Method and system for providing a context-dependent knowledge base to validate at least one perceptual function

Abstract

Methods for providing a context-dependent knowledge base (20) to validate at least one perceptual function (30), wherein participants (40) of a vehicle fleet (51) each locally execute at least one perception function (30) and generate at least one context-dependent knowledge datum (10) based on at least one output of the at least one perception function (30), wherein the generated context-dependent knowledge data (10) are transmitted by the participants (40) of the vehicle fleet (51) to a backend server (3), wherein the transmitted context-dependent knowledge data (10) are collected on the backend server (3) and compared with each other in terms of content, whereby a context-dependent knowledge base (20) is created or updated on the basis of a comparison result, wherein the generated or updated context-dependent knowledge base (20) is transmitted to the participants (40) of the vehicle fleet (51), and wherein the transmitted context-dependent knowledge base (20) is used locally by the participants (40) to validate and / or support the at least one perceptual function (30), whereby a result of the at least one perceptual function (30) is validated on the basis of the context-dependent knowledge base (20).

Description

[0001] The invention relates to a method and a system for providing a context-dependent knowledge base for verifying the plausibility of at least one perception function.

[0002] It is known to create a knowledge base based on manual knowledge modeling from existing knowledge in the form of graphs, for example, and to form hypotheses about an environment or objects on this basis. This can be used, for example, in classification and object recognition. However, this is problematic because all relevant semantic relationships must be modeled manually. Furthermore, it is known to generate knowledge in the form of class-specific occupancy maps (also known as heat maps) from data sets and to use these to weight a decision, for example in semantic segmentation. However, the creation and use of occupancy maps is data set-specific, so the occupancy maps can differ significantly depending on the context.It is also known to use machine learning methods to extract patterns from a dataset and then use the extracted patterns to improve the performance of a function, e.g., when extracting shapes or functions in two-dimensional or three-dimensional object mapping. Here, too, the extracted patterns are highly dependent on the dataset used during training.

[0003] EP 3 511 920 A1 discloses a method for operating a parking assistance system for at least one motor vehicle, comprising detecting the surroundings of the motor vehicle in a parking facility for motor vehicles by an environment detection device of the one motor vehicle; transmitting environmental data generated by the environment detection device, which represent the detected environment, and position data associated with the respective environmental data, which represent at least one position in the environment driven through or parked by the one motor vehicle, to a computing device external to the vehicle; training a neural network for detecting driveable or parkable areas in the detected environment based on the environmental data, using a supervised learning process based on the environmental data and the associated position data, in the computing device external to the vehicle.a return transmission of profile information characterizing the trained neural network from the vehicle-external computing device to one motor vehicle or to another motor vehicle; and a configuration of a local neural network of a parking assistance device of the respective motor vehicle, which serves to recognize drivable or parkable areas in an environment detected by an environment detection device of the respective motor vehicle, with the returned profile information in order to reduce the infrastructure required by the parking assistance device of the motor vehicle during parking of the motor vehicle assisted by the parking assistance device.

[0004] DE 10 2017 223 632 A1 discloses a system and method for calculating an error probability of a vehicle-based sensor data set, comprising a sensor unit and sensors, a reference database containing reference data, and a central computer. The reference data has a temporal and spatial relationship to the sensor data set. The central computer calculates an error probability of the sensor data set based on suitable reference data from the reference database.

[0005] The invention is based on the object of improving a method and a system for providing a context-dependent knowledge base for verifying the plausibility of at least one perception function.

[0006] The object is achieved according to the invention by a method having the features of patent claim 1 and a system having the features of patent claim 10. Advantageous embodiments of the invention emerge from the subclaims.

[0007] In particular, a method is provided for providing a context-dependent knowledge base for verifying the plausibility of at least one perception function, wherein participants of a vehicle fleet each locally execute at least one perception function and generate at least one context-dependent knowledge datum based on at least one output of the at least one perception function, wherein the generated context-dependent knowledge data are transmitted from the participants of the vehicle fleet to a backend server, wherein the transmitted context-dependent knowledge data are collected on the backend server and compared with each other in terms of content, wherein a context-dependent knowledge base is generated or updated based on a comparison result, wherein the generated or updated context-dependent knowledge base is transmitted to the participants of the vehicle fleet,and wherein the transmitted context-dependent knowledge base is used locally by the participants to verify and / or support the at least one perception function.

[0008] Furthermore, in particular, a system for providing a context-dependent knowledge base for verifying the plausibility of at least one perception function is created, comprising control devices of participants in a vehicle fleet, and a backend server, wherein the control devices are each designed to generate at least one context-dependent piece of knowledge data on the basis of at least one output of at least one perception function executed locally by the participant and to transmit the generated at least one context-dependent piece of knowledge data to the backend server, wherein the backend server is designed to collect the knowledge data transmitted by the participants and to compare their content and to generate or update a context-dependent knowledge base based on a comparison result, and to transmit the generated or updated context-dependent knowledge base to the participants in the vehicle fleet,and wherein the control devices are further each designed to use the transmitted context-dependent knowledge base locally to verify the plausibility and / or support the at least one perception function.

[0009] The method and the system make it possible to provide an improved context-dependent knowledge base with which the at least one perception function can be verified or validated. This is achieved by generating the knowledge for the knowledge base on a fleet-by-fleet basis. Participants in a vehicle fleet each locally execute at least one perception function and generate at least one context-dependent piece of knowledge data based on at least one output of the at least one perception function. The generated context-dependent knowledge data is transmitted by the participants of the vehicle fleet to a backend server. There, the transmitted context-dependent knowledge data is collected. The collected context-dependent knowledge data is compared with each other in terms of content in the backend server, i.e., in particular, merged.Based on a comparison result, a context-dependent knowledge base is generated if no knowledge base yet exists, or an existing context-dependent knowledge base is updated based on the currently transmitted context-dependent knowledge data and its comparison. Updating particularly includes expanding the context-dependent knowledge base. The context-dependent knowledge base particularly includes a set of merged context-dependent knowledge data. The generated or updated context-dependent knowledge base is transmitted from the backend server to the participants of the vehicle fleet. The transmitted context-dependent knowledge base is used locally by the participants to verify plausibility and / or support the at least one perception function.

[0010] One advantage of the invention is that a context-dependent knowledge base can be created and updated on a fleet-by-fleet basis. This means that the knowledge contained in the knowledge base is not limited in scope and can be continuously updated and expanded depending on the context. Updating and expanding can be automated. Furthermore, the creation of the knowledge base by the fleet ensures that only knowledge accessible to fleet members is stored in the knowledge base.

[0011] A context, or the term context-dependent, should in particular mean that the knowledge data is generated or valid for a specific context. The context can include a local (e.g., in the form of coordinates, or through an infrastructural or geographical reference), temporal (e.g., time of day, day of the week, week, season), situation-related (e.g., village, city, country, parking lot), weather-related (e.g., sunshine, rain, snow, fog, wind), and / or traffic-related (e.g., light traffic, heavy traffic, traffic jam, many different road users, traffic light phase) context. In principle, it is also possible to consider other context dimensions, in particular dimensions to which humans have no intuitive access, but which can only be determined using machine learning methods.

[0012] A perception function refers, in particular, to a function that is performed to detect and / or interpret the environment. The perception function can, for example, serve for semantic segmentation, object recognition, and / or predicting the behavior of other road users. The perception function can, for example, be implemented based on a trained deep neural network and / or a Bayesian network. In principle, however, a perception function can also be implemented differently.

[0013] The at least one perceptual function used to generate the at least one piece of knowledge data can, in principle, be a different perceptual function than the at least one perceptual function verified or validated using the context-dependent knowledge base. Ideally, however, the perceptual functions are the same.

[0014] A context-dependent piece of knowledge specifically refers to a context-dependent relationship, i.e., a relationship between objects and / or environmental properties in a specific context. The context-dependent piece of knowledge or context-dependent knowledge data can be generated, in particular, in the form of occupancy maps, object occurrences, object relationships, object properties, and / or object behaviors. Simply put, the context-dependent piece of knowledge or context-dependent knowledge data should provide or enable a statement about which environment, with which objects and their object properties, etc., can typically be expected in which context.

[0015] For example, an occupancy map indicates how likely it is that certain objects (e.g., pedestrians or traffic signs) are present in certain areas of the environment. For example, pedestrians are more likely to be found on the roadway near a street crossing or zebra crossing than on the roadway of a highway.

[0016] Object occurrences generally indicate which objects, or which types or classes of objects, can occur in a context. In the previous example, for example, in the context of a highway, the object "pedestrian" would not occur or would be associated with an extremely low probability. The object "motor vehicle," on the other hand, would be considered extremely likely.

[0017] Object relationships indicate relationships between different or similar objects, or relationships within or between object classes, particularly the co-occurrence of objects in a particular context. Examples include the following statements: a person has legs, people can be found at a bus stop, or people can be found at a zebra crossing.

[0018] Object properties refer, for example, to typically occurring colors, shapes, reflection properties or relations between classes (person on a moving bicycle, trucks are larger than pedestrians, children move faster than trees) of objects.

[0019] Behaviors refer to the typical behavior of objects in a specific context. For example, a pedestrian at a zebra crossing can be expected with a high probability that the pedestrian intends to cross the street or will cross. Another example is the assumption of a continuous / changing speed of a motor vehicle or a pedestrian, etc. Another motor vehicle in a right-turn lane, for example, is highly likely to turn right, etc.

[0020] A participant in a vehicle fleet is, in particular, a motor vehicle. However, a participant can, in principle, be any other land, water, air, or space vehicle that forms a vehicle fleet with other participants.

[0021] The participants' control devices and the backend server communicate primarily via dedicated communication interfaces and a communication connection. This is typically a wireless communication connection (WLAN, 4G, 5G, etc.). The context-dependent knowledge data and the context-dependent knowledge base are transmitted, for example, in the form of data packets.

[0022] The transmission of at least one context-dependent knowledge data from the participants to the backend server takes place in particular at temporal and / or spatial intervals and / or depending on progress in generating the context-dependent knowledge data, in particular depending on a number and / or a data volume of the generated context-dependent knowledge data.

[0023] The comparison of the context-dependent knowledge data transmitted to the backend server is carried out in particular by consolidating the context-dependent knowledge data. This can be done, in particular, by offsetting the individual pieces of knowledge data against each other, for example, by averaging and / or mapping the relationships using statistical distributions and / or expected values, etc. In particular, the comparison or consolidation of the knowledge data can include correlation analyses, on the basis of which the knowledge data is generalized. The correlation analysis answers the question of which constellations of objects and / or environmental properties, etc., occur simultaneously (positive correlation) and which do not occur simultaneously (negative correlation). If the corresponding correlations are identified in the knowledge data of several participants, the underlying relationships are generalized.Simply put, after performing the comparison, the comparison result should be averaged or generalized relationships between objects and / or environmental properties, etc., for the contexts under consideration across the participants in the vehicle fleet. The context-dependent knowledge base is then generated from the comparison result, i.e., the averaged relationships, or the context-dependent knowledge base is updated. During the update, it can be specifically planned that an existing context-dependent knowledge base is taken into account during the comparison.

[0024] The generated or updated context-dependent knowledge base is used by the participants in the vehicle fleet to check the plausibility of and / or support the at least one perception function. This means that a result of the at least one perception function is checked for plausibility on the basis of the context-dependent knowledge base. On the other hand, parameters of the perception function can also be selected and / or set on the basis of the context-dependent knowledge base before the perception function is applied, so that the at least one perception function is supported in a current context. If the perception function is, for example, a function for detecting pedestrians in camera images captured of the environment, support can be provided by identifying an area in which pedestrians are usually found on the basis of the context-dependent knowledge base.The perception function can then evaluate this area with particularly high resolution. Context-dependent changes in pedestrian properties can also be searched for in order to identify pedestrians in the surrounding area. For example, if the current context includes rainy weather, it can be extracted from the context-dependent knowledge base that pedestrians are, depending on the context, about half a meter taller due to open umbrellas than in dry weather. For example, targeted search tasks for pedestrians in certain areas of the surrounding area can also be specified based on the context-dependent knowledge base (e.g., "find pedestrian / object y in area x"). The perception function is then adapted accordingly, or a result of the perception function is checked for plausibility.

[0025] In particular, it is provided that a plausibility value is calculated for a result of the at least one perception function based on the context-dependent knowledge base. This plausibility value can then be used to evaluate the result of the at least one perception function. Based on the plausibility value, it can be decided, for example, whether a result of the at least one perception function can be trusted or that the result should be rejected. This can increase safety, especially when the result is used to implement an automated driving function.

[0026] Parts of the control devices can be implemented individually or collectively as a combination of hardware and software, for example, as program code executed on a microcontroller or microprocessor. However, it can also be provided that parts are implemented individually or collectively as an application-specific integrated circuit (ASIC).

[0027] In one embodiment, the relevance of the context-dependent knowledge data transmitted to the backend server is assessed by the backend server. This can improve the matching process, as, for example, only knowledge data that achieves a certain rating is considered. Within the scope of the assessment, for example, legal or conventional compliance can be assessed. Furthermore, an ethical assessment can also be performed.

[0028] The basis for such an assessment can be in particular: - Regulatory requirements (minimum quality to be achieved is prescribed and any knowledge that leads to an expected increase in functional quality is considered relevant); - Latency requirements (one allows a maximum knowledge representation that allows plausibility check in every context at a given maximum latency); - Usability requirements (only knowledge data that occurs in a minimal number of plausibility checks is used); - generality requirements (only knowledge that is relevant to a minimal number of different situations is considered relevant); - Ethical requirements (certain knowledge, although potentially factually correct, can introduce negative bias into the behavior of systems - such knowledge should not have any plausibility influence).

[0029] In one embodiment, the backend server evaluates the accuracy of the context-dependent knowledge data transmitted to the backend server. The evaluation of accuracy should, in particular, include checking the logical consistency of the context-dependent knowledge data. Logical consistency can be checked, for example, based on physical laws and / or traffic regulations and / or other behavioral conventions and social rule systems.

[0030] The basis for such an assessment can be in particular: - World knowledge: for certain knowledge data, an assessment based on existing rule-based knowledge can be used; this applies in particular to physical, geometric, or traffic rule-related knowledge; - Expert assessments: certain interpretable knowledge data may not be covered by existing rule-based knowledge, but can certainly be evaluated by experts; such experts can be: lawyers, government authorities, scientists, technicians (engineers, computer vision experts, etc.), ethics committees, etc.; - Evaluations based on a large base of training data or fleet experience; - Logic (applications of logical operations such as connections, logical reasoning, generalization or specialization).

[0031] In one embodiment, it is provided that the at least one piece of context-dependent knowledge data generated locally by the participants is used locally to verify the plausibility of and / or support at least one perception function. This allows context-dependent knowledge data collected locally by the individual participants to be used to verify the plausibility of and / or support the at least one perception function. This improves the performance of the at least one perception function even without the context-dependent knowledge base being compared or consolidated by the backend server.

[0032] In particular, it can be provided that the participants of the vehicle fleet locally create or update a context-dependent knowledge base on the basis of the at least one generated knowledge data and use this to check and / or support the at least one perception function.

[0033] In one embodiment, it is provided that context-dependent knowledge data is annotated by the participants in the vehicle fleet depending on the degree to which it has contributed to an increase in the performance of the at least one perception function and / or to the plausibility of the at least one perception function, wherein at least annotations that exceed a predetermined threshold are transmitted to the backend server. Annotating should in particular mean that a context-dependent piece of knowledge data is linked to a value that is representative of the degree to which it has contributed to an increase in the performance of the at least one perception function and / or to the plausibility of the at least one perception function. This makes it possible to evaluate the context-dependent knowledge data stored in the context-dependent knowledge base according to its relevance ora level of support for the at least one perceptual function. The context-based knowledge base can thus be improved step by step, since irrelevant context-dependent knowledge data that enable no or only a slight increase in the performance of the at least one perceptual function can be removed in favor of context-dependent knowledge data that enable an improvement in performance. In the simplest case, it is only annotated whether a context-dependent piece of knowledge data has contributed to an improvement in the performance of the at least one perceptual function and / or to the plausibility of the at least one perceptual function (positive annotation) or not (negative annotation). However, it can also be provided that further intermediate stages are provided for the evaluation.

[0034] It can also be provided that only context-dependent knowledge data for which an associated annotation exceeds the specified threshold is transmitted to the backend server. This allows a pre-selection to be made, thus reducing the number of context-dependent knowledge data items to be transmitted and compared. This can save storage space and data volume during communication between the participant and the backend server.

[0035] In a further embodiment, the context-dependent knowledge base is updated taking into account the transmitted annotations. In particular, context-dependent knowledge data that enable a high degree of support are given priority.

[0036] In one embodiment, a probability-based evaluation of the context-dependent knowledge data is performed to update the context-dependent knowledge base. In particular, this can include all known methods of probabilistic, i.e., probability-based, reasoning.

[0037] It can also be provided that in order to update the context-dependent knowledge base, the context-dependent knowledge data are evaluated or linked together using fuzzy logic.

[0038] In one embodiment, a contextual refinement of the context-dependent knowledge base is performed to update the context-dependent knowledge base. This allows the context-dependent knowledge data to be provided even more specifically for a specific, more finely defined, context. The contextual refinement includes, in particular, the dimensions of location, time, situation, weather, and / or traffic. In principle, it is also possible to consider other context dimensions, in particular dimensions to which humans do not have intuitive access but which can only be determined using machine learning methods.

[0039] In one embodiment, only a selected portion of the participants in the vehicle fleet generates context-dependent knowledge data and transmits it to the backend server. This allows these selected participants to be equipped, for example, with improved (and therefore generally more expensive) environmental sensors and increased data storage capacity. In other words, a distinction is then made between an extraction fleet, whose participants generate the context-dependent knowledge data, and an overall fleet, which includes all other participants in the vehicle fleet in addition to the extraction fleet, whereby the other participants only receive the context-dependent knowledge base from the backend server but do not generate or provide any context-dependent knowledge data themselves. Overall, this can save resources and costs.

[0040] It can be provided that the backend server transmits at least one specification to the participants of the vehicle fleet regarding the context for which context-dependent knowledge data should be generated, whereby the participants generate the at least one piece of context-dependent knowledge data based on the at least one transmitted specification. For example, context-dependent knowledge data can only be generated in contexts that correspond to the specification of the backend server. This allows context-dependent knowledge data to be specifically generated and collected for certain contexts for which, for example, little or no knowledge is available in the context-dependent knowledge base.

[0041] Further features of the system's design are described in the various process configurations. The advantages of the system are the same as those of the process configurations.

[0042] The invention will be explained in more detail below using preferred embodiments with reference to the figures. Fig. 1 a schematic representation of an embodiment of the system for providing a context-dependent knowledge base for verifying the plausibility of at least one perception function; Fig. 2 a schematic flow diagram of an embodiment of the method for illustrating the method steps in creating or updating the context-dependent knowledge base; Fig. 3 a schematic flow diagram of an embodiment of the method to illustrate the method steps on the part of the participants.

[0043] In Fig. Figure 1 shows a schematic representation of an embodiment of the system 1 for providing a context-dependent knowledge base 20 for verifying the plausibility of at least one perception function. The system 1 comprises control devices 2 in motor vehicles 50, which form participants 40 of a vehicle fleet 51, and a backend server 3.

[0044] The control devices 2 are designed, for example, as a combination of hardware and software, for example as program code that is executed on a microcontroller or microprocessor.

[0045] The backend server 2 comprises a computing device 4 which can access a storage device 5.

[0046] At least one perception function, for example, for object detection in the respective local environment of the motor vehicle 50, is executed in each of the motor vehicles 50. The perception function can be executed, for example, on the control devices 2. The control devices 2 receive sensor data acquired by an environmental sensor system (not shown), for example, a camera, on which the perception function is then executed. The perception function can, for example, detect pedestrians in the environment of the motor vehicle and determine their position.

[0047] The control devices 2 generate at least one context-dependent piece of knowledge data 10 in a context-dependent manner based on at least one output of the at least one locally executed perception function. The generated at least one context-dependent piece of knowledge data 10 is then transmitted by the control devices 2 to the backend server 2. This occurs in particular at specific temporal and / or spatial intervals or depending on specific criteria, such as the number of pieces of knowledge data collected or the distance traveled since the last transmission, etc.

[0048] The backend server 2 receives and collects the knowledge data 10 transmitted by the control devices 2 and compares the content of the knowledge data 10 transmitted by the participants 40. Based on the comparison result, the backend server 2 generates or updates a context-dependent knowledge base 20. The backend server 2 then transmits the generated or updated context-dependent knowledge base 20 back to the participants 40 of the vehicle fleet 51.

[0049] The control devices 2 of the participants 40 or the motor vehicles 50 receive the transmitted context-dependent knowledge base 20 and use it locally to verify the plausibility and / or support the at least one perception function.

[0050] Subsequently, the described steps are repeated, i.e. the participants 40 continue to collect context-dependent knowledge data 20 and transmit it to the backend server 2. There, the context-dependent knowledge base 20 is updated by comparison and the updated knowledge base 20 is again transmitted to the control devices 2 of the participants 40.

[0051] It can be provided that the relevance of the context-dependent knowledge data 10 transmitted to the backend server 2 is evaluated by the backend server 2. The comparison is then carried out taking into account the relevance of the context-dependent knowledge data 10.

[0052] It can be provided that the accuracy of the context-dependent knowledge data 10 transmitted to the backend server is evaluated by the backend server 2. The comparison is then carried out taking into account the accuracy of the context-dependent knowledge data 10.

[0053] It can be provided that the at least one context-dependent piece of knowledge 10 locally generated by the participants 40 is used locally to verify the plausibility of and / or support at least one perception function. This is done in particular by means of the control devices 2, but can in principle also be done by means of other devices of the motor vehicle 50.

[0054] It can be provided that context-dependent knowledge data 10 are annotated by the participants 40 of the vehicle fleet 51 depending on the degree to which they have contributed to an improvement in the performance of the at least one perception function and / or to the plausibility of the at least one perception function, wherein at least annotations 11 that exceed a predetermined threshold value are transmitted to the backend server 2.

[0055] Further development can be provided for the backend server 2 to update the context-dependent knowledge base 10 taking into account the transmitted annotations 11.

[0056] It can be provided that the backend server 2 performs a probability-based evaluation of the context-dependent knowledge data 10 in order to update the context-dependent knowledge base 20.

[0057] It can be provided that the backend server 2 performs a contextual refinement of the context-dependent knowledge base 20 in order to update the context-dependent knowledge base 20.

[0058] It can be provided that only a selected portion of the participants 40 of the vehicle fleet 51 generates context-dependent knowledge data 10 and transmits it to the backend server 2. The motor vehicles 50 of the selected portion have, in particular, improved environmental sensors and a larger memory.

[0059] In Fig. Figure 2 shows a schematic flow diagram of an embodiment of the method for illustrating the method steps when creating or updating the context-dependent knowledge base 10. Some of the method steps shown are executed by the participants 40 of the vehicle fleet 51, while others are executed by the backend server 2. After starting 100 of the method, the perception functions and the local knowledge bases 12 of the participants 40 (or in the motor vehicles 50) of the vehicle fleet 51 are initialized in a method step 101.

[0060] In a method step 101, the participants 40 of the vehicle fleet 51 each locally execute the at least one perception function and generate at least one context-dependent piece of knowledge datum 10 based on at least one output of the at least one perception function. The context-dependent piece of knowledge datum 10 or context-dependent knowledge data 10 can be generated, in particular, in the form of occupancy maps 13, object occurrences 14, object relationships 15, object properties 16, and / or behaviors 17 of the objects.

[0061] In a method step 103, the generated context-dependent knowledge data 10 is transmitted from the participants 40 of the vehicle fleet 51 to the backend server 2. This occurs via a communication connection. The transmission can occur at predetermined temporal and / or spatial intervals and / or depending on the number of generated context-dependent knowledge data 10.

[0062] In a method step 104, the backend server 2 collects the transmitted context-dependent knowledge data 10 of the participants 40 of the vehicle fleet 51.

[0063] In a method step 105, the backend server 2 performs an evaluation of the relevance and accuracy of the collected context-dependent knowledge data 10.

[0064] If a context-dependent piece of knowledge 10 is not relevant and not correct, it is discarded in a process step 106. If, however, the context-dependent piece of knowledge 10 is assessed as relevant and correct, the context-dependent knowledge base 10 is updated in a process step 107.

[0065] The updated context-dependent knowledge base 10 is then transmitted to the participants 40 of the vehicle fleet 51. The participants 40 can then, for example, each replace the local knowledge base 12 with the updated context-dependent knowledge base 10 transmitted by the backend server 2.

[0066] The transmitted updated context-dependent knowledge base 10 is then used locally by the participants 40 to verify the plausibility and / or support the at least one perception function.

[0067] In Fig. 3 shows a schematic flow diagram of an embodiment of the method to illustrate the method steps on the part of the participants.

[0068] After process step 101 (cf. Fig.2) In a method step 200, it is checked whether the method is being executed for the first time. If this is the case, a local knowledge base 12 and the at least one perception function 30 are initialized in a method step 201.

[0069] In a method step 202, the at least one perception function 30 is executed, i.e., a result is inferred. For example, object recognition can be performed in camera data captured from the environment using a trained neural network.

[0070] In a method step 203, it is checked whether a result of the at least one perception function 30 is to be plausibly checked or whether a context-dependent knowledge datum is to be generated.

[0071] If a context-dependent knowledge datum is to be generated, it is generated or collected in a method step 204 on the basis of the inferred result and a current context.

[0072] In a method step 205, a check is made to determine whether a time interval has expired or whether a certain amount of context-dependent knowledge data has already been collected. If this is not the case, the system returns to method step 202.

[0073] If this is the case, the context-dependent knowledge data is transmitted to the backend server in a process step 206.

[0074] If, however, method step 203 indicates that a plausibility check of the inferred result of the at least one perception function should be performed, then following method step 204, a check is carried out in method step 207 to determine whether the result inferred by the at least one perception function agrees with the local knowledge base 12. Agreement should here denote the existence of plausibility between the inferred result and the context-dependent knowledge base and should not be understood as a strict identity. If this is not the case, a warning is generated in method step 208, and the system returns to a fallback state. The fallback state can, for example, include more defensive driving behavior. If, however, the inferred result and the local knowledge base 12 agree, the method continues in method step 209.In particular, a plausibility check result is positive in this case. As part of the check in method step 207, it may be provided that a plausibility check value is generated and checked against a threshold value.

[0075] Furthermore, each time the backend server transmits an updated context-dependent knowledge base 20 to the participants, the local knowledge base 12 is updated in a method step 210. List of reference symbols 1 system 2 Control device 3 backend servers 4 Computing device 5 Storage device 10 context-dependent knowledge data 11 Annotation 12 local knowledge base 13 Occupancy map 14 object occurrences 15 Object relationship 16 Object property 17 Behavior 20 context-dependent knowledge base 30 Perceptual function 40 participants 50 motor vehicles 51 vehicle fleet 100-107 Procedural steps 200-210 procedural steps

Claims

[1] Method for providing a context-dependent knowledge base (20) for verifying the plausibility of at least one perception function (30), wherein participants (40) of a vehicle fleet (51) each locally execute at least one perception function (30) and generate at least one context-dependent knowledge datum (10) based on at least one output of the at least one perception function (30), wherein the generated context-dependent knowledge data (10) are transmitted by the participants (40) of the vehicle fleet (51) to a backend server (3), wherein the transmitted context-dependent knowledge data (10) are collected on the backend server (3) and compared with each other in terms of content, wherein a context-dependent knowledge base (20) is generated or updated on the basis of a comparison result, wherein the generated or updated context-dependent knowledge base (20) is transmitted to the participants (40) of the vehicle fleet (51), and wherein the transmitted context-dependent knowledge base (20) is used locally by the participants (40) to check the plausibility and / or support the at least one perception function (30), wherein a result of the at least one perception function (30) is checked for plausibility on the basis of the context-dependent knowledge base (20). [2] Method according to claim 1, characterized by that a relevance of the context-dependent knowledge data (10) transmitted to the backend server (3) is assessed by means of the backend server (3). [3] Method according to claim 1 or 2, characterized by that the correctness of the context-dependent knowledge data (10) transmitted to the backend server (3) is assessed by means of the backend server (3). [4] Method according to one of the preceding claims, characterized bythat the at least one context-dependent knowledge datum (10) generated locally by the participants (40) is used locally to verify the plausibility and / or support at least one perception function (30). [5] Method according to one of the preceding claims, characterized by that context-dependent knowledge data (10) are annotated by the participants (40) of the vehicle fleet (51) depending on the degree to which they have contributed to an increase in the performance of the at least one perception function (30) and / or to the plausibility of the at least one perception function (30), wherein at least annotations (11) that exceed a predetermined threshold value are transmitted to the backend server (3). [6] Method according to claim 5, characterized by that the context-dependent knowledge base (20) is updated taking into account the transmitted annotations (11). [7] Method according to one of the preceding claims, characterized by that a probability-based evaluation of the context-dependent knowledge data (10) is carried out to update the context-dependent knowledge base (20). [8] Method according to one of the preceding claims, characterized by that in order to update the context-dependent knowledge base (20), a contextual refinement of the context-dependent knowledge base (20) is carried out. [9] Method according to one of the preceding claims, characterized by that only a selected part of the participants (40) of the vehicle fleet (51) generates context-dependent knowledge data and transmits it to the backend server (3). [10] System (1) for providing a context-dependent knowledge base (20) for verifying the plausibility of at least one perception function (30), comprising: Control devices (2) of participants (40) of a vehicle fleet (51), and a backend server (3), wherein the control devices (2) are each designed to generate at least one context-dependent piece of knowledge data (10) on the basis of at least one output of at least one perception function (30) executed locally by the subscriber (40) and to transmit the generated at least one context-dependent piece of knowledge data (10) to the backend server (3), wherein the backend server (3) is designed to collect the knowledge data (10) transmitted by the participants (40) and to compare their content with each other and to generate or update a context-dependent knowledge base (20) on the basis of a comparison result, and to transmit the generated or updated context-dependent knowledge base (20) to the participants (40) of the vehicle fleet (51), and wherein the control devices (2) are further each designed to use the transmitted context-dependent knowledge base (20) locally to check the plausibility and / or support the at least one perception function (30), and to check the plausibility of a result of the at least one perception function (30) on the basis of the context-dependent knowledge base (20).

Images