Method for the relational storage of object properties of an object, for training artificial intelligence and for generating a class relating to an object, computer program product, storage medium, data carrier signal, devices and vehicle

Abstract

A computer-implemented method for relationally storing object properties 116 of an object 102, 104 comprises a step of obtaining 118 sensor data 172, 174, 176, 178 in which the object 102, 104 is detected, wherein the sensor data 172, 174, 176, 178 are generated by at least one sensor 106, 168, 170 arranged on a vehicle 100. Furthermore, the method 116 comprises a step of detecting and classifying 120 the object 102, 104 using the sensor data 172, 174, 176, 178, wherein the object 102, 104 is assigned a first class and at least one object property is determined, wherein a confidence value 191 is generated. Furthermore, the method 116 comprises a step of requesting 124 an occupant of the vehicle 100 to provide a second class relating to the object 102,104 if the confidence value 191 is below a first threshold value 192.Furthermore, the method 116 comprises a further step of receiving 126 a first feedback from the occupant. Furthermore, the method 116 comprises a further step of extracting 128 the second class from the first feedback from the occupant. Furthermore, the method 116 comprises a further step of storing 130 a data set 193 comprising at least the second class, the at least one object property, and a derivative of the sensor data 172, 174, 176, 178.

Description

Technical FieldThe present invention relates to the field of data processing. More particularly, the present invention relates to computer-implemented methods for relational storage of object properties of an object. The present invention further relates to a computer-implemented method for training an artificial intelligence, to computer-implemented methods for generating a class relating to an object, to a computer program product, to a computer-readable storage medium, to a data carrier signal, to devices for classifying an object, and to a vehicle.Technical Background and ObjectDetection of objects outside a vehicle is used to enable automated driving. For this purpose, the objects must also be classified in order to calculate corresponding automated driving maneuvers. However, the technique is also used to transmit information to occupants of a vehicle. For example, systems are known which detect and classify vehicles located in front of the own vehicle and which are then displayed on display by a digital graphical model based on the class of the vehicle. A truck is then also shown as such and not as a generic vehicle.The detection and classification are usually carried out by means of artificial intelligences, such as neural networks (NN). The accuracy with which an NN can detect and classify an object depends on the training dataset. If a specific object is missing in this object, it can be detected and classified reliably only with a low probability. If an occupant of a vehicle, for example the driver, is presented with a misinterpreted vehicle on a display in the vehicle, this can obscure the driver and, in the worst case, provoke an accident.DE 10 2018 219 125 A1 relates to a method for classifying objects by means of an automatically driving motor vehicle, wherein the automatically driving motor vehicle has a device for detecting and classifying objects and output and input means, wherein the detected but not classifiable or not unambiguously classifiable object is made known to at least one passenger of the automatically driving motor vehicle via the output means, wherein the passenger is requested to classify the detected object, wherein the classifications of the at least one passenger are stored, and an automatically driving motor vehicle.DE 10 2014 015 075 A1 relates to a method for operating an automatically guided, rodless motor vehicle, in particular a passenger car, wherein sensor data recorded by surrounding sensors, comprising at least one camera, of the motor vehicle are evaluated with respect to objects to be taken into account in the trajectory planning, which objects can be classified as an obstacle or no obstacle via at least one classifier evaluating the associated sensor data, wherein, in the case of an object which cannot be classified as an obstacle or no obstacle or cannot be classified with inadequate safety and / or, in the case of at least one object preventing the motor vehicle from moving further to a current destination, at least one camera image of the corresponding object is recorded using at least one of the at least one camera, transmitted to a portable mobile communication device carried by a user of the motor vehicle and displayed there, an input of a user classifying the object as an obstacle or no obstacle is received as classification information, the classification information is transmitted back to the motor vehicle and is taken into account in the further automatic guidance of the motor vehicle.DE 10 2015 007 493 A1 relates to a method for training a decision algorithm which is used in a control device of a motor vehicle and is based on machine learning, wherein the decision algorithm has been determined as a function of input data describing the current operating state and / or the current driving situation for controlling the operation of the motor vehicle and an output data describing the reliability of the output data and trained on the basis of a basic training data set before use in the motor vehicle, wherein, in the case of a reliability value which falls below a threshold value, the input data on which the determination of the output data assigned to the reliability value is based is stored as evaluation input data and is presented to a human evaluation person at a later point in time, wherein output data corresponding to judgment output data is received by an operation input of the judgment person, and the decision algorithm is trained on an improvement training data set formed from the judgment input data and the associated judgment output data.In the context of automated driving, a mis-interpretation may result in an over- or under-reaction. For example, a safety-relevant object is not recognized as such and no corresponding avoidance maneuver is started. Or a safety-irrelevant object is classified as safety-relevant and the vehicle makes an unnecessary driving maneuver that other road users can oversee.It is therefore the object of the present invention to provide computer-implemented methods for relational storage of object properties of an object, which overcome at least one of the aforementioned disadvantages. It is furthermore an object of the invention to provide a computer-implemented method for training an artificial intelligence, computer-implemented methods for generating a class relating to an object, a computer program product, a computer-readable storage medium, a data carrier signal, devices for classifying an object and a vehicle.Disclosure of the InventionThe object is achieved according to the invention by the features of the main claims. Advantageous embodiments can be gathered from the dependent claims.According to a first aspect of the invention, a computer-implemented method for training an artificial intelligence comprises a step in which sensor data are obtained in which an object is detected. The sensor data are generated by at least one sensor arranged on a vehicle.The sensor data can be present, for example, as image data of a camera or radar data of a radar sensor. A vehicle may be, for example, a motor vehicle, a truck, a train, a helicopter, an aircraft or the like. Object characteristics are certain information in the sensor data, typically related to a region in the sensor data. This may be a physical region such as coordinates, or another region such as a frequency. Examples of object properties are points, lines or corners, i.e. geometric properties. A relationship between such information, for example a density of geometric properties or their specific arrangement, is also to be understood as object properties.In a further step of the method, the object is detected and classified by means of the sensor data, wherein a first class is assigned to the object and at least one object property is determined. In this step, a confidence value is generated.This step is a well known task of data and image processing. The skilled person is familiar with various methods available to him to solve this problem. Typically, such an object is achieved by a (convolutional) neural network ((C)NN).In essence, a neural network is trained here to recognize patterns and features within an image or other data indicative of a particular class. In this case, conventional methods generate a confidence value which specifies how likely the class was correctly assigned to the object. The higher the confidence value, the more secure the object is associated with the correct class. Lower values indicate, on the other hand, that the object is not associated with the correct class. This can occur, for example, if the (C)NN has not been trained on the class and this is thus unknown. The first class can therefore also indicate an unknown class.In a further step of the method, an occupant of the vehicle is requested to provide a second class relating to the object if the confidence value is below a first threshold value.The first threshold value represents a boundary, but an assignment to the class is considered to be certain, for example 0.95 or 95%. If the confidence value is below this, an incorrect class may have been assigned to the object. Therefore, in this case, an occupant of the vehicle, for example, the driver, is requested to provide the correct class corresponding to the second class. The request can be made, for example, via an audio system of the vehicle in which the driver is asked "what is this for an object in front of us?". A display may also be used to address such a prompt to the occupant. For example, objects may be displayed to the driver as icons / icons and the driver taps on the appropriate icon / icon.In a further step of the method, a first feedback of the occupant is received.This can be done, for example, via a microphone as an audio feedback or by pressure via a touch display.In a further step of the method, the second class is extracted from the first feedback of the occupant.If the feedback is present as audio feedback, speech recognition may need to analyze the audio feedback to extract the second class. For example, if the occupant responds "This is a chute", only the word "chute" needs to be extracted as the second class.In a further step of the method, a data record is stored which has at least the second class, the at least one object property, a derivative of the sensor data and an identifier of the occupant and / or of the vehicle.The data record can also have only one correlation between the second class, the at least one object property and the derivative of the sensor data, i.e. the actual data are not stored together in a data structure. Thus, the different components of the data set can be stored in a distributed manner.A derivative of the sensor data may be the sensor data itself or a modification thereof, for example compressed sensor data or only that part of the sensor data which contains object properties. This may save memory space.The identifier leaves conclusions about the user of the method. Although the identifier is unique, it can be anonymous. For example, this information may be used to exclude occupants or vehicles from the method attempting to intentionally establish false classes. The occupant can be recognized, for example, via speech recognition, his connected mobile terminal such as a smartphone or face recognition by means of interior cameras, whereby the assignment to an existing identifier or the generation of a new one is possible. An identifier for a vehicle can be realized, for example, by means of a serial number of the vehicle.The artificial intelligence is trained at least with the derivative of the sensor data or the object properties as input and with the second class as output.Methods of machine learning, for example a neural network (NN) or a convolutional NN (CNN), are particularly suitable for this purpose. These can receive the sensor data, for example image data, as input. At the same time, the correct class, namely the second class, is known and can be specified as output to the NN or CNN. In a training phase, the weights of the connections between the artificial neurons of the (C)NN are updated on the basis of an error function.Furthermore, the method has further steps if the confidence value is above a second threshold value:prompting the occupant of the vehicle to confirm the first class of the object,receiving a second feedback of the occupant,extracting an indicator from the second feedback of the occupant, the indicator indicating a truth value whether the object is associated with the first class, andstoring the indicator.For example, the request can be made via an audio system of the vehicle in which the driver is asked "Is you a click?". Here again, it may be necessary to analyze the feedback in order to extract the indicator. For example, if the occupant responds "yes, that is a cut, this may be converted to a binary value of a Boolean variable, for example.The occupant is only prompted to provide feedback if the object has been detected with a confidence above a second threshold value. The second threshold value is typically higher than the first threshold value, for example 0.99 or 99%. However, if the occupant's response also closes an incorrectly recognized class, the occupant may follow the goal of introducing additional or incorrect classes into the method. The indicator can thus be used to identify such users (roles) of the method.The indicator is preferably stored in correlation with the occupant. Storage in correlation with the underlying data set, in particular the sensor data and the object properties, is also advantageous. With the aid of this correlation or information, an artificial intelligence can be trained more specifically.Furthermore, the computer-implemented method for training an artificial intelligence comprises a step in which the previously described steps are carried out for a plurality of occupants. In this case, a plurality of indicators are generated and stored.In a further step of the method, a number of indicators is calculated for each occupant or for each vehicle. Indicators can be assigned to the occupants or vehicles by means of the identifier. In particular, the number of times a particular occupant or vehicle has given feedback indicating that an object has not been correctly classified, even though there is a high confidence value for the classification.In a further step of the method, the artificial intelligence is therefore trained only with data sets for whose identifier the number of indicators is below a third threshold value. The third threshold value defines a limit to which an occupant or vehicle is still deemed trusted. For example, starting from a third threshold value of "5", any further feedback of the occupant concerned is not used for training the artificial intelligence.A generation of a class relating to an object can be provided by first carrying out a method as described above. In this case, a multiplicity of data sets are generated and stored.The data sets can then be divided into groups, wherein data sets having similar object properties and / or similar second classes are grouped in a specific group. For example, data sets with an identical second class, i.e. the class suggested by the occupants, are grouped. Data sets with similar object properties, for example similar geometric patterns such as points or lines, can also be added to the group. A similarity of such patterns can be calculated, for example, by means of a geometric deviation. A similarity of second classes can be calculated, for example, via a translation if the plurality of occupants originates from different countries and therefore does suggest different words for the second class, but actually mean the same. The same applies to occupants who use synonymous words for the same object class, for example "kutsche" and "horse train".Now, a number is calculated for each group. In other words, it is calculated how many occupants or vehicles have already classified certain object properties in a matching manner.A third class is then generated for those groups whose number exceeds a fourth threshold value. In other words, a new class should not be generated from a single feedback. Rather, a swarm intelligence is to be implemented, i.e. a new class is only generated if a plurality of feedbacks have correspondingly classified a previously unknown object. For example, the fourth threshold value may be "10" for this purpose. Thus, 10 occupants must classify an unknown object in a matching manner before an assignment of this class to the object takes place.In a further step of the method, the third class is assigned to the data records in the corresponding groups. A correlation is thus obtained between the feedbacks of the occupants, i.e. the class of the previously unknown object, to the data sets on which the feedback is based.According to a second aspect of the invention, a computer program product comprises instructions which, when the program is executed by a computer, cause the computer to execute a method as described above. The computer program product may be written in a programming language, for example, python or C++.According to a third aspect of the invention, a computer-readable storage medium comprises instructions which, when the program is executed by a computer, cause the computer to execute a method as described above. The computer-readable storage medium can be embodied, for example, as an SSD (solid-state disk) or as flash memory. The computer-readable storage medium can additionally store other data, for example sensor data and / or data which are (temporarily) stored during the execution of the method.According to a fourth aspect of the invention, a data carrier signal transmits the computer program product as described above. The data carrier signal can be transmitted in a cable-bound manner, for example by means of a CAN-BUS (controller area network) or Ethernet cable, or wirelessly by means of Wifi, Bluetooth or the like.According to a fifth aspect of the invention, a device for classifying an object has an evaluation unit which is configured in such a way that it can carry out a method as described above. The evaluation unit can have a processor, for example. The evaluation unit can be distributed as a single component or over a plurality of components. For example, portions of the method may be executed on a server and other portions may be executed on a local processor in a vehicle.Furthermore, the device has at least one memory unit which is communicatively connected to the evaluation unit. The storage unit can store, for example, the computer program product as described above. It is also possible for the data to be stored in databases.The storage unit may be distributed as a single component or among a plurality of components. For example, parts of the data can be stored locally in the vehicle, other parts in turn on a server.Furthermore, the device has at least one sensor which is communicatively connected to the evaluation unit, wherein the sensor generates sensor data in which the object is detected, wherein the at least one sensor is arranged on a vehicle.In an advantageous embodiment, the at least one sensor is a lidar, a radar or a camera. These sensors are easily and inexpensively available and are usually already installed in many vehicles for other functions.According to a sixth aspect of the invention, a vehicle has a device as described above.Summary of the FiguresThe invention is explained in more detail below with reference to exemplary embodiments with the aid of figures. The figures show: FIG. 1 shows a driving scene for describing the problem underlying the invention; FIG. 2 : a first exemplary embodiment of a method for relational storage of object properties; FIG. 3 : a second exemplary embodiment of a method for relational storage of object properties; FIG. 4 : shows an exemplary embodiment of a method for training an artificial intelligence; FIG. 5 : shows an exemplary embodiment of a method for generating a class relating to an object; and FIG. 6 : An apparatus for classifying an object.Detailed Description of the FiguresFIG. 1 shows a driving scene for describing the problem underlying the invention.FIG. 1 shows the driving scene from the perspective of a driver of a vehicle 100. A first road user 102 and a second road user 104 are driving in front of the vehicle 100 on the road. The first road user 102 is a cut road and the second road user 104 is a motor vehicle.The vehicle 100 has a camera 106, in the field of view 108 of which objects 102, 104 on the road in front of the vehicle 100 can be detected. Furthermore, the vehicle 100 has a display device 110. An algorithm has recognized and classified road users 102, 104 from image data 172 of camera 106 by means of an object recognition and classification algorithm. In addition, a digital model of road users 102, 104 has been created with a first digital image 112 and a second digital image 114, which are displayed on display device 110.However, the first road user 102 has been incorrectly classified. Instead of a chute, a truck is displayed as the first digital image 112 because the classification algorithm has not been trained accordingly. This can obscure the driver or lead an automatically driving vehicle to a risky driving maneuver.FIG. 2 shows a first exemplary embodiment of a method for relational storage of object properties 116 of an object 102, 104.In a first obtaining step 118, sensor data 172, 174, 176, 178 are obtained in which the object 102, 104 is detected.In a detection and classification step 120, the object 102, 104 is detected or localized and classified in the sensor data 172, 174, 176, 178. This step can be carried out by conventional methods of image processing. Typically, a neural network (NN) is trained with corresponding data and predefined class.In this step, object property such as geometric patterns and their distribution are determined for object 102, 104. Furthermore, a class is assigned to the object 102, 104, for example "pedestrian", "bicycle", "tree", "motor vehicle", "truck" and the like. In this case, a confidence value is generated which is a measure of how reliable the classification into the class is. The confidence value is typically given as a floating point number such as 0.97 or a percent value such as 97%.In a first decision step 122, a check is made as to whether the confidence value is below a first threshold value 192. If this is not the case (f branch), a safe classification is present and method 116 simply continues continuously. However, if the confidence value is less than the first threshold value 192, then there is no secure classification. The object 102, 104 is considered unknown.In this case (t branch), a first request step 124 is carried out. In this case, an occupant of the vehicle 100 is requested to provide a second class relating to the object 102, 104. The request can be made, for example, by audio signal in which the occupant is asked "What is this for an object in front of the vehicle?". The second class is the correct class to which the object 102, 104 belongs.The occupant's feedback is received in a first receiving step 126. For example, this can again be done by an audio signal. For example, the occupant may answer "That is a cut.".In a first extraction step 128, the second class is extracted from the feedback of the occupant. For example, the word "Kutsch" is isolated from the audio signal and used as a second class.In a first storage step, a data record 193 is stored which has at least the second class, the at least one object property and a derivative of the sensor data 172, 174, 176, 178. A derivative may be the sensor data 172, 174, 176, 178 itself, but may also be a compressed version or a relevant part of the sensor data 172, 174, 176, 178.FIG. 3 shows a second exemplary embodiment of a method for relational storage of object properties 116.Most steps in Fig. 3 (see dotted line) have already been explained in Fig. 2 and will therefore not be discussed again here. In addition, FIG. 3 shows further steps, which will now be explained.If it is determined in the first decision step 122 that the confidence value is above the first threshold value 192, a sufficiently safe classification of the object 102, 104 can be assumed. For example, the first threshold value may be 0.95 or 95%.A check is then made in a second decision step 134 as to whether the confidence value is above a second threshold value 194. For example, this is 0.99 or 99%. If this is the case (t branch), the object 102, 104 is very securely classified into the first class. Subsequent steps 136, 138, 140, 142 are to ensure that the method 116 is not intentionally abused by generating false classifications.For this purpose, the occupant is first requested in a second request step 136 to confirm the first class. This can be realized again, for example, via an audio signal in which the occupant is asked "Is you a cookie?".A second feedback of the occupant is then received in a second receiving step 138. The feedback could be, for example, "yes that is a click" or "no that is a bicycle".Therefore, in a second extraction step 140, the second feedback from the occupant is subsequently evaluated and it is assessed whether or not it is an agreement. This is represented by an indicator, for example a Boolean variable. The indicator thus indicates a truth value whether the object 102, 104 is associated with the first class. The indicator may have binary states (true / false), for example, or have a value along a spectrum. It is thus stored in the indicator whether the occupant gives trusted responses, because it is to be expected that there is a positive response at the present high confidence.The indicator is now stored in a second storage step 142. The storage is effected in such a way that the indicator is linked to the identifier of the occupant or of the car. In other words, information is generated which links the occupant's response to his identity.Subsequently, the training step 132 can be configured more specifically by providing data records 193 of occupants for the training phase of the artificial intelligence only if the occupant is classified as trusted via the indicator.FIG. 4 shows an exemplary embodiment of a method for training an artificial intelligence 144.In a first execution step 146, the method from FIG. 3 is carried out continuously (illustrated by the circular arrow) for a plurality of occupants or vehicles 100. In this case, a corresponding plurality of indicators is also generated and stored.In a first calculation step 148, a number is calculated for the plurality of indicators. The number indicates how frequently an occupant has given a trusted response, since the indicator is only generated for objects 102, 104 which have been classified with very high certainty. The number is thus a measure of the trustworthiness of the occupant.The number is then compared with a third threshold value 196 in a third decision step 150. If the number is lower than the third threshold value 196 (f branch), all data records 193 which originate from the occupant or vehicle 100 can be used for training the artificial network in the training step 132. If the number is higher than the third threshold 196, the data records 193 associated with the occupant or vehicle 100 are excluded from the training step 132. The third threshold 196 is thus a measure of where an occupant or vehicle 100 is to be deemed trusted.FIG. 5 : shows an exemplary embodiment of a method for generating a class 152 relating to an object.In a second execution step 154, the method from FIG. 2 is carried out continuously (illustrated by the circular arrow) for a multiplicity of occupants or vehicles 100. In this case, a corresponding plurality of data sets 193 is also generated and stored.In a classification step 156, the data records 193 are classified into groups. The members of a group are distinguished by a similarity of their object properties and / or their second classes. For example, a particular arrangement (e.g., distances) of points, lines, or corners may cause similarity.In a second calculation step 158, a number is then calculated for each group. The number is thus a measure of how many different occupants or vehicles 100 have transmitted a matching classification for an object 102, 104.A check is then made in a fourth decision step 160 as to whether the number exceeds a fourth threshold value 198. The fourth threshold value 198 is a measure for sufficiently similar feedback with respect to a class of an object 102, 104. If this is the case (t branch), a generation step 162 is carried out.In the first generation step 162, a third class is generated for the records in the groups. The third class can correspond in particular to the second class, i.e. the class transmitted by the occupants is adopted.In an assignment step 164, the third class is assigned to the datasets 193. Thus, the data records 193 are now provided with a reliable class.Subsequently, the artificial intelligence can be trained with the data records 193 provided with the third class. The sensor data 172, 174, 176, 178 (or a derivative) present in the data sets 193 or else the object properties or a combination thereof serve as input for the artificial intelligence, wherein the class present in the data sets 193 serves as specification for the output. In training step 132, the internal characteristics of the artificial intelligence, for example weights between artificial neurons, may be adjusted.FIG. 6 shows a device 166 for classifying an object 102, 104.The device 166 includes a camera 106, a lidar sensor 168, and a radar sensor 170. The camera 106 generates image data 172, the lidar sensor 168 generates lidar data 174, and the radar sensor 170 generates radar data 176.The sensor data 172, 174, 176 that these sensors 106, 168, 170 generate are very well suited for detecting objects 102, 104 with them, since they contain object characteristic patterns.Before the sensor data 172, 174, 176 are further processed, they are still fused to form fused sensor data 178 in the example of FIG. 6. These are then forwarded to a high performance computer (HPC) 180. The HPC 180 has an evaluation unit 182, which is realized as a processor 182 in the example of FIG. 6. In addition, the HPC 180 includes a non-volatile memory 184 communicatively coupled to the processor 182 by the involvement of a volatile memory 200.Stored on the non-volatile memory 184 is a computer program product 186 which includes instructions which, when executed by the processor 182, cause the latter to execute the methods 116, 144, 152 of FIGS. 3-5. Furthermore, a first database 188 is stored on the non-volatile memory 184, in which database, inter alia, the fused sensor data 178 can be stored. In addition, a first neural network (NN) 190 is stored on the nonvolatile memory 184, which has been trained to detect and classify objects 102, 104. Furthermore, a first threshold value 192 and a second threshold value 194 are stored on the nonvolatile memory 184, which threshold values are necessary for executing the methods 116, 144, 152 from FIGS. 3 to 5.The device 166 furthermore has an external memory 202, which is realized as cloud memory 202. This has a second neural network 204 and a second database 206. Furthermore, a third threshold value 196 and a fourth threshold value 198 are stored on the cloud memory 202, which threshold values are necessary for executing the methods 116, 144, 152 from FIGS. 3 to 5.The device 166 is communicatively connected to an audio system 208 of a vehicle 100 in a bi-directional manner.The computer program product 186 is now executed on the processor 182. In this case, the fused sensor data 178 are evaluated by means of the first NN 188. In this case, the object 102, 104 is detected and classified in the fused sensor data 178. In this case, object properties such as geometric patterns are recognized in the fused sensor data 178. The object properties are characteristic of a first class which is assigned to the object 102, 104. The first NN 188 has implicitly learned these characteristics in a previous training phase. The first NN 188 also generates a confidence value 191, which specifies how safe the classification was in the first class.It is now checked whether the confidence value 191 is below the first threshold value 192, for example 0.95 or 95%. If so, there is no safe classification and an occupant of the vehicle 100 is prompted via the audio system 208 to provide a second class by inquiring "What is for an object in front of the vehicle?". The response is received and extracted from it the second class. For example, if the occupant responds "This is a chute", then "chute" is extracted as the second class. The second class is then stored together with the object properties, the fused sensor data 178 and an identifier as a data record 193 in the first database 186. The identifier is a specific value assigned per occupant or vehicle 100. It ensures that data sets 193 can be assigned to a particular source. This is preferably performed anonymously.However, if the confidence value 191 is above the second threshold value 194, for example 0.99 or 99%, a very safe classification is present and the occupant is requested to provide a confirmation by inquiring "Is this a truck?". For example, if the occupant answers "yes, that is a truck", or "no, that is no truck", or "that is an aircraft", an indicator 195 is extracted from this feedback. In the example of FIG. 6, the indicator 195 is generated as a binary Boolean variable, i.e. with the values "true" or "false". The indicator 195 is then stored together with the underlying data record 193. In this case, it can be added to the data record 193 or a reference to the data record 193 can be appended to the indicator 195.The records 193 and indicators 195 in the first database 186 are wirelessly added to the second database 206 in the cloud 202. A plurality of feedback from a plurality of occupants or vehicles 100 are collected in the cloud 202.In the cloud 202, it is now checked how many data records 195 of an occupant or of a vehicle 100 have an indicator 195 which indicates negative feedback. A negative feedback indicates that the occupant or vehicle 100 is not providing reliable responses because the classification of the object 102, 104 was very likely correct. If the number of these data sets 195 is above the third threshold value, for example "10", these data sets 195 are not used to train the second NN 204.Further, grouping is performed in the cloud 202. A group is associated with records 193 if they have similar second classes and similar object properties, i.e., if the plurality of occupants or vehicles 100 have made a matching classification. If the number of data records in a group exceeds a fourth threshold value, for example "20", it can be assumed that the object properties in the data records 193 actually correspond to the second class of the data records 193.Therefore, a third class is generated for the group which corresponds to the second in the example of FIG. 6. The third class can also differ from the second, however, for example because different languages were used for the second class. The third class is assigned to the respective members in the group.The second NN 204 is now trained with the data sets 193 as long as they have not been excluded from the training by too large a number of indicators. After the training phase, the second NN 204 has implicitly learned to recognize the new third class and can be transmitted back to the vehicle 100. In other words, the first NN 188 is replaced by the second NN 204, or the internal properties such as weights are transmitted to the first NN 188 for update. The vehicle 100 can now recognize the new, third class and, on the basis thereof, create better planning of driving maneuvers or generate more accurate representations of digital objects 112, 114 for a display device 110 of the vehicle 100.List of reference characters100 Vehicle 102 First road user; Coaster 104 Second road user; Car 106 Camera 108 Field of view 110 Display device 112 First digital image 114 Second digital image 116 Method for relational storage of object properties 118 First obtaining step 120 Detection and classification step 122 First decision step 124 First requesting step 126 First receiving step 128 First extracting step 130 First storing step 132 Training step 134 Second decision step 136 Second requesting step 138 Second receiving step 140 Second extracting step 142 Second storing step 144 Method for training an artificial intelligence 146 First performing step 148 First calculating step 150 Third decision step 152 Method for generating a class 154 relating to an object Second performing step 156 Dividing step 158 Second calculating step 160 Fourth decision step 162 Generating step 164 Assigning step 166 Device 168 Lidar sensor 170 radar sensor 172 image data 174 lidar data 176 radar data 178 fused sensor data 180 central computer; HPC 182 evaluation unit; processor 184 nonvolatile memory 186 computer program product 188 first database 190 first neural network 191 confidence value 192 first threshold value 193 data set 194 second threshold value 195 indicator 196 third threshold value 198 fourth threshold value 200 volatile memory 202 external memory; cloud 204 second neural network 206 second database 208 audio system

Claims

Computer-implemented method for training an artificial intelligence (144), comprising the steps of: a) obtaining (118) sensor data (172,174,176,178) in which an object (102, 104) is detected, wherein the sensor data (172,174,176,178) are generated by at least one sensor (106, 168, 170) arranged on a vehicle (100), b) detecting and classifying (120) the object (102, 104) by means of the sensor data (172,174,176,178), wherein a first class is assigned to the object (102, 104) and at least one object property is determined, thereby generating a confidence value (191), c) prompting (124) an occupant of the vehicle (100) to provide a second class relating to the object (102, 104), if the confidence value (191) is below a first threshold value (192), d) receiving (126) a first feedback of the occupant, e) extracting (128) the second class from the first feedback of the occupant, and f) storing (130) a data set (193) which comprises at least the second class, the at least one object property, a derivative of the sensor data (172,174,176,178) and an identifier of the occupant and / or of the vehicle (100), wherein the method (116) comprises the following further steps if the confidence value (191) is above a second threshold value (194): g) requesting (136) the occupant of the vehicle (100) to confirm the first class of the object (102, 104), h) receiving (138) a second feedback of the occupant, i) extracting (140) an indicator (195) from the second feedback of the occupant, the indicators (195) indicating truth values whether the object (102, 104) is associated with the first class, j) storing (142) the indicator (195), k) performing (146) steps g) to j) for a plurality of occupants, thereby generating a plurality of indicators (195) that are stored, l) calculating (148) a number of the indicators (195) for each occupant or for each vehicle (100), and m) training (132) the artificial intelligence (190, 204) at least with the derivative of the sensor data (172,174,176,178) or the object properties as input and with the second class as output, characterized in that, using only those data sets (193) for the training for whose identifier the number of indicators (195) is below a third threshold value (196).A computer program product (186) comprising instructions which, when the program is executed by a computer, cause the computer to perform a method (116, 144, 152) according to claim 1.A computer readable storage medium (184) comprising instructions which, when the program is executed by a computer, cause the computer to perform a method (116, 144, 152) according to claim 1.A data carrier signal carrying the computer program product (186) of claim 2.An apparatus for classifying an object (166), comprising: a) an evaluation unit (182) configured to be able to execute a method (116, 144, 152) according to claim 1, b) at least one storage unit (184, 202) communicatively connected to the evaluation unit (182), and c) at least one sensor (106, 168, 170) communicatively connected to the evaluation unit (182), wherein the at least one sensor (106, 168, 170) generates sensor data (172,174,176,178) in which the object (102, 104) is detected, wherein the at least one sensor (106, 168, 170) is arranged on a vehicle (100).The device according to claim 5, characterized in that the at least one sensor (106, 168, 170) is a lidar (168), a radar (170) or a camera (106).Vehicle (100) comprising a device (166) according to claim 5 or 6.

Images