Method for determining a parking space and a target position for a vehicle in the parking space

Abstract

The invention relates to a method for determining a parking space and a target position for a vehicle (1) in the parking space by means of a neural network (4), the method comprising the following steps: - capturing a scene in the area surrounding a vehicle (1) by means of a sensor system (2) and providing surroundings information (S10); - generating a grid map (3) of the scene containing a multiplicity of cells (3.1) on the basis of the surroundings information, the cells (3.1) of the grid map (3) each comprising information regarding whether or not the area of the surroundings that corresponds to the cell (3.1) is occupied by an object (O1, O2) (S11); - transmitting the grid map (3) to a neural network (4), the neural network (4) being trained to take the grid map (3) as a basis for providing information relating to a parking space boundary frame (P) and information relating to a target position (Z) for the vehicle (1) in the parking space boundary frame (P) (S12).

Description

[0001] Description

[0002] Method for determining a parking space and a target position of a vehicle in the parking space

[0003] The invention relates to a method and a system for determining a parking space and a target position of a vehicle in the parking space by means of a neural network.

[0004] Methods for determining a parking space and a vehicle's target position within the parking space are known from the prior art. Based on the environmental information acquired by sensors, an environmental model is first created. Based on this environmental model, the scene is then interpreted, a parking space is detected, and the vehicle's target position within the parking space is determined. Complex algorithms are used for this process, employing a multitude of if-then queries and lookup tables.

[0005] The problem with existing methods is that the software code implementing the process for determining the parking space or the vehicle's target position is very complex due to the multitude and often varying customer requirements, making it difficult to maintain or expand. Furthermore, the existing methods cannot handle all parking situations satisfactorily.

[0006] Based on this, the object of the invention is to provide a method for determining a parking space and a target position of a vehicle in the parking space, which avoids complex program structures and enables reliable and satisfactory parking space detection and determination of the vehicle's target position. This object is achieved by a method having the features of independent patent claim 1. Preferred embodiments are the subject of the dependent claims. A system for determining a parking space and a target position of a vehicle is the subject of the independent patent claim 12.

[0007] According to a first aspect, the invention relates to a method for determining a parking space and a target position of a vehicle using a neural network. The method comprises the following steps:

[0008] First, a scene in the surrounding area of ​​a vehicle is captured using sensors, and environmental information is provided. The sensors can comprise any type of sensor suitable for environmental detection, such as ultrasonic sensors, radar sensors, LIDAR sensors, cameras, etc. The environmental information is, for example, two-dimensional data in which the detections recorded by the sensors are mapped.

[0009] A raster map of the scene is then generated with a large number of cells based on the environmental information. The cells of the raster map each contain information as to whether the surrounding area corresponding to the cell is occupied by an object or not. The raster map thus essentially forms a digital image, and the cells form the pixels of the image, with the pixels containing occupancy information, for example "0" for unoccupied and "1" for occupied. The cells can preferably also contain further information, for example height information on the detected objects, the density and / or intensity of the reflections received at the sensor, etc. The density of the reflections received at the sensor can indicate how many feature points fall into a certain cell.The intensity of the reflections received at the sensor can indicate the signal strength of the received reflection, which in turn depends on the material and / or surface texture of the reflecting object.

[0010] The raster map is then transmitted to a neural network. The neural network is trained to provide information about a parking space bounding box and information about a target position of the vehicle in the parking space bounding box based on the raster map.

[0011] The technical advantage of the method according to the invention is that the neural network achieves reliable and satisfactory parking space detection and determination of the vehicle's target position. Furthermore, the use of the neural network allows several state-of-the-art software components to be replaced simultaneously, namely the software component for interpreting the scene and the software component for calculating the target position.

[0012] According to one embodiment, the cells of the raster map contain information indicating whether the respective cell is occupied by an object that is higher or lower than a specified threshold. This allows the neural network to be provided with additional information regarding the object's height or the height class (high or low relative to a specified threshold) into which the object should be classified. This significantly improves parking space detection, as, for example, tall objects such as vehicles, etc., can be distinguished from short objects such as curbs.

[0013] According to one embodiment, the cells of the raster map

[0014] Information about the density and / or intensity of the reflections received at the sensor is recorded. This information can be used to weight the cells, such that those cells with a higher density and / or intensity are given greater weight by the neural network than those with a lower density and / or intensity.

[0015] According to one embodiment, the neural network comprises several sections, namely a first section for detecting structures of the objects contained in the raster map, a second section for determining one or more areas relevant for parking space detection, and a third section that receives the at least one area relevant for parking space detection and, within the at least one area relevant for parking space detection, determines information about a parking space boundary frame and information about a target position of the vehicle in the parking space boundary frame. The neural network thus has a Fast R-CNN architecture, or Faster R-CNN architecture, which offers high efficiency and detection accuracy in parking situations.

[0016] According to one embodiment, the first section has a convolutional neural network (CNN) with multiple layers, wherein the layers each have a convolutional layer (i.e. a convolutional layer) and a pooling layer (i.e. a pooling layer or bundling layer), and wherein each layer provides a feature map (also called an activation map or feature map) as output information. The layers of the convolutional neural network are designed to recognize features of varying complexity from the information contained in the raster map and to output them in the feature map. The different feature maps thus provide output information of varying complexity. According to one embodiment, the output information of the respective convolutional layer is modified by an activation function.The activation function can, in particular, be a ReLU activation function. This can increase the computational efficiency and convergence capability of the neural network.

[0017] According to one embodiment, the second section comprises a selective search algorithm or a convolutional neural network configured to detect and select one or more sections in the feature maps provided by the respective layers of the convolutional neural network of the first section and containing one or more relevant areas for detecting parking spaces. Thus, the second section of the neural network detects those areas in the feature maps that are relevant for parking space detection. Such an area can be, for example, a parking space with the objects bordering this parking space.

[0018] According to one embodiment, the selective search algorithm or the convolutional neural network receives multiple feature maps from different layers of the convolutional neural network of the first section. Based on the overall information contained in the feature maps, one or more relevant areas for parking space detection are determined. In other words, information from several or all feature maps is used to determine relevant areas for parking space detection.

[0019] According to one embodiment, the third section has at least a first fully connected layer. The third section generates at least a portion of the feature maps provided by the first section of the neural network based on the at least one relevant region provided by the second section. This at least one portion of the feature maps is further processed by the at least one first fully connected layer. Thus, in the third section, the information contained in the feature maps is reduced to the regions relevant for parking space detection and further processed.

[0020] According to one embodiment, the third section comprises a pooling layer that generates the at least one section of the feature maps. The at least one section has a predefined size. Preferably, the pooling layer provides multiple sections that are the same size, regardless of the size of the relevant regions that led to the generation of the sections. The equal size of the sections makes it possible to simplify the further processing of the information. In particular, all sections can be processed simultaneously.

[0021] According to one embodiment, the third section comprises at least a second fully-connected layer for calculating the parking space boundary frame and at least a third fully-connected layer for calculating the target position of the vehicle, wherein the second and the third fully-connected layer are connected to the at least one first fully-connected layer and receive output information therefrom.

[0022] According to one embodiment, the first and / or second section provides a plurality of different pieces of information. These different pieces of information are processed in parallel by a plurality of third sections of the neural network. In particular, information from different feature maps generated by the first section can be further processed in different third sections of the neural network. The output information generated thereby can then be combined in a further information processing step to provide the information relating to a parking space boundary frame and information relating to a target position of the vehicle within the parking space boundary frame.

[0023] According to a further aspect, the invention relates to a system for determining a parking space and a target position of a vehicle using a neural network. The system is coupled to a sensor system designed to detect a scene in the surrounding area of ​​the vehicle and to provide environmental information. The system has a computing unit designed to generate a raster map of the scene with a plurality of cells based on the environmental information. The cells of the raster map each contain information as to whether or not the surrounding area corresponding to the cell is occupied by an object. The neural network is trained to determine information about a parking space boundary frame and information about a target position of the vehicle in the parking space boundary frame based on the raster map.

[0024] The technical advantage of the system according to the invention is that the neural network achieves reliable and satisfactory parking space detection and determination of the vehicle's target position. Furthermore, the neural network can perform data-driven determination of the parking space boundary frame and the vehicle's target position, rather than using complex geometric algorithms.According to one embodiment of the system, the neural network has several sections, namely a first section designed to recognize structures of the environmental objects contained in the raster map, a second section designed to determine one or more areas relevant for parking space detection, and a third section designed to receive the at least one area relevant for parking space detection and to determine information about a parking space boundary frame and information about a target position of the vehicle in the parking space boundary frame within the at least one area relevant for parking space detection. The neural network thus has a Fast R-CNN architecture, or Faster R-CNN architecture, which offers high efficiency and detection accuracy in parking situations.

[0025] According to one embodiment of the system, the second section comprises a selective search algorithm or a convolutional neural network configured to detect and select one or more sections in the feature maps provided by the respective layers of a convolutional neural network of the first section and containing one or more relevant areas for detecting parking spaces. Thus, the second section of the neural network detects those areas in the feature maps that are relevant for parking space detection. Such an area can, for example, be a parking space with the objects bordering this parking space.

[0026] According to one embodiment of the system, the selective search algorithm or the convolutional neural network is configured to receive multiple feature maps from different layers of the convolutional neural network of the first section and to determine one or more relevant areas for parking space detection based on the overall information contained in the feature maps. In other words, information from several or all feature maps is used to determine relevant areas for parking space detection.

[0027] According to one embodiment of the system, the third section has at least one first fully connected layer. At least one section of the feature maps provided by the first section of the neural network is generated based on the at least one relevant region provided by the second section. This at least one section of the feature maps is further processed by the at least one first fully connected layer.

[0028] According to one embodiment of the system, the third section comprises at least a second fully connected layer for calculating the parking space boundary frame and at least a third fully connected layer for calculating the target position of the vehicle. The second and third fully connected layers are connected to the at least one first fully connected layer and receive output information therefrom.

[0029] The terms “approximately”, “essentially” or “about” mean, in the sense of the invention, deviations from the exact value by + / - 10%, preferably by + / - 5% and / or deviations in the form of changes that are insignificant for the function.

[0030] Further developments, advantages, and possible applications of the invention will become apparent from the following description of exemplary embodiments and from the figures. All described and / or illustrated features, individually or in any combination, are fundamentally part of the invention, regardless of their summary in the claims or their reference back to them. The content of the claims is also incorporated into the description.

[0031] The invention is explained in more detail below with reference to exemplary embodiments and the figures. They show:

[0032] Fig. 1 shows, by way of example, a schematic plan view of a vehicle with a sensor system and a computing unit connected to this sensor system;

[0033] Fig. 2 shows, by way of example, a schematic representation of a raster map illustrating a parking situation and in which an area relevant for parking space detection, a parking space boundary frame and a target position of the vehicle in the parking space boundary frame are outlined;

[0034] Fig. 3 shows an exemplary schematic representation of a convolutional neural network with several layers;

[0035] Fig. 4 shows, by way of example, a schematic representation of a neural network designed to determine a parking space boundary frame and a target position of the vehicle in the parking space boundary frame; and

[0036] Fig. 5 shows an example of a block diagram illustrating the sequences of a method for determining a parking space boundary frame and a target position of the vehicle within the parking space boundary frame. Figure 1 shows a vehicle 1 by way of example and in a highly schematic manner. The vehicle 1 has a sensor system comprising a plurality of individual sensors distributed around the vehicle, by means of which environmental detection is possible. The sensors can, for example, include ultrasonic sensors, at least one camera, at least one radar sensor, and / or at least one LIDAR sensor.

[0037] The sensor system 2 is coupled to at least one computing unit R, by means of which the method described below for detecting a parking space and a target position of a vehicle in the parking space is carried out. In particular, a neural network 4 is implemented in the computing unit R, by means of which information about a parking space boundary frame P and information about a target position Z of the vehicle 1 in the parking space boundary frame P can be determined.

[0038] The sensor system 2 of the vehicle 1 provides environmental information about a scene in the surrounding area of ​​the vehicle 1. To enable the neural network 4 to process the environmental information quickly and efficiently (in terms of computing resources), a raster map 3 is generated that depicts the scene. Figure 2 shows an exemplary and schematic representation of a raster map 3 with a scene depicting a parking situation.

[0039] The raster map 3, for example, is a two-dimensional, discretized representation of the scene. It has a plurality of cells 3.1, each of which is assigned to an environmental area. Each cell is assigned a piece of digital information indicating whether or not the cell 3.1 is occupied by an object. If the sensor system provides environmental information that includes a height classification of the objects, the cell can also contain information about the height class of the section of the object that falls into the respective cell. For example, the cell can contain information about whether the object is a tall or short object (high / low with respect to a height threshold).

[0040] In the raster map 3 according to Fig. 2, two lines are drawn which outline the contours of objects 01, 02 detected by the sensor system 2.

[0041] The neural network 4 determines an area B1 relevant for parking space detection, which area at least partially contains the two objects O1, O2. In addition, Fig. 2 also shows a parking space boundary frame P and a target position Z of the vehicle 1 in the parking space boundary frame P, wherein the parking space boundary frame P and the target position Z of the vehicle 1 are output at the output interface of the neural network 4. For the parking space boundary frame P and / or the target position Z, the neural network 4 preferably also determines an angle which indicates the orientation of the parking space boundary frame P or the target position Z in the scene.

[0042] As shown in Fig. 4, the neural network 4 has several sections that contribute to determining the position and size of the parking space boundary box P and the position of the target position Z of the vehicle 1 within this parking space boundary box P. In particular, the neural network 4 forms a so-called Fast R-CNN or Faster R-CNN, where R-CNN is a so-called Region Based Convolutional Neural Network.

[0043] In particular, the neural network 4 has a first section 4.1 which comprises a convolutional neural network 5 (CNN). A second section 4.2 connected to the first section 4.1 has means for determining areas that are relevant for parking space detection. In Fig. 2, such an area is labeled B1. Such an area is characterized in particular by the fact that it has at least one parking space that is laterally delimited by one or more objects. A third section 4.3 of the neural network 4, connected to the first and second sections 4.1, 4.2, has means for determining the position and size of the parking space boundary frame P and the position of the target position Z of the vehicle 1 in this parking space boundary frame P.

[0044] Fig. 3 shows an exemplary and schematic view of a convolutional neural network 5 that can be used in the first section of the neural network 4.

[0045] The convolutional neural network 5 receives the raster map 3 at its input. The convolutional neural network 5 has several layers 5.1, 5.2. In the illustrated embodiment, only two layers are provided. It should be noted that the convolutional neural network 5 can also have more than two layers 5.1, 5.2, depending on the level of structures to be recognized.

[0046] Each layer 5.1 has at least one convolutional layer CL and one max-pooling layer MPL. The convolutional layer CL has a filter (so-called kernels) to detect features in the information of the raster map 3. The layered arrangement of several convolutional layers CL allows increasingly complex structures to be recognized in the scene. For example, a first convolutional layer CL can recognize basic structures such as horizontal, vertical, or diagonal edges. Based on this, a second convolutional layer CL, which follows the first convolutional layer CL in the direction of information flow, can recognize patterns such as curves, rectangles, or circles. A third convolutional layer CL, if required, which follows the second convolutional layer CL in the direction of information flow, can then, based on this information, recognize more complex structures such as vehicles, gaps between objects, etc.

[0047] For example, in each layer 5.1, 5.2, a convolutional layer CL is followed by a max-pooling layer MPL. The max-pooling layer MPL serves to reduce the information processed by the subsequent layer in order to allow the kernel of the subsequent layer to provide a zoom-like perspective on the scene, which has already been reduced to detected features. For example, the respective max-pooling layer MPL reduces the information by the scaling factor ß, i.e. the convolutional layer of the first layer 5.1 has the dimension I * w, which is equal to the dimension of the raster map 3 and provides m feature maps, whereas the convolutional layer of the second layer 5.2 has the dimension and delivers n2 feature maps.

[0048] The convolutional layer CL preferably uses an activation function, for example a ReLU activation function of the form:

[0049] Fig. 4 shows the overall structure of the neural network 4. The information flow in Fig. 4 is from bottom to top, as indicated by the arrows.

[0050] A raster map 3 containing the discretized environmental information is transmitted to the convolutional neural network 5, which, as previously described, has several layers 5.1, 5.2 and generates several feature maps as output information based on these layers 5.1, 5.2. Sections 4.1 to 4.3 of the neural network 4 are also outlined in Fig. 4, with the convolutional neural network 5 located in the first section 4.1. The convolutional neural network 5 transmits the feature maps to the second section 4.2 and to the third section 4.3 of the neural network 4.

[0051] The second section 4.2 of the neural network 4 implements a selective search algorithm or has another convolutional neural network, also known as a regional proposal network (RPN). The selective search algorithm or the regional proposal network is designed to define regions of interest (ROI) to be examined in the respective feature maps generated by the convolutional neural network 5. The regions to be examined are those areas in the feature maps in which parking spaces are expected. The regional proposal network, for example, is a convolutional network pre-trained using labeled training data that is adapted to recognize parking spaces. The training data, for example, includes parking lot scenes in which areas containing one or more objects with an adjacent or enclosed open space, which could be a parking space, are labeled.By training with this data, the regional proposal network can be trained to identify areas to be examined for parking.

[0052] The areas to be examined and the feature maps are then transferred to the third section 4.3 of the neural network 4. Preferably, the components of the neural network 4 shown in Fig. 4 in the third section 4.3 are provided multiple times, namely once for each feature map generated by the convolutional neural network 5. As a result, a parking space boundary frame P and a target position Z of the vehicle 1 in the parking space boundary frame P can first be determined separately based on each feature map, with this information then being linked to one another, thereby determining a final parking space boundary frame P and a final target position Z of the vehicle 1 in this final parking space boundary frame P.

[0053] The following describes information processing based on a single feature map. The information processing of the other feature maps is carried out in the same way.

[0054] A feature map generated by the convolutional neural network 5 is linked to the at least one region to be examined, which is provided by the regional proposal network or the selective search algorithm. This defines one or more relevant regions B1, B2, B3 within the feature map that fall within the at least one region to be examined. In other words, by linking the feature map and the at least one region to be examined, at least one section A1, A2, A3 of the feature map is defined, as indicated in Fig. 4 in the lower section of Section 4.3.

[0055] Subsequently, a so-called Region of Interest (Rol) pooling is performed by a pooling layer 7. For example, a max pooling operation is used to generate a partial feature map (sections A1, A2, A3) for each region under investigation. The partial feature maps are all the same size, meaning that even if sections of the original feature map are not the same size, pooling layer 7 provides partial feature maps that are all the same size.

[0056] These partial feature maps are then forwarded to a first fully connected layer 6, which further processes the information contained in the partial feature maps. The information provided by the fully connected layer 6 is then transmitted in parallel to a second fully connected layer 8 and a third fully connected layer 9. The second fully connected layer 8 is trained, for example, to determine the parking space boundary frame P for a detected parking space. The third fully connected layer 8 is trained, for example, to determine the target position of vehicle 1 in the parking space boundary frame P or the parking space. This information is then output by the second and third fully connected layers 8, 9.

[0057] For example, the neural network 4 provides the information about the parking space boundary frame P and the target position Z of the vehicle 1 as follows. The target position Z of the vehicle 1 in a parking space is output, for example, by a vector with the following values:

[0058] Z = [tx ty t©]; where, for example, the coordinates t x and t y a corner of the vehicle is defined or an offset of this vehicle corner to the parking space boundary frame P is specified and t© specifies the orientation of the vehicle in the parking space or relative to the orientation of the parking space boundary frame P.

[0059] For example, the parking space boundary frame P is output by a vector with the following values:

[0060] P = [b x b y bi bw b©]; where, for example, the coordinates b x and b ya corner of the parking space boundary box P is defined, the values ​​bi and bw define the length and width of the parking space boundary box P and b© indicates the orientation of the parking space boundary box P in the raster map 3.

[0061] It is understood that this information is provided separately for each detected parking space boundary frame P or for each target position Z of the vehicle 1.

[0062] The training of the neural network 4 is carried out using training data that has labeled parking situation scenes, wherein the labels each indicate the aforementioned output information of the neural network 4, ie the target position Z and orientation of the vehicle 1 in the parking space and the position, size and orientation of the parking space boundary box P.

[0063] During the training of the neural network 4, an attempt is made to select the parameters or weights of the neural network 4 such that the information provided by the neural network 4 corresponds as closely as possible to the labels specified in the training data.

[0064] As explained above, the neural network 4 provides two pieces of output information, namely information on the parking space boundary frame P and the target position Z of the vehicle 1 .

[0065] The neural network 4 is trained in such a way that the total error from the information on the parking space boundary box P and the target position Z determined by the neural network 4 is minimized compared to the training data. In addition, validation data may be available that is used not as training data but for validating the training. The validation data can be used to determine when the training of the neural network 4 has led to sufficient quality. This can reduce the training time and prevent overfitting of the neural network 4, so that the neural network 4 leads to more general, i.e., less restricted, solutions.

[0066] For this purpose, a method based on minimizing the multi-task loss L is used, which takes into account both the loss due to the error in determining the parking space boundary box P and the loss due to the error in determining the target position Z. Since this is a regression problem, the loss is also referred to as regression loss.

[0067] The total loss L is defined as follows:

[0068] L (t k , and k , b k , v k ) = Lz (t k , and k ) + LP (b k , v k ); where t k over the vector [t x t y t©] the position and orientation of the target position Z of vehicle 1 for the k-th parking space t, u k over the vector [u x u y u©] the position and orientation of the target position Z of vehicle 1 for the k-th parking space in the training data, b kover the vector [b x b y bi bw b©] the position, size and orientation of the parking space boundary frame P for the k-th parking space and v k over the vector [v x v y vi v w v©] define the position, size, and orientation of the parking space bounding box P for the k-th parking space in the training data. Lz (t k , and k ) is the loss resulting from the difference in the information about the target position Z of vehicle 1 compared to the training data and Lp (b k , v k ) is the loss resulting from the difference in the information about the parking space boundary box P relative to the training data.

[0069] The calculation of the losses Lz (t k , and k ) and Lp (b k , v k ) can be done as follows:

[0070] The function smooth L1a modified loss function that allows a mixture of absolute distance and squared distance. It is defined as follows: if Ixl < 1 0.5 otherwise

[0071] Fig. 5 shows a diagram illustrating the steps of the method for determining a parking space and a target position of a vehicle in the parking space using a neural network.

[0072] First, a scene in the surrounding area of ​​a vehicle is captured using sensors and environmental information is provided (S10).

[0073] Subsequently, a raster map of the scene is generated with a plurality of cells based on the environmental information, wherein the cells of the raster map each contain information as to whether the environmental area corresponding to the cell is occupied by an object or not (S11).

[0074] Finally, the raster map is transmitted to a neural network, which is trained to provide information about a parking space boundary frame and information about a target position of the vehicle within the parking space boundary frame based on the raster map (S12). The invention has been described above using exemplary embodiments. It is understood that numerous changes and modifications are possible without thereby departing from the scope of protection defined by the patent claims.

[0075] List of reference symbols 1 vehicle

[0076] 2 Sensor technology

[0077] 3 Raster map

[0078] 3.1 Cells

[0079] 4 neural network 4.1 first section

[0080] 4.2 second section

[0081] 4.3 third section

[0082] 5 convolutional neural network

[0083] 5.1 , 5.2 Layer 6 first fully-connected layer

[0084] 7 pooling layers

[0085] 8 second fully-connected layer

[0086] 9 third fully-connected layer A1, A2, A3 cutout

[0087] B1, B2, B3 relevant area

[0088] CL convolutional layer

[0089] MPL Max-pooling layer

[0090] 01 , 02 Object P Parking space boundary frame

[0091] R arithmetic unit

[0092] Z Target position of the vehicle

Claims

Patent claims 1 ) Method for determining a parking space and a target position of a vehicle (1 ) in the parking space by means of a neural network (4), the method comprising the following steps: - detecting a scene in the surrounding area of ​​a vehicle (1) by means of a sensor system (2) and providing environmental information (S10); - generating a raster map (3) of the scene with a plurality of cells (3.1) based on the environmental information, wherein the cells (3.1) of the raster map (3) each have information as to whether the environmental area corresponding to the cell (3.1) is occupied by an object (01, 02) or not (S11); - transmitting the raster map (3) to a neural network (4), wherein the neural network (4) is trained to provide, based on the raster map (3), information about a parking space boundary frame (P) and information about a target position (Z) of the vehicle (1) in the parking space boundary frame (P) (S12). 2) Method according to claim 1, characterized in that the cells (3.1) of the raster map (3) have information as to whether the respective cell (3.1) is occupied by an object (01, 02) which is higher or lower than a specified threshold value and / or that the cells (3.1) of the raster map (3) have information on the density and / or intensity of the reflections received by the sensor system (2). 3) Method according to claim 1 or 2, characterized in that the neural network (4) has several sections (4.1, 4.2, 4.3), namely a first section (4.1) by means of which a recognition of structures of the objects contained in the raster map (3) (01, 02), a second section (3.2) by means of which one or more areas (B1, B2, B3) relevant for parking space detection are determined, and a third section (4.3) which receives the at least one area (B1, B2, B3) relevant for parking space detection and which, within the at least one area (B1, B2, B3) relevant for parking space detection, determines information about a parking space boundary frame (P) and information about a target position (Z) of the vehicle (1) in the parking space boundary frame (P). ) Method according to claim 3, characterized in that the first section (4.1) has a convolutional neural network (5) with a plurality of layers (5.1, 5.2), wherein the layers each have a convolutional layer and a pooling layer, and wherein each layer (5.1, 5.2) provides a feature map as output information.) Method according to claim 4, characterized in that the output information of the respective convolutional layer is modified by an activation function. ) Method according to one of claims 3 to 5, characterized in that the second section (4.2) comprises a selective search algorithm or a convolutional network which is designed to recognize and select one or more sections in the feature maps which are provided by the respective layers of the convolutional neural network of the first section and contain one or more relevant areas (B1, B2) for detecting parking spaces. ) Method according to claim 6, characterized in that the selective search algorithm or the convolutional network comprises several. Receives feature maps from different layers (5.1, 5.2) of the convolutional neural network (5) of the first section (4.1) and, based on the overall information contained in the feature maps, determines one or more relevant areas (B1, B2) for the detection of parking spaces. ) Method according to one of claims 4 to 7, characterized in that the third section (4.3) has at least a first fully connected layer (6), that at least one section (A1, A2, A3) of the feature maps provided by the first section (4.1) of the neural network (4) is generated based on the at least one relevant area (B1, B2, B3) provided by the second section (4.2), and that the at least one section (A1, A2, A3) of the feature maps is further processed by the at least one first fully connected layer (6). ) Method according to claim 8, characterized in that the third section (4.3) comprises a pooling layer (7) which generates the at least one section (A1, A2, A3) of the feature maps, wherein the section (A1, A2, A3) has a predefined size. 0) Method according to claim 8 or 9, characterized in that the third section (4.3) has at least a second fully-connected layer (8) for calculating the parking space boundary frame and at least a third fully-connected layer (9) for calculating the target position (Z) of the vehicle (1), wherein the second and the third fully-connected layer (8, 9) are connected to the at least one first fully-connected layer (6) and receive output information from it. 1) Method according to one of claims 3 to 10, characterized in that by the first and / or second section. (4.1, 4.2) a plurality of different partial information items are provided, and that the different partial information items are processed in parallel by a plurality of third sections (4.3) of the neural network (4). ) System for determining a parking space and a target position of a vehicle (1) by means of a neural network (4), wherein the system is coupled to a sensor system (2) which is designed to detect a scene in the surrounding area of ​​the vehicle (1) and to provide environmental information, wherein the system has a computing unit (R) which is designed to generate a raster map (3) of the scene with a plurality of cells (3.1) based on the environmental information, wherein the cells (3.1) of the raster map (3) each have information as to whether the surrounding area corresponding to the cell (3) is occupied by an object (01, 02) or not, and wherein the neural network (4) is trained to determine, based on the raster map (3), information about a parking space boundary frame (P) and information about a target position (Z) of the vehicle (1) in the parking space boundary frame (P). ) System according to claim 12, characterized in that the neural network (4) has a plurality of sections (4.1, 4.2, 4.3), namely a first section (4.1) which is designed to recognize structures of the surrounding objects contained in the raster map (3), a second section (4.2) which is designed to determine one or more areas (B1, B2, B3) relevant for parking space detection, and a third section (4.3), which is used to receive the at least one area (B1, B2, B3) relevant for parking space detection and to determine information about a parking space boundary frame (P) and information about a target position (Z) of the vehicle (1) in the. Parking space boundary frame (P) within which at least one area (B1, B2, B3) relevant for parking space detection is formed. 14) System according to claim 13, characterized in that the second section (4.2) comprises a selective search algorithm or a convolutional network designed to recognize and select one or more sections (A1, A2, A3) in the feature maps provided by the respective layers (5.1, 5.2) of a convolutional neural network (5) of the first section (4.1) and containing one or more relevant areas (B1, B2, B3) for detecting parking spaces. 15) System according to claim 14, characterized in that the selective search algorithm or the convolutional network is designed to receive a plurality of feature maps from different layers (5.1, 5.2) of the convolutional neural network (5) of the first section (4.1) and to determine one or more relevant areas (B1, B2, B3) for the detection of parking spaces based on the overall information contained in the feature maps. 16) System according to claim 14 or 15, characterized in that the third section (4.3) has at least one first fully-connected layer (6), that at least a section of the feature maps provided by the first section (4.1) of the neural network (4) is generated based on the at least one relevant area (B1, B2, B3) provided by the second section (4.2), and that the at least one section (A1, A2, A3) of the feature maps is further processed by the at least one first fully-connected layer (6). ) System according to claim 16, characterized in that the third section (4.3) has at least a second fully-connected layer (8) for calculating the parking space boundary frame (P) and at least a third fully-connected layer (9) for calculating the target position (Z) of the vehicle (1 ), wherein the second and the third fully- Connected layer (8, 9) with which at least one first fully connected layer (6) is connected and receives output information from it.

Images