Method for generating a raster map of the static infrastructure of an environment, method for generating a trained learning model for generating such a raster map, computer program product and electronic computing device

Abstract

The invention relates to a method for generating a raster map of the static infrastructure (10) of an environment (12) for an assistance system (14) of a motor vehicle (10) by means of an electronic computing device (18) of the assistance system (14), comprising the steps of: - providing a trained learning model (20) for generating the raster map of the static infrastructure (10) by means of the electronic computing device (18); - receiving a plurality of fleet data (22) for the raster map of the static infrastructure (10) by means of the electronic computing device (18); - generating a raster map of the fleet data (26) depending on the received fleet data (22) by means of the electronic computing device (18); and - generating the raster map of the static infrastructure (10) by evaluating the raster map of the fleet data (26) by means of the trained learning model (20) by means of the electronic computing device (18).Furthermore, the invention relates to a method for generating a trained learning model (20), a computer program product and an electronic computing device (18).

Description

[0001] The invention relates to a method for generating a raster map of the static infrastructure of an environment for an assistance system of a motor vehicle by means of an electronic computing device of the assistance system according to claim 1. Furthermore, the invention also relates to a method for generating a trained learning model for generating a raster map of the static infrastructure of an environment for an assistance system of a motor vehicle, a corresponding computer program product and an electronic computing device.

[0002] Behavioral planning, particularly for partially or fully automated vehicles, and prediction of other vehicles participating in traffic are essential components of autonomous vehicles. State-of-the-art approaches to behavioral planning and prediction are based on raster maps. The environment is represented in a bird's-eye view raster map. These raster maps are generated from the automated vehicle's perception of surrounding vehicles, raw sensor data, and high-definition (HD) maps.

[0003] The input data for raster map-based approaches is therefore the raster map of the environment. A raster map can be seen as a generalization of an image. Different information is represented in different layers of the raster map, which are also called channels. A typical image, for example, has three layers (RGB), while a raster map can contain significantly more layers, depending on the available information.

[0004] A state-of-the-art, machine-learning-based method is subsequently used to train a convolutional neural network (CNN) or other machine-learning-based image processing methods, such as Vision Transformer, in order to either plan a trajectory from this given input data (e.g., for behavioral planning) or predict the movement of surrounding agents (prediction). For raster map-based approaches, current methods transfer the information from the high-definition map into the raster map layer, a process known as rendering. This information can include, for example, road geometries, crosswalks, or other elements of the static infrastructure. These layers containing this information can be referred to as the raster map of the static infrastructure.For example, one layer of the raster map of the static infrastructure can contain the lane boundaries that come from the HD map.

[0005] HD maps already exist for some areas of the world. The raster map layers of the static infrastructure listed above can easily be derived from these HD maps. However, it is questionable how raster map layers, which are typically derived from HD maps, can be provided when no HD map is available.

[0006] Therefore, there is a need in the state of the art to generate a raster map of the static infrastructure, even if no HD map information is available.

[0007] DE 10 2023 003 226 A1 relates to a method for generating raster maps. The method according to the invention is characterized by the following process steps: providing aerial photographs of a target region; reading the aerial photographs into an artificial intelligence-based conversion model, wherein the conversion model is configured to read aerial photographs, recognize static infrastructure in the aerial photographs, and generate a raster map based on the recognized static infrastructure, storing different types of static infrastructure on a respective individual layer of the raster map; and generating the raster map of the target region by the conversion model.

[0008] The object of the present invention is to provide a method for generating a raster map of the static infrastructure of an environment for an assistance system of a motor vehicle, a method for generating a trained learning model for generating a raster map of the static infrastructure of an environment for an assistance system of a motor vehicle, a computer program product and an electronic computing device by means of which a raster map of the static infrastructure can be reliably realized even without the provision of an HD map.

[0009] This problem is solved by a method for generating a raster map of the static infrastructure of an environment for a motor vehicle assistance system using an electronic computing device, by a method for generating a trained learning model for generating a raster map of the static infrastructure of an environment for a motor vehicle assistance system, using a computer program product, and an electronic computing device according to the independent claims. Advantageous embodiments are specified in the dependent claims.

[0010] One aspect of the invention relates to a method for generating a raster map of the static infrastructure of an environment for a motor vehicle assistance system using an electronic computing device of the assistance system. A trained learning model is provided for generating a raster map of the static infrastructure using the electronic computing device. A large number of fleet data for the static infrastructure are received by the electronic computing device. A raster map of the fleet data is generated by the electronic computing device based on the received fleet data. The raster map of the static infrastructure is generated by evaluating the raster map of the fleet data using the trained learning model via the electronic computing device.

[0011] This makes it possible to generate the raster map of the static infrastructure even without an HD map and solely based on the received fleet data.

[0012] In particular, the invention proposes using fleet data instead of HD maps. Compared to HD maps, this fleet data can be obtained globally and without manual labeling. This approach uses a neural network to approximate the corresponding raster map layer of the static infrastructure from fleet data. Training can be performed on a dataset for which actual HD maps are available. During training, the learning model learns how to derive the raster map of the static infrastructure, which was generated using an actual HD map, from fleet data.

[0013] In particular, the invention offers the further advantages that an HD map is only required for training in a local area. Fleet data, on the other hand, is readily available and readily available in large quantities. Specifically, the invention thus enables the use of raster map-based algorithms in areas not covered by HD maps. HD maps are considered particularly complex to generate. Generating the corresponding raster map of the static infrastructure directly using the proposed method results in a significant saving of effort.

[0014] It is also planned that an essential component of a corresponding raster map of the static infrastructure will include the encoding of lane directions using unit vectors. This is precisely where fleet data offers a decisive advantage compared to the sole use of, for example, aerial photographs. Fleet data contains dynamic information. This information is converted into a format that can be processed by, for example, an encoder-decoder model using the described algorithmic transformation into the fleet data raster map. In particular, the normalized direction vectors of the fleet data provide enormous added value for the encoder-decoder model, enabling it to derive the lane directions relatively easily and robustly.

[0015] As mentioned earlier, an HD map is specifically a high-definition map. This is a digital map that offers a high level of detail and resolution. The data points are densely distributed, providing detailed information about streets, buildings, trees, and other objects. HD maps often contain various layers with different information, such as traffic areas, footpaths, building plans, terrain profiles, and the like.

[0016] High-definition (HD) maps are updated regularly to reflect changes in infrastructure. Sources for HD maps include digital surveying and mapping by government or private companies. Satellite imagery and aerial photography, processed using specialized techniques, can also be used. HD maps are particularly well-known in navigation systems, where they display road networks, traffic signs, and other essential information for drivers. Furthermore, HD maps can be used to visualize infrastructure projects, plan new developments, and manage transportation.

[0017] Fleet data, also known as fleet information, refers specifically to data collected from multiple vehicles, containing information about their movement, operation, and environment. Examples of fleet data include driving trajectory, speed, acceleration and braking, navigation and route planning, and vehicle condition data. Furthermore, data on engine power, fuel consumption, tire temperature, and similar metrics can also be used. This data can be collected, for example, via sensors on the respective vehicles and transmitted to an electronic computing unit. Examples of sensors used include radar, lidar, cameras, and positioning devices.

[0018] In the present embodiment, the electronic computing device is to be regarded in particular as part of the assistance system and can, for example, be generated within the vehicle. Alternatively, it can also be provided that the electronic computing device is supplied as a central electronic computing device, and the generation of the raster map is thus carried out externally to the vehicle, and the corresponding generated raster map can then be transferred to a corresponding vehicle or assistance system of the vehicle.

[0019] According to an advantageous embodiment, a control signal for a vehicle function is generated based on the raster map of the static infrastructure, and / or a prediction of the movement of an object in the environment is generated. In other words, the raster map of the static infrastructure can be used, for example, for the navigation of a vehicle that is at least partially automated or fully automated. The control signal can be used, for example, as a function for a longitudinal acceleration device and / or a lateral acceleration device of the vehicle. In other words, the vehicle can be controlled, for example, essentially at least partially autonomously or fully autonomously (automated / fully automated), based on the raster map of the static infrastructure.Furthermore, the raster map of the static infrastructure can be used to predict the movement of an object, such as a cyclist, pedestrian, or other motor vehicle in a given environment. This enables reliable navigation for at least partially automated or fully automated vehicles based on improved generation of the raster map of the static infrastructure.

[0020] Another advantageous implementation involves deploying the trained learning model as an encoder-decoder architecture. An encoder-decoder architecture is particularly well-suited for learning complex relationships between inputs and outputs. This allows the model to be applied to various tasks instead of being trained for a single, specific one. Furthermore, the architecture enables the model's inputs and outputs to be adapted to different tasks. By modifying the encoder-decoder layers, the model can be used for different data types and formats. The encoder-decoder architecture offers particular advantages in generating raster maps, as these maps are characterized by a high level of detail. An encoder-decoder architecture can learn this detail based on various input data and represent it in the raster map.Furthermore, a trained model can be continuously fed with new data, dynamically updating the raster map. This allows changes such as new buildings, road construction, or landscaping to be detected and reflected in the map. The encoder-decoder architecture also enables the use of various data sources to create a comprehensive and accurate raster map. The encoder-decoder architecture also allows for the customization of the training dataset and model parameters to meet specific needs.

[0021] It is also advantageous if the raster map of the static infrastructure is generated for a predefined spatial area. In particular, the raster map of the fleet maps describes a specific size, for example, 512 x 512 pixels, and covers a specific local area of, for example, 50 x 50 meters. This allows the raster map of the static infrastructure to be generated computationally efficiently.

[0022] In a further advantageous embodiment, at least two raster maps of the static infrastructure are combined to form a single environmental raster map for a given spatial area. In particular, a large number of infrastructure raster maps can be combined to create the environmental raster map. This can also be accomplished, for example, using a suitable neural network of the electronic computing device. Thus, the raster map of the static infrastructure can be generated for a larger local area, a process that is repeated for different local areas. The raster maps are then linked together to form the environmental map, especially using the method of image stitching.

[0023] Furthermore, it has proven advantageous to generate the raster map of the static infrastructure based on at least one aerial image of the surrounding area. In particular, aerial photographs, such as satellite or drone images, can be used to further specify local regions. The actual input data is then used, namely the raster map of the fleet data, which, for example, has a format of [C, H, W], where C is the number of channels, and H and W are the altitude and latitude. An aerial photograph, for example, with three channels, has the format [3, H, W]. This means that the extended input raster map for the encoder-decoder model can be a concatenation of the raster map of the fleet data and the aerial photograph, i.e., [C + 3, H, W]. However, the training and generation of the corresponding training target itself do not need to be changed.

[0024] It is also advantageous if trajectory data from at least one other vehicle is received as fleet data. The fleet data is therefore primarily in the form of trajectory data, i.e., a temporal sequence of vehicle states. Components of this state include at least the spatial position (x and y) and the speed (vx and vy). Furthermore, it is only necessary for the vehicles in the fleet to record their position at regular intervals, for example, every second, and send it directly or at a later time to a central computer / server. Over time, a vast amount of such driven trajectories accumulates on the server. This enables the reliable generation of the raster map of the static infrastructure based on the trajectory data, which does not represent a high data load.

[0025] A further aspect of the invention relates to a method for generating a trained learning model for creating a raster map of the static infrastructure of an environment for a motor vehicle assistance system according to the preceding aspect, using an electronic computing device. A high-resolution map (HD map) of the environment is provided by the electronic computing device. Fleet data for the environment is also provided by the electronic computing device. An HD raster map of the static infrastructure is generated by the electronic computing device based on the high-resolution map, with the HD raster map serving as a reference map for generating the trained learning model. The learning model is then trained by the electronic computing device based on the HD raster map and the fleet data.

[0026] In particular, this enables the provision of so-called supervised learning. The learning model uses a raster map representation of the fleet data as input. This raster map representation is generated from fleet data in several preprocessing steps. During supervised training, the training target (ground truth) is a raster map of the static infrastructure, created from an HD map and available during training.

[0027] Initially, many pairs of input data are required for training, in particular the raster map of the fleet data, which is derived from the fleet data using the following algorithm, and the training target, in particular the so-called ground truth, i.e., the raster map of the static infrastructure created from the HD map. The creation of the input data, i.e., the raster map of the fleet data, works as follows. In principle, the invention uses fleet data in the form of trajectories. The corresponding training target for the learning model can be derived directly from the existing HD map. This means that the HD map can be directly converted into a ground truth raster map of the static infrastructure. Such techniques for rendering an HD map are already known in the prior art. After training, the trained learning model can be used in any area of ​​the world.Given fleet data, an input raster map is first created from the fleet data and is then able to generate the raster map of the static infrastructure. During training, the high-definition (HD) map is only required in a local area to derive the training target. This results in a training dataset with many pairs of the input raster map, which aggregates fleet information from multiple channels, and the training target, which is in turn generated from the locally available HD map. An encoder-decoder model is then trained to approximate the training target as closely as possible using the input raster map derived from the fleet data. It therefore learns how to generate a raster map of the static infrastructure from the algorithmically generated fleet data.This raster map of the static infrastructure is similar, due to the training process, to the raster map of the static infrastructure that can be derived from the HD map. Possible architectures for the encoder-decoder model are based, for example, on Convolutional Neural Networks (CNNs) or Vision Transformers.

[0028] Another aspect of the invention relates to a computer program product with program code means, wherein an electronic computing device is configured to perform, when the computer program product is processed by the electronic computing device, a method according to the first aspect of the invention and / or a method according to the second aspect of the invention.

[0029] Furthermore, the invention therefore also relates to a computer-readable storage medium with at least the computer program product according to the preceding aspect.

[0030] The invention also relates to an electronic computing device for generating a raster map of the static infrastructure of an environment for an assistance system of a motor vehicle and / or for generating a trained learning model for generating a raster map of the static infrastructure of an environment for an assistance system of a motor vehicle, wherein the electronic computing device is configured to carry out a method according to the first aspect of the invention and / or a second aspect of the invention.

[0031] The process for generating the raster map and the process for generating the trained learning model can be performed using a single electronic computing device or on separate electronic computing devices. The electronic computing devices, or the single electronic computing device, can be located on or in a vehicle, or they can be provided as a central, server-based system, for example, cloud-based.

[0032] Advantageous embodiments of the method according to the first aspect of the invention are also to be regarded as advantageous embodiments of the method according to the second aspect of the invention. Furthermore, the advantageous embodiments according to the first aspect of the invention are also to be regarded as applicable to the computer program product and the electronic computing device. The electronic computing device has, in particular, material features to enable it to carry out the corresponding method steps.

[0033] A computing unit / electronic computing device can be understood, in particular, as a data processing device containing a processing circuit. The computing unit can therefore process data to perform arithmetic operations. This may also include operations to perform indexed access to a data structure, such as a lookup table (LUT).

[0034] The computing unit may, in particular, contain one or more computers, one or more microcontrollers, and / or one or more integrated circuits, for example, one or more application-specific integrated circuits (ASICs), one or more field-programmable gate arrays (FPGAs), and / or one or more systems on a chip (SoCs). The computing unit may also contain one or more processors, for example, one or more microprocessors, one or more central processing units (CPUs), one or more graphics processing units (GPUs), and / or one or more signal processors, in particular one or more digital signal processors (DSPs). The computing unit may also include a physical or virtual array of computers or other units of the aforementioned type.

[0035] In various embodiments, the computing unit includes one or more hardware and / or software interfaces and / or one or more storage units.

[0036] A storage unit can be volatile data storage, for example as dynamic random access memory (DRAM) or static random access memory (SRAM), or as non-volatile data storage, for example as read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or flash EEPROM, ferroelectric random access memory (FRAM), or magnetoresistive random access memory.It can be designed as MRAM (magnetoresistive random access memory) or as phase-change random access memory, PCRAM (phase-change random access memory).

[0037] Further advantages, features, and details of the invention will become apparent from the following description of a preferred embodiment and from the drawings. The features and combinations of features mentioned above in the description, as well as those mentioned below in the figure description and / or shown in the figure alone, can be used not only in the combinations specified, but also in other combinations or individually, without departing from the scope of the invention.

[0038] This shows: Fig. 1. A schematic block diagram for generating a grid map of the static infrastructure of an environment for a motor vehicle assistance system; and Fig. 2. Another schematic block diagram for generating a trained learning model for generating a raster map of the static infrastructure according to Fig. 1.

[0039] In the figures, identical or functionally equivalent elements are provided with the same reference symbols.

[0040] Fig. Figure 1 shows a schematic block diagram according to an embodiment of a method for generating a raster map of the static infrastructure 10 of an environment 12 for an assistance system 14 of a motor vehicle 16 using an electronic computing device 18. In particular, it is provided that a trained learning model 20 is specified for generating the raster map of the static infrastructure 10 using the electronic computing device 18. A multitude of fleet data 22, in particular a multitude of additional motor vehicles 24, is received for the raster map of the static infrastructure 10 using the electronic computing device 18. A raster map of the fleet data 26 is generated using the electronic computing device 18. The raster map of the static infrastructure 10 is then generated by evaluating the raster map of the fleet data 26 using the trained learning model 20 via the electronic computing device 18.

[0041] In an advantageous embodiment, it can be provided that, depending on the grid map of the static infrastructure 10, a control signal 28 for a functional device 30, for example in the form of a longitudinal acceleration device and / or lateral acceleration device, of the motor vehicle 16 is generated. Furthermore, a prediction of the movement of an object in the environment 12 can also be generated.

[0042] The trained learning model 20 can be implemented in particular in the form of an encoder-decoder architecture 32, as for example in the Fig. Figure 2 is provided. Furthermore, it may be provided that the raster map of the static infrastructure 10 is generated for a given spatial area of ​​the environment 12. In particular, it may be provided that at least two raster maps of the static infrastructure 12 of the respective spatial area of ​​the environment 12 are combined to form an environment raster map.

[0043] Furthermore, the Fig. 1. Furthermore, the raster map of the static infrastructure 10 can also be generated based on at least one aerial photograph 34 of the surroundings 12. It can also be provided that trajectory data 22 from at least one motor vehicle, and in particular from the multitude of other motor vehicles 24, are received as fleet data.

[0044] Fig. Figure 2 shows a schematic block diagram for generating a trained learning model 20, in particular for generating the raster map of the static infrastructure 10 of the environment 12 from the Fig.1. In this process, a high-resolution map 36 for the environment 12 is provided by means of the electronic computing device 18 or by means of another electronic computing device. Furthermore, the fleet data 22 or other fleet data for the environment 12 are provided by means of the electronic computing device 18. A high-resolution raster map of the static infrastructure 38 is generated based on the high-resolution map 36 by means of the electronic computing device 18, whereby the high-resolution raster map is used as a reference map for the trained learning model 20. The training of the learning model then takes place based on the high-resolution raster map and the fleet data 22 by means of the electronic computing device 18. For this purpose, an encoder-decoder architecture 32 can be provided, as shown here.

[0045] Part of the invention is thus based on the trained learning model 20, which is based on supervised learning. The model uses a raster map representation of the fleet data 22 as input. This raster map representation is obtained from the fleet data 22 in several preprocessing steps. During supervised training, a raster map of the static infrastructure 10, created from the HD map available during training, is used as the training target (ground truth).

[0046] Therefore, for the training, many pairs of input data (raster map of fleet data 26, which are obtained from fleet data 22 using the following algorithm) and the training target (ground truth, i.e., raster map of the static infrastructure created from HD map 36) are necessary.

[0047] The creation of the input data, i.e., the raster map of the fleet data 22, works as follows. In principle, the invention uses fleet data 22 in the form of trajectories. For this, it is only necessary that vehicles in the fleet record their position at regular intervals (e.g., every second) and send it directly or at a later time to a central server. Over time, a huge number of such driven trajectories accumulate on the server. This arbitrarily large number of trajectories is then used effectively to generate high-definition maps.

[0048] Part of the invention therefore provides for the creation of an input raster map, namely the raster map of the fleet data 22. The input raster map has a specific size (e.g., 512x512 pixels) and covers a specific spatial area (e.g., 50x50 meters). Using the neural network presented later, and after successful training, the raster map of the static infrastructure 10 for the same spatial area can be generated from the input raster map. If the environmental map is to be created for a larger spatial area, the process can be repeated multiple times with different spatial areas, and the resulting raster maps of the static infrastructure 10 can be linked (image stitching).

[0049] The creation of the input grid map is therefore based on fleet data 22. The fleet data 22 is in the form of trajectories, i.e., a temporal sequence of vehicle states. At least one of the state parameters is the local position (x and y) and the speed (vx and vy).

[0050] Each pixel i of the input raster map is calculated as follows: Pixel i covers a specific local area (in world coordinates). For example, with 512x512 pixels and 50x50 meters, each pixel covers a local area of ​​9.77x9.77 centimeters. A query is performed on the fleet data database for this local area: All vehicle states whose position (x and y) lies within the local area of ​​pixel i are queried. This results in a number N of vehicle states for the local area of ​​pixel i. An aggregation then takes place. The aggregated values ​​are written directly to the corresponding pixel i of the input raster map. The pixel's channels are calculated as follows: The first channel contains the number of states found in the database for the area of ​​pixel i.This corresponds to N; in a second channel, the mean of all N values ​​of vx is calculated; in a third channel, the mean of all N values ​​of vy is calculated.

[0051] The mean values ​​of vx and vy are then normalized: (nxny)=1(vx2+vy2)(vxvy) nx and ny can thus be viewed as normalized direction vectors. These indicate the direction of traffic flow at pixel i. nx forms the fourth channel of the raster map. ny forms the fifth channel of the raster map.

[0052] In a sixth channel, the variance or standard deviation of all N values ​​of vx is calculated. In a seventh channel, the variance or standard deviation of all N values ​​of vx is calculated. In eighth and ninth channels, the standard deviations of vx and vy are also multiplied by the factor. 1(vx2+vy2) Normalized. In further channels, mean values ​​and standard deviations of all N values ​​of any components of the state can be used. Examples of this are turn signal states. At any point in the trajectory, a vehicle in the fleet can also have a turn signal state (right or left). This is a binary state (on or off). This can also be averaged, and a standard deviation can be calculated. Steering angle, acceleration in the x-direction, and acceleration in the y-direction also represent such states.

[0053] The training target can be derived directly from an existing HD map 36. This means that the HD map 36 can be directly converted into a ground-truth raster map of the static infrastructure 10. Such techniques for "rendering" an HD map 36 are state of the art. After training, the trained learning model 20 can be used in any area of ​​the world: Given fleet data 22, it first creates an input raster map from the fleet data 22 and is then able to generate the raster map of the static infrastructure 10. During training, the HD map 36 is only necessary in a local area to derive the training target.

[0054] This results in a training dataset with many pairs: an input raster map (which aggregates fleet information in multiple channels) and a training target (generated from a locally available HD map 36). An encoder-decoder model is then trained to approximate the training target as closely as possible, given the input raster map (from fleet data 22). It learns how to generate the raster map algorithmically from fleet data 22 into the raster map of the static infrastructure 10. Due to the training process, this raster map of the static infrastructure 10 resembles the HD raster map of the static infrastructure 38, which can be derived from HD maps 36. Possible architectures for the encoder-decoder model are based on convolutional neural networks (CNNs) or vision transformers.

[0055] In an extension of the invention disclosure, an aerial photograph 34 (e.g., satellite image or drone footage) of the local region can also be used. This can be easily combined: The actual input data (i.e., the raster map of the fleet data 26) has, for example, a format of [C, H, W], where C is the number of channels, and H and W are the height and width. An aerial photograph (here with three channels) has the format [3, H, W]. This means that the extended input raster map for the encoder-decoder model can be a concatenation of the raster map of the fleet data 26 and the aerial photograph (i.e., then [C+3, H, W]). No changes need to be made to the training or the generation of the training target. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2023 003 226 A1

[0007]

Claims

[1] Method for generating a raster map of the static infrastructure (10) of an environment (12) for an assistance system (14) of a motor vehicle (10) by means of an electronic computing device (18) of the assistance system (14), comprising the steps: - providing a trained learning model (20) for generating the raster map of the static infrastructure (10) by means of the electronic computing device (18); - receiving a plurality of fleet data (22) for the raster map of the static infrastructure (10) by means of the electronic computing device (18); - generating a raster map of the fleet data (26) as a function of the received fleet data (22) by means of the electronic computing device (18); and - generating the raster map of the static infrastructure (10) by evaluating the raster map of the fleet data (26) by means of the trained learning model (20) by means of the electronic computing device (18). [2] Method according to claim 1, characterized by that, depending on the raster map of the static infrastructure (10), a control signal (28) for a functional device (30) of the motor vehicle (10) is generated and / or a prediction of a movement of an object in the environment (12) is generated. [3] Method according to claim 1 or 2, characterized by that the trained learning model (20) is provided as an encoder-decoder architecture (32). [4] Method according to one of the preceding claims, characterized by that the raster map of the static infrastructure (10) is generated for a given spatial area of ​​the environment (12). [5] Method according to claim 4, characterized by that at least two raster maps of the static infrastructure (10) for a respective predetermined spatial area of ​​the environment (12) are combined to form an environmental raster map. [6] Method according to one of the preceding claims, characterized by that the raster map of the static infrastructure (10) is additionally generated as a function of at least one aerial photograph (34) of the surroundings (12). [7] Method according to one of the preceding claims, characterized by that trajectory data from at least one further motor vehicle (24) are received as fleet data (22). [8] Method for generating a trained learning model (20) for generating a raster map of the static infrastructure (10) of an environment (12) for an assistance system (14) of a motor vehicle (10) according to one of claims 1 to 7 by means of an electronic computing device (18), comprising the steps: - providing a high-resolution map (36) for the environment (12) by means of the electronic computing device (18); - Providing fleet data (22) for the environment (12) by means of the electronic computing device (18); - generating an HD raster map of the static infrastructure (38) as a function of the high-resolution map (36) by means of the electronic computing device (18), wherein the HD raster map is used as a reference map for generating the trained learning model (20); and - Training the learning model as a function of the HD raster map and the fleet data (22) by means of the electronic computing device (18). [9] Computer program product with program code means, wherein an electronic computing device (18) is designed to carry out a method according to one of claims 1 to 7 and / or a method according to claim 8 when the computer program product is processed by the electronic computing device (18). [10] Electronic computing device (18) for generating a raster map of the static infrastructure (10) of an environment (12) for an assistance system (14) of a motor vehicle (16) and / or for generating a trained learning model (20) for generating a raster map of the static infrastructure (10) of an environment (12) for an assistance system (14) of a motor vehicle (16), wherein the electronic computing device (18) is designed to carry out a method according to one of claims 1 to 7 and / or a method according to claim 8.

Images