a very low computation effort, whereby the resources of a corresponding control unit provided in the vehicle are spared.

LIST OF REFERENCE NUMBERS

10

captured image

11

Convolutional Autoencoder

12

position power

13

route

14

location current

15

section

16

comparator

20

captured image

21

another vehicle

22

building

23

Street

24

road marking

25

sky

26

landscape

30

diagram

31

starting point

32

destination

33

Position of a first vehicle

34

Position of a second vehicle

35

distance

36

Pixel number of sky

37

Pixel number of buildings

38

Pixel number landscape

Claims (10)

A method for optically determining a position of a vehicle, comprising the following steps: - establishing a mobile radio connection between the vehicle and a back-end server by means of a mobile radio module provided in the vehicle; Receiving at least one software module (11) for encoding / decoding optical data from the backend server via the established mobile radio connection and installing the at least one received software module (11) in a control unit of the vehicle; - determining a route (13) for driving the vehicle; Receiving, from the backend server via the established cellular connection, at least one localization stream (14) for each installed software module (11) associated with the particular route (13) and storing each localization stream (14) in the controller; - Forming a position current (12) to each installed software module (11) by multiple, in particular periodically performing the following steps while driving the vehicle on the particular route (13): - Capturing an image (10, 20) of an environment of the vehicle an outdoor camera provided on the vehicle; - Calculating each one of the captured image (10, 20) corresponding coded position value by means of each installed software module (11); and - storing the calculated encoded position value by the controller; and - determining the position of the vehicle by continuous, in particular periodic abstract Comparing the formed position current (12) with a portion (15) of the received at least one localization current (14). Method according to Claim 1 in which a correlation function for comparing the generated position current (12) with the section (15) of the at least one localization current (14) is calculated. Method according to Claim 2 in which satellite-based position data, in particular GPS data, are received by means of a satellite receiver provided in the vehicle, and the section (15) of the at least one localization current (14) is determined by means of the received position data. Method according to one of Claims 1 to 3 in which from the determined position of the vehicle and at least one localization current (14) a length of a section of road traveled by the vehicle on the route (13) and / or a distance of the vehicle to another vehicle (21) on the route (13) is calculated. A method of generating at least one localization stream (14) for use in a method according to any one of Claims 1 to 4 during a journey of a measuring vehicle, which comprises the following steps: - establishing a mobile radio connection between the measuring vehicle and a back-end server by means of a mobile radio module provided in the measuring vehicle; Receiving at least one software module (11) for encoding / decoding optical data from the backend server via the established mobile radio connection and installing the at least one received software module (11) in a control unit of the measuring vehicle; - Determining a route (13) for a journey of the measuring vehicle; - Forming at least one localization stream (14) to each installed software module (11) by continuously, in particular periodically performing the following steps while driving the measuring vehicle on the particular route (13): - Capturing an image (10, 20) of an environment of the measuring vehicle by an outdoor camera provided on the measuring vehicle; - Calculating each one of the captured image (10, 20) corresponding coded position value by means of each installed software module (11); Detecting a speed value of the measuring vehicle by means of a speed sensor provided in the measuring vehicle; Detecting a time by means of a timer provided in the measuring vehicle; Forming a localization value by associating the calculated position value with the detected velocity value and the detected time; and - storing the formed localization value by the controller; and - sending each formed localization stream (14) to the backend server via the established cellular connection. Method according to Claim 5 in which at least two localization streams to an installed software module (11) are formed by simultaneously capturing at least two images (10, 20) of the surroundings of the vehicle by at least two external cameras, in particular exactly two images (10, 20) by a stereoscopic camera , Method according to one of Claims 5 or 6 in which satellite-based position data, in particular GPS data and preferably DGPS data, are received by means of a satellite receiver arranged in the measuring vehicle, and the localization value is formed by additionally linking the calculated position value with the received satellite-based position data. A method for providing at least one software module (11) for encoding / decoding optical data on a backend server and at least one localization stream (14) for use in a method according to any one of Claims 1 to 7 comprising the steps of: receiving the at least one localization stream (14) from a measurement vehicle by the backend server; - storing the received localization stream (14); - providing the stored localization current (14) for receiving by a vehicle; - Training the at least one software module (11) by means of the stored localization stream (14); and - providing the at least one trained software module (11) for receiving by a vehicle. Method according to one of Claims 1 to 8th in which a convolutional auto-encoder is used as a first software module for encoding / decoding optical data and preferably a deep neural network is used as a second software module for encoding / decoding optical data, with which to compute one to the acquired image (10, 20) corresponding to the encoded position value, the following steps are carried out: identifying an object which is associated with at least one predetermined object class, which in particular includes signs, posts or buildings, in the captured image (10, 20) by means of the deep neural network; - Determining, by the controller, a number of pixels (36, 37, 38) within a minimum, in particular rectangular, contour enclosing the identified object; and - using the determined number of pixels (36, 37, 38) as coded position value. A system for optically determining a position of a vehicle, comprising a vehicle, which comprises at least one outdoor camera, a control device and a mobile radio module and for carrying out a method according to one of the Claims 1 to 4 is configured, a measuring vehicle, which comprises at least one outdoor camera, a control unit and a mobile radio module and for carrying out a method according to one of Claims 5 to 7 is configured to a backend server, which is used to perform a method Claim 8 and, in particular, a convolutional auto-encoder forming a first software module (11) for encoding / decoding optical data, and preferably a deep neural network for encoding / decoding optical data forming a second software module (11).

Images