Camera-based vehicle docking with the help of artificial intelligence

Abstract

The present invention relates to docking a vehicle based on a camera by means of artificial intelligence. An evaluation device for a docking station, the evaluation device comprising: an entry port for obtaining at least one image of the docking station captured by an imaging sensor that can be arranged on a vehicle, wherein the evaluation device is implemented by implementing an artificial neural network that is trained to obtain image coordinates of a support point of the docking station based on the image, to obtain a position and / or orientation of the imaging sensor relative to the support point based on a known geometric shape of the support point, and to obtain a position and / or orientation of the docking station relative to the vehicle based on the known position and / or orientation of the imaging sensor and the known arrangement of the imaging sensor on the vehicle; an exit port for providing a signal for a vehicle control device based on the known position of the docking station relative to the vehicle, so as to control the docking of the vehicle to the docking station.

Description

Technical Field

[0001] The invention relates to an evaluation device for locating a support point of a docking station. The invention also relates to a method for locating a support point of a docking station. The invention also relates to an evaluation device for automatically docking a vehicle to a docking station. The invention also relates to a vehicle for automatically docking to a docking station. The invention also relates to a method for automatically docking a vehicle to a docking station. Finally, the invention relates to a computer program product for docking a vehicle to a docking station. Background Art

[0002] On the one hand, the challenge of automated driving is flexible driving in road traffic. On the other hand, especially in the field of commercial vehicles, the task of automated driving is docking processes, in which, for example, goods are loaded and unloaded and / or tools are changed.

[0003] The publication GB 2 513 393 describes a device with a camera and a target. The camera is fixed to the vehicle. The target, for example a patterned panel, is fixed to the trailer. By identifying and locating the target in the image captured by the camera, a trajectory can be calculated. The trajectory describes the path in the direction of the trailer that the vehicle must pass through in order to dock with the trailer.

[0004] The publication DE 10 2006 035 929 B4 discloses a method for driving under or into an object, in particular into a changeover bridge, with the aid of a commercial vehicle, wherein surrounding information is detected by at least one sensor arranged at the rear of the commercial vehicle, and wherein the relative position between the object and the commercial vehicle is determined based on the detected surrounding information, wherein object features of a hierarchical model of the object are selected in a sensor-controlled manner in at least two stages in dependence on the distance, wherein during the approach of the commercial vehicle to the object, individual modeling of the object is achieved by model adaptation based on the individual object features. In this case, the hierarchical model changes with the distance from the changeover bridge. For example, "coarse" features are detected in the long-distance range, while the model is refined in the short-distance range for more precise positioning. Summary of the invention

[0005] The present invention is proposed here. The object of the invention is to improve the automated docking of vehicles.

[0006] The object is achieved by an evaluation device according to the invention for locating a support point of a docking station. The object is also achieved by a method according to the invention for locating a support point of a docking station. The object is also achieved by an evaluation device according to the invention for automatically docking a vehicle to a docking station. The object is also achieved by a vehicle according to the invention for automatically docking to a docking station. The object is also achieved by a method according to the invention for automatically docking a vehicle to a docking station. Finally, the object is achieved by a computer program product according to the invention for docking a vehicle to a docking station.

[0007] The evaluation device for locating the support points of the docking station in the image of the docking station according to the present invention comprises a first entry port for obtaining actual training data. The actual training data comprises an image of the docking station. For training, there are position data of the support points as separate information. The evaluation device also comprises a second entry port for obtaining target training data. The target training data comprises target position data of the corresponding support points in the image. The evaluation device is implemented as follows: the actual training data is fed forward to the artificial neural network, and the actual position data of the corresponding support points known by the artificial neural network are obtained in the forward feed. The evaluation device is also implemented as follows: by backfeeding the deviation between the actual position data and the target position data, the weight factor for the connection of the neurons of the artificial neural network is adjusted to minimize the deviation so as to learn the target position data of the support points. The evaluation device also comprises an exit port for providing the actual position data.

[0008] The following definitions apply to the entire subject matter of the present invention.

[0009] An evaluation device is a device which processes input information and outputs a result obtained from the processing. The evaluation device is in particular an electronic circuit, such as a central processing unit or a graphics processor.

[0010] Support points, also called key points in English, are, for example, corner points of the trailer and / or other distinct points on the trailer or the docking station. According to the invention, the support points are therefore detected directly, which are characteristic of the docking station. This means that, in particular, the free space between the supports of the transfer bridge is not taken into account for the classification, so that it is not necessary to use complex object-free space models, which vary with the distance from the transfer bridge.

[0011] A docking station is an object with which a vehicle can dock. In the docked state, the vehicle is connected to the docking station. Examples of docking stations are trailers, containers, transfer bridges or docks, such as loading and unloading boards. The vehicle is a land vehicle, such as a passenger car, a commercial vehicle, such as a truck or a tractor, such as a tractor or a rail vehicle. The vehicle can also be a watercraft, such as a ship.

[0012] An image is a picture taken by an imaging sensor. A digital camera includes an imaging sensor. An image is especially a color image.

[0013] Artificial intelligence is a general concept for the automation of intelligent behavior. For example, learning intelligent algorithms in order to react to new information in a targeted manner. Artificial neural networks, also known as artificial neural networks in English, are intelligent algorithms. Intelligent algorithms are implemented to learn in order to react to new information in a targeted manner.

[0014] In order to be able to react to new information in a targeted manner, artificial intelligence needs to first learn the meaning of predetermined information. To this end, artificial intelligence is trained with valid data. Valid data is generally a general concept for training data or test data. In addition to real data, training data also contains information about the meaning of the corresponding data. In other words, the training data learned by artificial intelligence and called actual training data is labeled. Target training data is real, given information. Target position data is especially the two-dimensional image coordinates of support points. This training phase is inspired by the learning process of the brain.

[0015] Specifically, useful data is a data set with which the algorithm was tested during the development period. Since the developers also make decisions that influence the algorithm based on the tests, a further data set, the test data set, is taken into account at the end of the development for the final evaluation. For example, images of the docking station against different backgrounds constitute a further data set.

[0016] Training with available data is called machine learning. A subset of machine learning is deep learning, in which a series of hierarchical layers of neurons, so-called hidden layers, are used to perform the machine learning process.

[0017] A neuron is a functional unit of an artificial neural network. The output of a neuron is usually obtained as the value of an activation function in an estimated manner via the sum of the inputs weighted by weight factors plus a systematic error, the so-called bias. Artificial neural networks with multiple hidden layers are deep neural networks.

[0018] Artificial neural network is, for example, a fully connected network, which is called Fully Connected Network in English. In a fully connected network, each neuron of a layer is connected to all neurons of the previous layer. Each connection has its own weight factor. Preferably, the artificial neural network is a fully convolutional network (Fully Convolutional Network). In a convolutional neural network, filters are applied to the layers of neurons using the same weight factor regardless of position. The convolutional neural network includes a plurality of pooling layers between convolutional layers. The pooling layer changes the size of the two-dimensional layer in terms of width and height. The pooling layer is also used for higher dimensional layers. Preferably, the artificial neural network is a convolutional neural network with an encoder-decoder architecture known to those skilled in the art.

[0019] Thereby, the evaluation device learns to recognize the support points in the image. Preferably, the output of the artificial neural network is a pixel-by-pixel probability for the support points, that is to say, a so-called predicted heat map is obtained for each support point, which heat map illustrates the pixel-by-pixel probability of the support points. The target position data, also called ground truth heat map, then preferably include a two-dimensional Gaussian clock with a standardized height, the maximum value of which is at the support point position. The deviation of the actual position data from the target position data is then minimized by means of the cross entropy between the ground truth heat map and the predicted heat map.

[0020] The method according to the invention for locating a support point of a docking station in an image of the docking station comprises the following method steps:

[0021] Obtain actual training data and position data of support points;

[0022] Obtaining target training data, wherein the target training data includes target position data of corresponding support points in the image;

[0023] Feeding the actual training data forward to the artificial neural network and using the artificial neural network to obtain the actual position data of the corresponding support points;

[0024] The deviation between the actual position data and the target position data is fed back in order to adjust the weight factors for the connections of the neurons of the artificial neuron network in such a way that the deviation is minimized in order to learn the target position data of the support point.

[0025] This method is a training method for artificial neural networks. In the so-called training phase, the connection between neurons is evaluated using weight factors. Forward feed, called forward propagation in English, refers to: feeding information into the input layer of the artificial neural network, passing through the next layer and outputting it in the output layer. Backward feed, called backward propagation in English, refers to: information passes through these layers in reverse, that is, from the output layer to the direction of the input layer. The deviation of the corresponding layer is obtained by gradually feeding the obtained deviation of the target data and the actual data from the output layer back to the corresponding previous layer until it reaches the input layer. These deviations are functions of weight factors. The deviation between the actual output and the target output is evaluated by the cost function. In the backfeed, the gradient of the error is fed back according to each weight. In this way, it is known whether the deviation between the actual output and the target output is minimized and how strongly it is minimized when the corresponding weight is increased or decreased. By minimizing the deviation in the training phase, for example, by means of the least squares method, the cross entropy or the gradient descent method known from information theory, the weight factor is changed. Thus, when the input is fed back, an output close to the desired output is achieved. Backfeeding is described in detail, for example, in Michael A. Nielsen, Neural Networks and Deep Learning, Determination Press, 2015.

[0026] Advantageously, an evaluation device according to the invention for locating a support point of a docking station is used to carry out the method.

[0027] Preferably, the training method is executed on a graphics processor while fully utilizing parallel computing.

[0028] The evaluation device for automatically docking a vehicle to a docking station according to the present invention comprises an entry port for obtaining at least one image of the docking station captured by an imaging sensor that can be arranged on the vehicle. The evaluation device is implemented by implementing an artificial neural network. The artificial neural network is trained to learn the image coordinates of the support points of the docking station based on the image. The evaluation device is also implemented by learning the position and / or orientation of the imaging sensor relative to the support points based on the known geometry of the support points. The evaluation device is also implemented by learning the position and / or orientation of the docking station relative to the vehicle based on the known position and / or orientation of the imaging sensor and the known arrangement of the imaging sensor on the vehicle. The evaluation device also comprises an exit port for providing a signal for a vehicle control device based on the known position of the docking station relative to the vehicle so that the vehicle automatically drives to the docking station for docking.

[0029] Imaging sensors provide images at every timestamp and not just point clouds, such as radar, lidar or laser.

[0030] The image coordinates are two-dimensional coordinates of an object from a three-dimensional object space in a reference space of a two-dimensional image of the object.

[0031] The vehicle control device comprises control circuits and / or actuators, with which the longitudinal and / or transverse guidance of the vehicle can be regulated and / or controlled.

[0032] The vehicle can thereby advantageously be driven automatically to the calculated position of the docking station and docked thereto. The signal is, for example, the steering angle. Thus, an end-to-end process can also be achieved. Advantageously, the pose estimation based on the support point is very precise and allows more control and thus greater algorithmic safety compared to end-to-end learning.

[0033] The vehicle control system preferably includes a trajectory control.

[0034] From the three-dimensional model of the docking station, for example, the geometry of the support points is known, for example, the relative position of the support points to one another. If no model exists in advance, the support points are previously measured according to the invention. The position and / or orientation of the imaging sensor relative to the support points is then derived from the knowledge of the geometry of the support points, preferably from intrinsic parameters of the imaging sensor. The intrinsic parameters of the imaging sensor define how optical measurements of the imaging sensor relate to the image points of the imaging sensor, in particular the pixel values. For example, the focal length of the optical system of the imaging sensor or the pixel density of the imaging sensor are intrinsic parameters of the imaging sensor.

[0035] Preferably, the artificial neuron network is trained according to the method according to the invention for locating support points of a docking station.

[0036] The vehicle according to the invention for automatic docking to a docking station comprises a camera with an imaging sensor. The camera is arranged on the vehicle in order to obtain an image of the docking station. The vehicle also comprises an evaluation device according to the invention for automatic docking of the vehicle to the docking station in order to provide a signal for a vehicle control device as a function of the known position and / or orientation of the docking station relative to the vehicle. The vehicle also comprises a vehicle control device in order to automatically drive the vehicle to the docking station for docking as a function of the signal.

[0037] By this, the vehicle can advantageously be driven automatically to the calculated position of the docking station and docked on the docking station. Therefore, the vehicle is preferably a vehicle that can be operated automatically, preferably semi-automatically. A vehicle that can be operated automatically is a vehicle that has the following technical equipment, which can control the corresponding vehicle with a vehicle control device after activating the corresponding automatic driving function, especially the highly automated or fully automated driving function according to the SAEJ3016 standard, in order to complete the driver's tasks, including longitudinal and lateral guidance. A vehicle that can be operated semi-automatically can undertake specific driving tasks. A fully automated vehicle replaces the driver. The difference between the SAEJ3016 standard SAE Level 4 and SAE Level 5 is that Level 4 is defined in the following way: the driving mode-specific implementation scheme of all aspects of the dynamic driving task is implemented by the automated driving system, even if the human driver does not respond appropriately to the system's inquiry. Level 5 is defined in the following way: all aspects of the dynamic driving task are implemented consistently by the automated driving system under all driving and surrounding conditions that the human driver can handle. Pure assistance systems, which are also the subject of the present invention, assist the driver in carrying out driving tasks. This corresponds to SAE Level 1. The assistance system assists the driver in steering maneuvers, for example, by means of a visual output on a human machine interface, HMI for short. The human machine interface is, for example, a screen, in particular a touchscreen screen.

[0038] The method according to the invention for automatically docking a vehicle to a docking station comprises the following method steps:

[0039] Obtaining at least one image of the docking station captured using an imaging sensor that can be disposed on the vehicle;

[0040] Implementing an artificial neural network that is trained to learn the image coordinates of the support points of the docking station from the images;

[0041] The position and / or orientation of the imaging sensor relative to the support points is known based on the known geometric shapes of the support points;

[0042] knowing the position and / or orientation of the docking station relative to the vehicle based on the known position of the imaging sensor and the known placement of the imaging sensor on the vehicle;

[0043] as well as

[0044] • Providing signals for vehicle controls based on the learned position and / or orientation of the docking station relative to the vehicle.

[0045] Thereby, a prior manipulation of the docking station, for example by installing sensors, markings or pattern panels, is no longer necessary. The position and / or orientation of the docking station is identified with the aid of the recorded support points.

[0046] Preferably, the vehicle control device automatically drives the vehicle to the docking station for docking according to the signal. By means of this method, the vehicle can be advantageously automatically docked on the docking station.

[0047] Advantageously, a known model of the docking station is used, with the position and / or orientation of the imaging sensor relative to the support points being known from the known geometry of the support points, wherein the model describes the relative position of the support points with respect to one another.

[0048] Preferably, using the known model, intrinsic parameters of the imaging sensor are used.

[0049] Particularly preferably, a coordinate transformation from the system of the imaging sensors to the system of the vehicle is performed when the position and / or orientation of the docking station relative to the vehicle is known based on the known position of the imaging sensors and the known arrangement of the imaging sensors on the vehicle. Trajectory planning towards the docking station can be achieved based on the coordinate system of the vehicle.

[0050] Thus, by means of the method according to the invention, based on the knowledge of the relative three-dimensional positions of the support points, a two-dimensional projection of the support points is obtained by the artificial neural network and the trajectory to the docking station is known.

[0051] Advantageously, an evaluation device according to the invention for automated docking of a vehicle to a docking station or a vehicle according to the invention for automated docking of a vehicle to a docking station is used to carry out the method.

[0052] The computer program product according to the invention for docking a vehicle to a docking station is designed to be loaded into a memory of a computer and comprises software code portions with which the method steps of the method according to the invention for automatically docking a vehicle to a docking station are implemented when the computer program product is run on the computer.

[0053] Programs belong to the software of a data processing system, such as an evaluation device or a computer. Software is a general term for programs and the associated data. Hardware supplements software. Hardware represents the mechanical and electronic refinement of a data processing system. A computer is an evaluation device.

[0054] A computer program product generally includes a sequence of instructions which, when loaded, causes the hardware to execute a certain method which leads to a certain result. When the program in question is used on a computer, the computer program product produces the technical effect of obtaining a trajectory plan for automated docking to a docking station.

[0055] The computer program product according to the invention is platform-independent. That is, it can be implemented on any arbitrary computing platform. Preferably, it is implemented on an evaluation device according to the invention for automatically docking a vehicle to a docking station.

[0056] The software code part is written in any programming language, for example, in Python. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The present invention is illustrated by way of example in the accompanying drawings, wherein:

[0058] Figure 1 An embodiment of a vehicle and an embodiment of a docking station according to the present invention are shown;

[0059] Figure 2 An embodiment of a docking station is shown;

[0060] Figure 3 An embodiment of an evaluation device for locating support points of a docking station according to the invention is shown;

[0061] Figure 4 A schematic diagram showing a method for locating support points of a docking station according to the present invention;

[0062] Figure 5 An embodiment of an evaluation device according to the invention for automatically docking a vehicle to a docking station is shown; and

[0063] Figure 6 An exemplary embodiment of the method according to the invention for the automated docking of a vehicle to a docking station is shown.

[0064] In the figures, the same reference numerals refer to identical or functionally similar reference parts. The respective associated reference parts are represented in the individual figures. DETAILED DESCRIPTION

[0065] Figure 1 A tractor is shown as a vehicle 30. The tractor transports a trailer, which is a docking station 10 for the tractor. When the vehicle reaches the docking station 10, it is coupled to the docking station. Approaching the docking station 10 and docking are performed automatically. For this purpose, the vehicle comprises a camera 33.

[0066] The camera 33 records images 34, which cover the rearward field of view of the vehicle 30. The docking station 10 is detected by means of these images 34. Figure 2 The support point 11 is shown in FIG. Figure 2 A graphic panel 9 is also shown, by means of which the position and / or orientation of the docking station 10 can also be detected. However, the graphic panel 9 is not absolutely necessary according to the invention. The camera 33 comprises an imaging sensor 31. The imaging sensor 31 sends an image 34 to the evaluation device 20 for the automated docking of the vehicle 30 to the docking station 10.

[0067] exist Figure 5 2 shows an evaluation device 20. The evaluation device 20 acquires images 34 of an imaging sensor 31 via an inlet port 21. These images 34 are provided to an artificial neuron network 4.

[0068] The artificial neural network 4 is a fully convolutional network. The artificial neural network 4 includes an input layer 4a, two hierarchical layers 4b and an output layer 4c. The artificial neural network may include a plurality of, for example, more than 1000 hierarchical layers 4b. The artificial neural network 4 is based on Figure 4 34 for locating the support point 11 of the docking station 10 in the image 34. That is, the artificial neural network 4 calculates the image coordinates of the support point from the image 34 and, based on this, calculates the position and orientation of the docking station 10 relative to the vehicle 30, knowing the geometry of the support point and the arrangement of the imaging sensor 31 on the vehicle 30. Based on the position and orientation of the docking station 10 relative to the vehicle 30, the evaluation device 20 calculates the trajectory of the vehicle 30 towards the docking station 10 and outputs a corresponding control signal to the vehicle control device 32. The evaluation device 20 provides the control signal to the vehicle control device 32 via the output port 22.

[0069] Use in Figure 3 The evaluation device 1 for locating the support point 11 of the docking station 10 in the image 34 of the docking station 10 is executed in Figure 4 2 shows a training method for locating support points 11 of a docking station 10 in images 34 of the docking station 10. In the training method, the images 34 are labeled, that is, the support points 11 are marked in these images.

[0070] The evaluation device 1 comprises a first inlet port 2. The evaluation device 1 obtains actual training data through the first inlet port 2. The actual training data is the image 34. The actual training data is obtained at Figure 4 The first method step V1 is shown in FIG.

[0071] The evaluation device 1 further comprises a second inlet port 3. The evaluation device 1 obtains target training data through the second inlet port 3. The target training data comprises target position data of the corresponding support point 11 in the marked image 34. The target training data is obtained at Figure 4 The second method step V2 is shown in FIG.

[0072] The evaluation device 1 also includes an artificial neural network 4. The artificial neural network 4 has, for example, a structure similar to that in Figure 5 The architecture of the artificial neural network 4 of the evaluation device 20 is shown in FIG.

[0073] The actual training data is fed forward to the artificial neural network 4. In the forward feeding, the artificial neural network 4 is used to obtain the actual position data of the corresponding support point 11. The forward feeding together with the obtaining of the actual position data is Figure 4 Method step V3 shown in .

[0074] The deviation between the actual position data and the target position data is fed back through the artificial neural network 4. In the backfeed, the weight factors 5 for the connections 6 of the neurons 7 of the artificial neural network 4 are adjusted so as to minimize the deviation. Here, the target position data of the support point 11 is learned. The target position data is learned in Figure 4 Method step V4 is shown in .

[0075] The evaluation device 1 further comprises an outlet port 8. Actual position data obtained with the artificial neuronal network 4, which actual position data approached the target position data during the training method, are provided via the outlet port.

[0076] Use in Figure 5 The trained evaluation device 20 shown in FIG. Figure 6 , a method for automatically docking a vehicle 30 to a docking station 10 is shown in FIG. In a first method step S1, at least one image 34 of the docking station 10 captured by an imaging sensor 31 arranged on the vehicle 30 is obtained. Here, the image 34 is a normal image without marking the support point 11.

[0077] In a second method step S2 , an artificial neural network 4 is implemented. The artificial neural network 4 is trained to determine the image coordinates of the support points 11 of the docking station 10 based on the images 34 .

[0078] In a third method step S3, the position and / or orientation of the imaging sensor 31 relative to the support points 11 is determined based on the known geometry of the support points 11. In method step S3a, the geometry of the support points 11 is determined with the aid of a known three-dimensional model of the docking station 10, wherein the model describes the relative positions of the support points 11 with respect to one another. In method step S3b, intrinsic parameters of the imaging sensor 31 are used using the known model.

[0079] In method step S4, the position and / or orientation of the docking station 10 relative to the vehicle 30 is determined based on the determined position of the imaging sensor 31 and the known arrangement of the imaging sensor 31 on the vehicle 30. In this case, in method step S4a, a coordinate transformation is performed from the system of the imaging sensor 31 to the system of the vehicle 30. The position and / or orientation of the docking station 10 in the system of the vehicle is known by means of this coordinate transformation, so that the vehicle can be automatically docked to the calculated position of the docking station 10 by means of trajectory adjustment.

[0080] In method step S5 , a signal is provided for the vehicle control device 32 as a function of the determined position and / or orientation of the docking station 10 relative to the vehicle 30 .

[0081] In method step S6 , vehicle control device 32 drives vehicle 30 automatically to docking station 10 for docking based on the signal.

[0082] Reference numerals list

[0083] 1 Evaluation device

[0084] 2First inlet port

[0085] 3 Second entry port

[0086] 4 Artificial Neural Networks

[0087] 4a Input layer

[0088] 4b Grading layer

[0089] 4c Output layer

[0090] 5Weight Factor

[0091] 6Connect

[0092] 7 Neurons

[0093] 8 outlet ports

[0094] 9 pattern panels

[0095] 10 Docking Station

[0096] 11 Support Points

[0097] 20 Evaluation Devices

[0098] 21 Entry Ports

[0099] 22 Exit Ports

[0100] 30 vehicles

[0101] 31 Imaging Sensor

[0102] 32 Vehicle Control Devices

[0103] 33 cameras

[0104] 34 images

[0105] V1-V4 Method Steps

[0106] S1-S6 Method Steps

Claims

1. An evaluation device (1) for locating a support point (11) of a docking station (10) in an image (34) of the docking station (10), the evaluation device comprising: a first entry port (2) for obtaining actual training data, wherein the actual training data comprises an image of the docking station (10), wherein the support point (11) is marked in the image (34); A second entry port (3) is used to obtain target training data, wherein the target training data includes target position data of the corresponding support point (11) in the image (34), and The output port (8) is used to provide the actual position data, Wherein, the evaluation device (1) is implemented as follows: The actual training data is fed forward to the artificial neural network (4), and the actual position data of the corresponding support point (11) obtained by using the artificial neural network (4) is obtained in the forward feeding, and The weight factor (5) of the connection (6) of the neuron (7) of the artificial neural network (4) is adjusted by back-feeding the deviation between the actual position data and the target position data to minimize the deviation, so as to learn the target position data of the support point (11).

2. A method for locating a support point (11) of a docking station (10) in an image (34) of the docking station (10), wherein: To carry out the method, an evaluation device (1) according to claim 1 is used, the method comprising the following method steps: Obtaining actual training data and position data (V1) of the support point (11); Obtaining target training data, wherein the target training data includes target position data (V2) of the corresponding support point (11) in the image (34); Feeding the actual training data forward to the artificial neural network (4) and using the artificial neural network (4) to obtain the actual position data (V3) of the corresponding support point (11); The deviation between the actual position data and the target position data is fed back so as to adjust the weight factor (5) of the connection (6) of the neurons (7) of the artificial neural network (4) as follows, i.e., to minimize the deviation so as to learn the target position data (V4) of the support point (11).

3. An evaluation device (20) for automatically docking a vehicle (30) to a docking station (10), the evaluation device comprising: an entry port (21) for obtaining at least one image (34) of the docking station (10) captured by an imaging sensor (31) that can be arranged on the vehicle (30), and an exit port (22) for providing a signal for a vehicle control device (32) based on the known position of the docking station (10) relative to the vehicle (30) so that the vehicle (30) automatically drives to the docking station (10) for docking, Wherein, the evaluation device is implemented as: Implementing an artificial neural network (4) which is trained to determine the image coordinates of the support points (11) of the docking station (10) based on the image (34), wherein the artificial neural network (4) is trained according to the method of claim 2, The position and / or orientation of the imaging sensor (31) relative to the support point (11) is determined based on the known geometry of the support point (11), and The position of the docking station (10) relative to the vehicle (30) is known based on the known position and / or orientation of the imaging sensor (31) and the known arrangement of the imaging sensor (31) on the vehicle (30).

4. A vehicle (30) for automatically docking to a docking station (10), the vehicle comprising: a camera (33) having an imaging sensor (31), the camera being arranged on the vehicle (30) so as to obtain an image (34) of the docking station (10); Evaluation device (20) according to claim 3, for providing a signal for a vehicle control device as a function of the known position and / or orientation of the docking station (10) relative to the vehicle (30); and A vehicle control device (32) is used to automatically drive the vehicle (30) to the docking station (10) for docking according to the signal.

5. A method for automatically docking a vehicle (30) to a docking station (10), wherein: To carry out the method, an evaluation device (20) according to claim 3 or a vehicle (30) according to claim 4 is used, the method comprising the following method steps: Obtaining at least one image (34) of the docking station (10) captured by an imaging sensor (31) that can be arranged on the vehicle (30) (S1); Implementing an artificial neural network (4) which is trained to obtain the image coordinates (S2) of the support points (11) of the docking station (10) based on the image (34); Obtaining the position (S3) of the imaging sensor (31) relative to the supporting point (11) based on the known geometric shape of the supporting point (11); Determining the position and / or orientation of the docking station (10) relative to the vehicle (30) based on the known position of the imaging sensor (31) and the known arrangement of the imaging sensor (31) on the vehicle (30) (S4); as well as A signal (S5) is provided for a vehicle control device (32) based on the determined position and / or orientation of the docking station (10) relative to the vehicle (30).

6. The method according to claim 5, wherein: The vehicle control device (32) drives the vehicle (30) automatically to the docking station (10) for docking according to the signal (S6).

7. The method according to claim 5 or 6, wherein: A known model of the docking station (10) is used when the position (S3) of the imaging sensor (31) relative to the support points (11) is known based on the known geometry of the support points (11), wherein the model describes the relative positions (S3a) of the support points (11) with respect to one another.

8. The method according to claim 7, wherein: When using the known model, intrinsic parameters (S3b) of the imaging sensor (31) are used.

9. The method according to claim 5 or 6, wherein: When the position and / or orientation of the docking station (10) relative to the vehicle (30) is known based on the known position of the imaging sensor (31) and the known arrangement of the imaging sensor on the vehicle (30) (S4), a coordinate transformation (S4a) from the system of the imaging sensor (31) to the system of the vehicle (30) is performed.

10. A computer program product for docking a vehicle (30) to a docking station (10), wherein: The computer program product is implemented to be loaded into a memory of a computer; and It comprises software code parts, by which the method steps of the method according to any one of claims 5 to 9 are implemented when the computer program product is run on the computer.

Images