Method, system, computer program and vehicle for planning trajectory

Abstract

The invention relates to a method for planning a trajectory for an at least partially automated vehicle, comprising the following steps: providing a pre-trained machine learning model, predicting an occupancy (G (t0)) of an occupancy grid (100) at a next point in time (t0) on the basis of an occupancy history; generating an occupancy sequence by means of the measurement data; evaluating the occupancy sequence by means of a machine learning model in order to predict an occupancy G (t0) of the occupancy grid (100) at the current point in time (t0); comparing the predicted occupancy (G (t0)) at the current point in time (t0) with the occupancy (B (t0)) determined by means of the measurement data, a location-dependent measure being determined for the reliability of the occupancy information as a function of the comparison; a trajectory is planned for the at least partially automated vehicle based on the reliability. The invention also relates to a system for planning a trajectory for an at least partially automated vehicle, a computer program and a vehicle.

Description

Field of the Invention

[0001] The present invention relates to a method and a system for planning a trajectory for at least a partially automated vehicle. Furthermore, the present invention also relates to a computer program and a vehicle. Background Art

[0002] Autonomous, i.e., at least partially automated guided, vehicles must be able to recognize and evaluate their surroundings. A common system chain generally includes a perception module that processes data detected by one or more environmental sensors about the vehicle's surroundings and generates information therefrom.

[0003] Two common representations of the surroundings are the object list and the occupancy grid (hereinafter also referred to simply as the grid). The object list is based on object hypotheses and describes detected objects, such as pedestrians, cyclists, or cars, using, for example, bounding boxes identifying the orientation with additional attributes such as speed or classification. On the other hand, the occupancy grid provides an image of the proximity sensors of the surroundings, in which, for example, specific lidar measurement points result in occupied cells. By directly converting the information of the proximity sensors into a grid representation, the complexity and especially the error-proneness of object formation are eliminated.

[0004] The downstream planning module utilizes this information about the vehicle's surroundings and can derive driving decisions therefrom. In this case, the conflict-freeness at the output of the perception module is checked, and a trajectory is determined for the driving task in the current situation through a cost function and, for example, by machine learning (ML) methods.

[0005] Unforeseen and unknown scenarios pose a danger to the entire system chain and thus to the safety of the driving function. This particularly concerns model-based modules, which by nature cannot reflect all possible situations, and also ML methods, which have to work with input data that was not known at design time / training time. To eliminate or mitigate these problems, out-of-distribution (OOD) techniques are suitable: Once it is ascertained at runtime that the system has reached the boundary of known situations, the situation is mitigated, for example, by intentional braking.

[0006] Such techniques are described, for example, in DE 102023205459.0, which was not published as of the filing date.

[0007] US11433922 B1 discloses a method for controlling a highly or fully automated vehicle, in which an ML model is used to output an uncertainty measure. The vehicle can determine a traveled trajectory associated with an object and determine the difference between the traveled trajectory and a set of candidate trajectories. The vehicle can determine an uncertainty measure related to the object based on the difference.

[0008] US2023 / 031375 A1 discloses a method for controlling a highly or fully automated vehicle using an ML model. The method includes a training system with training data located away from the vehicle (on-vehicle system). In addition, the ML model also uses sensor data of the vehicle to establish an obtained intent prediction or a predicted probability. Summary of the Invention

[0009] Therefore, providing a reliable method for controlling at least partially automated guided vehicles can be regarded as a task of the present invention.

[0010] Providing a highly reliable system for driving assistance for at least partially automated guided networked motor vehicles can be regarded as another task of the present invention.

[0011] In the present invention, an occupancy grid (grid) is used to represent the vehicle environment. The occupancy grid provides an image of the surrounding environment of the proximity sensor, in which for example specific lidar measurement points result in occupied cells of the occupancy grid. By directly converting the information of the proximity sensor into a grid representation, the complexity and especially the error-proneness of object formation are eliminated. In this regard, the grid presents a great advantage compared to a representation in the form of an object.

[0012] Providing reliable and safe trajectory planning for at least partially automated vehicles based on occupancy grid information can be regarded as a task of the present invention.

[0013] According to a first aspect of the present invention, a method for planning a trajectory for at least partially automated vehicles is proposed, the method comprising the following steps:

[0014] a) Providing a pre-trained machine learning model for determining the occupancy of an occupancy grid. The machine learning model is pre-trained to predict the occupancy of the occupancy grid at a subsequent time point t0 based on a history of a determined length H, i.e., a time series of the occupancy of the given occupancy grid. Here, the occupancy of the occupancy grid can be understood, for example, as a function that assigns a value to each cell (ij) of the occupancy grid, where this value indicates whether the cell is occupied or with what probability it is occupied.

[0015] b) Generate an occupancy time series by detecting measurement data from the vehicle's surroundings using at least one environmental sensor of the vehicle at times t0 - H …… t0 - 2, t0 - 1, t0 following one another and calculating the occupancy B(t0 - H) ……

[0016] B(t0 - 2), B(t0 - 1), B(t0) from the measurement data. The measurement data can be, for example, the measurement data of the vehicle's lidar sensor system. These measurement data can be collected and cached, for example, in regular 100 - millisecond time intervals.

[0017] c) Analyze and evaluate the thus - determined occupancy sequence using a pre - trained machine - learning model to predict the occupancy G(t0) of the occupancy grid at the current time point t0.

[0018] d) Now, the predicted occupancy G(t0) of the occupancy grid at the current time point t0 can be compared with the occupancy B(t0) determined from the measurement data, where, in particular, the deviation between the measurement and the prediction is identified. Based on this comparison, a location - related reliability measure of the occupancy information can be determined. The occupancy information should be understood here, in particular, as a statement about whether the spatial region represented in reality by a corresponding area (e.g., a single cell and / or a group of cells of the occupancy grid) is occupied.

[0019] e) Based on the reliability measure, a trajectory can be planned for the at least partially automated vehicle, in particular based on the identified deviation and / or uncertainty measure.

[0020] In a preferred embodiment, based on the deviation between the measurement B(t0) and the prediction G(t0), a reliability measure can be determined for the determined current occupancy spatial region. This measure can be related, in particular, to the deviation and / or measurement uncertainty, where a region includes at least one cell (ij) and / or a plurality of consecutive cells (ij) of the occupancy grid.

[0021] In a preferred embodiment, to determine the reliability, a Subjective Logic Opinion can be determined for each cell (ij) of the occupancy grid. According to the subjective logic opinion principle, for this purpose, a tuple (b ij , d ij , u ij , a ij ) can be determined, in particular, for each cell (ij), where b ij represents the degree of agreement between the measurement and the prediction of the occupancy of this cell (ij), d ij represents the deviation between the measurement and the prediction of the occupancy of this cell (ij), u ijRepresents the occupancy uncertainty of the cell (ij), and a ij Describes the base probability of the occupancy of the cell (ij). This uncertainty can also be referred to as epistemic uncertainty, which here represents the uncertainty caused by the model. Thus for example, a cell (ij) with high uncertainty or low quality in its measurement data and / or a cell (ij) with high uncertainty in its prediction has a high u ij value. By thus applying subjective logic and location assessment of the occupancy grid, a very informative and persuasive analysis of the certainty / uncertainty of the presence can be generated. This analysis can for example be directly used in the planning module in such a way that the trajectory is advantageously planned so that the trajectory avoids cells (ij) with high uncertainty u ij or the trajectory is driven at a reduced speed.

[0022] Thus in step e), the trajectory can preferably be planned for the vehicle in such a way that cells and / or cell regions with high deviation d ij and / or high uncertainty u ij are avoided, and / or such that regions with high consistency b ij are preferably selected. Here, individual cells (ij) or continuous cell regions can be considered. In this way, a trajectory that is as safe as possible can be planned advantageously and efficiently.

[0023] In other words, generating an epistemic assertion about the situation or the development of the situation in the environment of an at least partially autonomously guided vehicle can be regarded as an object of the present invention. To achieve this object, a machine learning model is used, the aim of which is to predict the occupancy G(t0) at the current time point t0 based on a defined history of length H of the occupancy grids G(t0 - 1), G(t0 - 2)... G(t0 - H) at a defined time point t0. The method is trained offline based on a suitable training data set. A great advantage of this method is that no manual labels are required for training the machine learning model. The input data (occupancy grids G(t - 1), G(t - 2)... G(t - H)) can simultaneously be used as labels for optimizing the prediction network. Finally, for implementation in an at least partially automated vehicle, the predicted occupancy is compared with the occupancy grid obtained from the current sensor information. The deviation between the prediction and the actual measurement can then be evaluated as a potentially dangerous area because the prediction cannot predict the future occupancy with sufficient quality.

[0024] Subsequently, the occupancy grid predicted in the past is compared with the occupancy grid aggregated from sensor data during runtime. This comparison can provide different outputs: In the simplest case, a cell-by-cell comparison is performed and the difference between the prediction and the sensor information is calculated for each cell (ij). This simple output already allows for a targeted identification of areas in the occupancy grid that are particularly poorly predicted. Finally, a statement about "how good the prediction is" can be derived based on this cell-by-cell criterion.

[0025] In an advantageous embodiment, the occupancy grid is transformed before being processed by the machine learning model and / or before the comparison such that the vehicle's own movement is compensated. This ensures that the occupancy grid is decoupled from the vehicle movement and can thus be compared at each point in time.

[0026] In an advantageous embodiment, the input data of the machine learning (ML) model includes 3D tensors, where each of these 3D tensors includes the occupancy grid of the determined surroundings at different points in time.

[0027] In this case, the output data of the machine learning model particularly includes an occupancy grid having the same dimensions as the occupancy grid of the input data, but only for a determined point in time t0.

[0028] In an alternative advantageous embodiment, the machine learning model has an architecture in which the spatial and temporal dimensions of the input data are processed separately.

[0029] In an alternative advantageous embodiment, the machine learning model includes a recurrent network.

[0030] According to a second aspect of the invention, a computer program is provided. The computer program includes instructions which, when executed by a processor, cause the processor to execute the method according to the first aspect.

[0031] According to a third aspect of the invention, a system for planning a trajectory for at least a partially automated vehicle is provided. The system includes a machine learning (ML) module, a perception module, an interface for receiving data from environmental sensing devices, a comparison module, and a planning unit, where the system is configured to implement the method according to the first aspect. The interface can be configured, for example, to receive or read measurement data from one or more environmental sensors of the vehicle, for example. It is also conceivable that the measurement data is provided by an external source, such as a cloud service or an infrastructure unit. For example, it can relate to measurement data of a lidar sensor and / or a camera system (mono or stereo).

[0032] According to a fourth aspect of the present invention, a vehicle is provided, which is configured to drive at least partially automatically and has a system according to the third aspect.

[0033] Thus, by means of the present invention, a judgment can be generated regarding "the perception or reliability of the state or the development of the state of the information represented by the occupancy grid", and the trajectory planning for the at least partially automated vehicle can be adapted to this perception. For example, the planning module can utilize this information and evaluate possible trajectories based on the determined uncertainty. Thus, areas with high uncertainty can be avoided, or rather, areas with high prediction quality and low perceptual uncertainty are preferably selected.

[0034] Different from the hitherto solutions for the topic of anticipatory driving, in the present invention, an out-of-distribution (OOD) method is used for the occupancy grid instead of an object list. The advantage presented by this is that the information of the proximity sensor can be directly registered in the occupancy grid without error-prone intermediate processing. Thereby, adverse false detections that often or easily occur in the object list can be avoided.

[0035] Furthermore, by combining subjective logic opinions with the location assessment of the occupancy grid, a very informative and persuasive analysis of the existing certainty / uncertainty can also be generated. This analysis can be directly used in the planning module to, for example, avoid or drive more slowly through areas with higher uncertainty.

[0036] Therefore, the present invention enables safe and efficient trajectory planning for at least partially automated vehicles.

[0037] The expression "at least partially automated" particularly includes one or more of the following cases: guiding the vehicle, especially a motor vehicle, assistively, partially automatically, highly automatically, or fully automatically.

[0038] "Guiding assistively" means that the driver of the vehicle continuously performs the lateral or longitudinal guidance of the vehicle. Different driving tasks (i.e., controlling the lateral or longitudinal guidance of the vehicle) are automatically executed respectively. That is, in the case of guiding the vehicle assistively, either the lateral guidance or the longitudinal guidance is automatically controlled.

[0039] "Partially automated guidance" means that, in specific situations (e.g., driving on a highway, driving within a parking lot, overtaking an object, driving within a lane determined by lane markings) and / or automatically controlling the longitudinal and lateral guidance of the vehicle for a certain period of time. The driver of the vehicle does not have to manually control the longitudinal and lateral guidance of the vehicle himself. However, the driver must continuously monitor the automatic control of the longitudinal and lateral guidance so that he can manually intervene when necessary. The driver must be ready to take over the vehicle guidance completely at any time.

[0040] "Highly automated guidance" means that, in specific situations (e.g., driving on a highway, driving within a parking lot, overtaking an object, driving within a lane determined by lane markings) for a certain period of time, automatically controlling the longitudinal and lateral guidance of the vehicle. The driver of the vehicle does not have to manually control the longitudinal and lateral guidance of the vehicle himself. The driver does not have to continuously monitor the automatic control of the longitudinal and lateral guidance so that he can manually intervene when necessary. When necessary, an automatic takeover request is output to the driver to request the driver to take over the control of the longitudinal and lateral guidance, especially with sufficient time margin. Therefore, the driver must potentially be able to take over the control of the longitudinal and lateral guidance. Automatically identify the boundaries of the automatic control of the longitudinal and lateral guidance. In the case of highly automated guidance, it is not possible to automatically reach the state with the least risk in every output situation.

[0041] "Fully automated guidance" means that, in specific situations (e.g., driving on a highway, driving within a parking lot, overtaking an object, driving within a lane determined by lane markings), automatically controlling the longitudinal and lateral guidance of the vehicle. The driver of the vehicle does not have to manually control the longitudinal and lateral guidance of the vehicle himself. The driver does not have to monitor the automatic control of the longitudinal and lateral guidance so that he can manually intervene when necessary. Before ending the automatic control of the longitudinal and lateral guidance, automatically request the driver to take over the driving task (control the lateral and longitudinal guidance of the vehicle), especially with sufficient time margin. If the driver does not take over the driving task, automatically return to the state with the least risk. Automatically identify the boundaries of the automatic control of the longitudinal and lateral guidance. In all situations, it is possible to automatically return to the system state with the least risk.

[0042] Driverless control or guidance means automatically controlling the longitudinal and lateral guidance of a vehicle regardless of a specific application scenario (e.g., driving on a highway, driving within a parking lot, overtaking an object, driving within a lane determined by lane markings). The driver of the vehicle himself / herself does not have to manually control the longitudinal and lateral guidance of the vehicle. The driver does not have to monitor the automatic control of the longitudinal and lateral guidance in order to be able to manually intervene when needed. Thus, for example, the longitudinal and lateral guidance of the vehicle is automatically controlled in all road types, speed ranges, and environmental conditions. Thus, the full driving task of the driver is automatically taken over. Thus, the driver is no longer necessary. Thus, the vehicle can also drive from an arbitrary starting position to an arbitrary target position without a driver. Potential problems are solved automatically, i.e., without the help of a driver.

[0043] Other advantages, features, and details of the present invention become clear from the following description, in which embodiments of the present invention are described in detail with reference to the accompanying drawings. In this regard, each feature mentioned in the technical solution and the description can be of substantial significance to the present invention individually or in any combination. Brief Description of the Drawings

[0044] Figure 1 Exemplarily shows an occupancy grid for the vehicle's surroundings.

[0045] Figure 2 Schematically shows the training process of a machine learning model that can be used within the scope of the present invention.

[0046] Figure 3 Shows a flowchart of an embodiment of the method of the present invention for planning a trajectory for at least a partially automated vehicle.

[0047] Figure 4 Schematically shows an embodiment of the system of the present invention for planning a trajectory for at least a partially automated vehicle. Detailed Description of the Embodiments

[0048] In the following description of the embodiments of the present invention, the same elements are labeled with the same reference numerals, and where necessary, the repeated description of these elements is omitted. These figures only schematically show the subject matter of the present invention.

[0049] Figure 1An exemplary display showing the occupancy grid 100 is presented. Each pixel in the display corresponds to a cell (ij) of the occupancy grid 100. The vehicle 120 is located at the center of the occupancy grid 100 (so-called Birds-Eye-View). The occupancy grid 100 or the occupancy information contained therein can be generated by measurement or by prediction with the aid of a machine learning model. The area 110 of the occupancy grid 100 includes cells that are considered occupied, which are generated, for example, by lidar reflections received by the lidar sensor of the vehicle 120. The area 114 is an unknown area or an area that cannot be seen by the lidar sensor. In this example, the occupancy grid 100 is attached with map information 116, here the road orientation. Within the scope of the present invention, the map information can be additionally introduced into the assessment and determination of the reliability of the occupancy information, for example as a boundary condition or as a prior probability in subjective logic opinions.

[0050] Figure 2 Illustrate the machine learning algorithms that can be preferably used in the present invention.

[0051] The algorithm includes a pre-trained (offline) prediction network ML that performs occupancy prediction in an autonomous vehicle during runtime (online). And in addition, it also performs the determination of subjective logic opinions. In both cases, i.e., during training and during runtime, the input data of the prediction network ML includes the history G(t0-1), G(t0-2),..., G(t0-H) of length H of the occupancy data of the occupancy grid. The output is respectively the predicted occupancy (G(t0)) at the current time point t0 of the occupancy grid.

[0052] In both cases, compensation for the vehicle's own motion is first performed in order to establish an association with the occupancy grid in terms of location. One possibility for compensation lies in transforming the occupancy data (G(t0-1), G(t0-2) …… G(t0-H)) of the occupancy grid to the time point t0 based on the vehicle's motion. Thereby, the prediction is decoupled from the own motion. The history of the occupancy data of the occupancy grid can for example be stored in a buffer and handed over to a suitable network. For this purpose, there are various architectures and forms with different advantages and disadvantages. These architectures and forms are described in the background art. The input data can for example be a 3D tensor consisting of occupancy grids of multiple time steps. Here, the number of input channels is equal to the number of time steps multiplied by the number of channels per occupancy grid. The encoder-decoder architecture ML leads to a grid at the network output that has the same dimensions as the occupancy grid at the input, but only for the time point t0. An alternative to this implementation is to apply an architecture such as that described in "Wu, P., Chen, S., ,, MotionNet: Joint Perception and Motion Prediction for Autonomous Driving Based on Bird’s Eye View Maps", IEEE Conference on Computer Vision and Pattern Recognition (CVPR), DOI: 10.1109 / CVPR42600.2020.01140, June 2020, pp. 11382-11392", in which the spatial dimension and the time dimension are processed separately. Another possibility is a recurrent network, such as for example in "M. Schreiber, V. Belagiannis, C. and K. Dietmayer, "Dynamic Occupancy Grid Mapping with Recurrent Neural Networks", 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China, 2021, pp. 6717-6724, doi: 10.1109 / ICRA48506.2021.9561375". In the simplest case, the output of the network can be the predicted occupancy grid G(t0) for the time point t0.

[0053] Subsequently, the predicted occupancy grid G(t0) from the past is compared with the occupancy grid B(t0) aggregated from sensor data during runtime. For this purpose, the ego-motion of the vehicle is also compensated first. This comparison can provide different outputs: In the simplest case, a cell-by-cell comparison is performed and the predicted difference is calculated for each cell (ij). This simple output already allows for a targeted identification of areas in the occupancy grid that are particularly poorly predicted. Finally, based on this cell-by-cell criterion, a statement about "how good the prediction is" can be derived.

[0054] Figure 3 An exemplary flow of a possible implementation of the inventive method for planning a trajectory for at least a partially automated vehicle is described.

[0055] In step 310, a pre-trained machine learning model is provided for determining the occupancy G(t0) of the occupancy grid. The machine learning model is trained to predict the occupancy G(t0) of the occupancy grid at a subsequent time point t0 based on an occupancy history G(t0-1), G(t0-2),..., G(t0-H) of a determined length H. A history can for example have a length of 3 to 5 seconds. In the case of a common image rate of e.g. 10 Hz, the resulting history length is 30 to 50 images or occupancies of the occupancy grid.

[0056] In step 320, a temporal occupancy sequence B(t0-H),..., B(t0-2), B(t0-1), B(t0) of the occupancy grid is generated with the aid of measurement data from the vehicle environment. Here, H also corresponds to the history length H. Occupancy can for example be generated with the aid of lidar data from the vehicle's lidar sensor, where advantageously a transformation of the occupancy data is performed based on the ego-motion of the vehicle.

[0057] In step 330, the occupancy sequence (B(t0-H),..., B(t0-2), B(t0-1), B(t0)) determined in step 320 is analyzed and evaluated with the aid of the pre-trained machine learning model in order to predict the overall occupancy G(t0) of the occupancy grid at the current time point (t0).

[0058] In step 340, the occupancy G(t0) of the predicted occupancy grid at the current time point (t0) is compared with the occupancy B(t0) determined with the aid of measurement data. Here, the deviation between the measurement and the prediction is identified, and a location-dependent measure for the reliability of the occupancy or for the reliability of the known information is determined on the basis of this comparison. Thus, for example, occupancy information can be considered to have low reliability in those areas where the prediction deviates significantly from the measurement. Here, in particular, values for the degree of agreement, deviation and uncertainty can be determined for the comparison of the measurement and the prediction by determining a subjective logical opinion for each cell, and a reliability measure can be derived from these values.

[0059] In step 350, a trajectory is planned for the at least partially automated vehicle on the basis of the reliability, in particular on the basis of the identified deviation and / or uncertainty measure. Here, for example, those areas of the occupancy grid with a high deviation and / or high uncertainty can be avoided.

[0060] Figure 4 A system 400 for planning a trajectory for an at least partially automated vehicle is shown. The system 400 includes a machine learning module 418, a perception module 414, an interface 412 for receiving data from environmental sensing devices, a comparison module 420, and a planning unit 430.

[0061] The machine learning model 418 is trained to predict the occupancy G(t0) of the occupancy grid at the next time point t0 on the basis of the occupancy history G(t0−1), G(t0−2),..., G(t0−H) of determined length H. The perception module 414 is configured to determine the occupancy (B(t)) of the occupancy grid with the aid of measurement data received or read from the interface 412 (cf. Figure 1 ).

[0062] The comparison module 420 is arranged to compare the occupancy 422 of the occupancy grid predicted at the current time point t0 by the machine learning module 418 with the occupancy 424 determined with the aid of measurement data, wherein, in particular, the deviation between the measurement and the prediction is identified, and a location-dependent measure for the reliability of the occupancy information is determined on the basis of this comparison.

[0063] The planning module 430 is configured to plan a trajectory for the at least partially automated vehicle on the basis of the reliability, in particular on the basis of the identified deviation and / or uncertainty measure, and to output 440 it to the vehicle.

Claims

1. A method for planning a trajectory for an at least partially automated vehicle (120), the method comprising the following steps: a) providing a pre-trained machine learning model for determining the occupancy (G(t)) of an occupancy grid (100), wherein: The machine learning model is trained to predict the occupancy (G(t0)) of the occupancy grid (100) at a subsequent time point (t0) based on the occupancy history G(t0-1), G(t0-2) ... G(t0-H) of a determined length H; b) generating a temporal occupancy sequence B(t0-H) ... B(t0-2), B(t0-1), B(t0) of the occupancy grid (100) using measurement data from the environment of the vehicle (120); c) evaluating the occupancy sequence (B(t0-H) ... B(t0-2), B(t0-1)) determined according to step b) with the aid of the pre-trained machine learning model in order to predict the entire occupancy G(t0) of the occupancy grid (100) at the current point in time (t0); d) Among them, comparing a predicted occupancy (G(t0)) of the occupancy grid (100) at a current point in time (t0) with an occupancy (B(t0)) determined with the aid of measurement data, wherein in particular deviations between the measurement and the prediction are detected and a location-dependent measure for the reliability of the occupancy information is determined based on the comparison; e) planning a trajectory for the at least partially automated vehicle (120) based on the reliability, in particular based on the identified deviations and / or uncertainty measures.

2. The method according to claim 1, wherein: For a determined currently occupied spatial region, a measure for the reliability is determined based on a deviation between the measurement (B(t0)) and the prediction (G(t0)), in particular based on the deviation and / or measurement uncertainty, wherein a region comprises at least one cell (ij) and / or a plurality of consecutive cells (ij) of the occupancy grid (100).

3. The method according to claim 1 or 2, wherein: In step d), in order to determine the reliability, a subjective logical opinion is determined for each cell (ij) of the occupancy grid, wherein in particular a tuple (b ij , d ij ,u ij , a ij ), where b ij represents the agreement between the measured and predicted occupancy of the cell (ij), d ij represents the deviation between the measured and predicted occupancy of the cell (ij), u ij represents the uncertainty of the occupancy of the cell (ij), and a ij Describes the base probability of occupancy of this cell (ij).

4. The method according to claim 3, wherein: In step e), a trajectory is planned for the vehicle in such a way that the vehicle avoids the occupancy grid (100) with a high deviation (d ij ) and / or high uncertainty (u ij ) of cells and / or cell regions, and / or preferably having a high degree of consistency (b ij ) area.

5. The method according to any one of the preceding claims, wherein: Before processing by the machine learning model and / or before the comparison, the occupancy grid is transformed in such a way that compensation for the vehicle's (120) own movement is performed.

6. The method according to any one of the preceding claims, wherein: Input data of the machine learning (ML) model includes 3D tensors, wherein the 3D tensors respectively include occupancy grids of the determined surrounding environment at different time points.

7. The method according to claim 6, wherein: The machine learning model has an encoder-decoder architecture, wherein the output data of the machine learning model includes an occupancy grid having the same dimensions as the occupancy grid of the input data, but only for a certain time point (t0).

8. The method according to any one of the preceding claims, wherein: The machine learning model has an architecture in which the spatial and temporal dimensions of the input data are processed separately.

9. The method according to any one of the preceding claims, wherein: The machine learning model includes a recurrent network.

10. A computer program having instructions which, when executed by a processor, cause the processor to perform the method according to one of claims 1 to 9.

11. A system (400) for planning a trajectory for an at least partially automated vehicle (120), the system comprising a machine learning module (418), a perception module (414), an interface (412) for receiving data from an environmental sensor device, a comparison module (420), and a planning unit (430), wherein: The system (400) is designed to carry out the method according to one of claims 1 to 9, wherein: The machine learning model (418) is trained to predict the occupancy (G(t0)) of the occupancy grid (100) at a subsequent time point (t0) based on the occupancy history (G(t0-1), G(t0-2) ... G(t0-H)) of a certain length H, The perception module (414) is configured to determine the occupancy (B(t)) of the occupancy grid by means of environmental sensor device measurement data received via the interface (412), The comparison module (420) is configured to compare the occupancy (422) of the occupancy grid at the current point in time (t0) predicted by means of the machine learning module (418) with the occupancy (424) determined by means of measurement data, wherein in particular deviations between the measurement and the prediction are identified and a location-related measure for the reliability of the occupancy information (G(t0), B(t0)) is determined based on the comparison; The planning module (430) is designed to plan a trajectory (440) for the at least partially automated vehicle based on the reliability, in particular based on the detected deviations and / or uncertainty measures.

12. A vehicle (120) arranged to be driven at least partially automatically, the vehicle comprising a system according to claim 11.

Images