Method for generating a vehicle environment data evaluation model of machine learning and environment data evaluation device for processing environment data to perform at least a semi-automatic vehicle function of the vehicle.

Abstract

Method for generating a trained vehicle environment data evaluation model of machine learning for processing environment data acquired using at least one environment sensing device for a vehicle (1) for performing at least a semi-automatic vehicle function of the vehicle on the basis of the acquired environment data, wherein the vehicle environment data evaluation model to be trained comprises a set of parameters which are adjustable to values ​​as a result of a training process.According to the invention, the training process of the vehicle environment data evaluation model to be trained is based on training data which is at least partially based on external environment data generated by at least one external vehicle and whose evaluation by the external vehicle was classified as faulty and / or critical with regard to the validity of an evaluation result and was transmitted to at least one external storage device for use in the training process.

Description

[0001] The present invention relates to a method for generating a machine learning vehicle environment data evaluation model for processing environment data to carry out an at least partially automatic vehicle function of the vehicle, as well as an environment data evaluation device for a vehicle with at least one environment detection device for detecting environment data to carry out an at least partially automatic vehicle function, preferably a perception function, of the vehicle on the basis of the detected environment data.

[0002] The present invention further relates to a method for generating a training data set for training a machine learning vehicle environment data evaluation model for processing environment data acquired by means of at least one environment detection device for a vehicle for carrying out an at least partially automatic vehicle function, preferably a perception function, of the vehicle on the basis of the acquired environment data.

[0003] It is known from the state of the art that errors occur in AI-based methods for recognizing traffic signs, for example, when real-world objects are distorted in a way that is barely or not at all perceptible to the human eye. For example, if a small sticker with colorful pixels is applied to a stop sign, the vehicle AI will no longer recognize the sign correctly and may therefore make a serious driving error, even leading to an accident with property damage or personal injury.

[0004] The Fraunhofer Society also reported on safety and robust learning methods in a brochure entitled "MACHINE LEARNING - AN ANALYSIS OF COMPETENCIES, RESEARCH, AND APPLICATION." One measure to increase safety in machine learning systems ("ML systems") is the incorporation of safety margins. A key aspect here is robustness, i.e., ensuring that small changes in the input only result in small changes in the output. This report also referenced a 2017 headline in which deep ANNs produced serious misclassifications due to a change in a few pixels in an image that was imperceptible to humans. Stop signs that had been changed imperceptibly to humans were no longer recognized as such by the system, and some rifles were categorized as helicopters. This was considered a serious problem with safety-relevant implications.The report also describes an approach to counteract this, called “adversarial training,” in which models are deliberately trained with noisy data so that they generalize more robustly in use.

[0005] So-called "adversarial" training is specifically a supervised brute-force learning method in which as many adversarial examples as possible are fed into the model and explicitly labeled as threats. In the state of the art, this approach is compared to the approach used by typical antivirus software on PCs, with multiple updates per day. Although highly effective, it requires continuous maintenance to keep up with new threats and still suffers from the fundamental problem that it can only prevent something that has already happened from happening again (see https: / / deepai.org / machine-learning-glossary-and-terms / adversarial-machine-learning; accessed August 17, 2023).

[0006] Other defense methods against attacks on neural networks are known from the state of the art (see https: / / www.heise.de / hintergrund / Neuronale-Netze-Wie-sieangegriffen-werden-und-wie-man-sie-verteidigt-6132752.html?seite=2; accessed on August 17, 2023). Among other things, autoencoders can be used to preprocess the altered image. Since the manipulations often occur in the high-frequency portion, the images can be smoothed or denoised using an autoencoder. Another option is gradient masking: Many methods are based on calculating gradients, and this approach pursues the strategy of intentionally concealing the model's associated information to confuse attackers. A final example is the detection of manipulated images: By training on normal and manipulated images, classifiers detect when an attack is taking place.

[0007] US 2021 300 433 A1 discloses a method for training image recognition in autonomous vehicles. A training dataset of input images is received and adversarial images are generated from selected input images. A DOA image classifier is then trained to defend against attacks using the training dataset of input images, the generated adversarial images, and machine learning / neural network techniques.

[0008] US 2021 397 896 A1 discloses a method in which an image with a known first classification is received by the machine learning model and iteratively modified using at least one perturbation algorithm. The modified images are input to the machine learning model until the machine learning model outputs a second classification that differs from the first classification. Data characterizing the modifications to the image that led to the second classification can be provided.

[0009] The present invention is based on the object of overcoming the disadvantages known from the prior art and of providing a method for training a machine learning vehicle environment data evaluation model for processing environment data to carry out an at least partially automatic vehicle function of the vehicle, as well as an environment data evaluation device for a vehicle with at least one environment detection device for detecting environment data to carry out an at least partially automatic vehicle function of the vehicle on the basis of the detected environment data, which offer a very reliable evaluation of the environment data.

[0010] The object is achieved according to the invention by the subject matter of the independent claims. Advantageous embodiments and further developments of the invention are the subject matter of the dependent claims.

[0011] In a method according to the invention (preferably at least partially and particularly preferably computer-implemented in its entirety) for generating a trained vehicle environment data evaluation model using machine learning for processing environmental data acquired by means of at least one environmental detection device for a vehicle in order to carry out an at least partially automatic (preferably fully automatic) vehicle function of the vehicle, preferably a perception function, on the basis of the acquired environmental data, the vehicle environment data evaluation model to be trained comprises a set of parameters which can be set to values ​​as a result of a training process.

[0012] The environment detection device for a vehicle is preferably an environment detection device of the vehicle, which is particularly preferably fixedly arranged on the vehicle and / or is vehicle-bound.

[0013] Preferably, to perform the vehicle function (particularly in real time), particularly in a computer-implemented method step, at least one control variable is determined (based on the acquired environmental data) and provided (to perform the vehicle function) and / or transmitted (to a control device of the vehicle). Preferably, the at least one control variable is determined by means of the (trained) vehicle environmental data evaluation model by processing the (respective) environmental data.

[0014] Preferably, the machine learning vehicle environment data evaluation model is based on an (artificial) neural network. Preferably, the determination of the control variable is based on (in particular) the (artificial) neural network. In particular, the vehicle sensor data is processed using the specifically trained (artificial) neural network.

[0015] Preferably, the neural network is designed as a deep neural network (DNN), in which the parameterizable processing chain has a plurality of processing layers, and / or a so-called convolutional neural network (CNN) and / or a recurrent neural network (RNN) and / or transformer networks.

[0016] Preferably, the data (to be processed), in particular the environmental data (or data derived therefrom), are fed to the vehicle environment data evaluation model or the (artificial) neural network as input variables. Preferably, the vehicle environment data evaluation model or the artificial neural network maps the input variables to output variables depending on a parameterizable processing chain (or depending on the set of parameters), wherein the at least one control variable is preferably selected as at least one output variable. The at least one output variable can be an evaluation result, for example a label (of an object detected in the environmental data). It is also conceivable that the at least one control variable is determined based on the output variable(s).

[0017] The input variable(s) and / or output variable(s) of the vehicle environment data evaluation model depend on the (application) task to be fulfilled by the vehicle environment data evaluation model, for example which control variable(s) are to be determined (to fulfill the vehicle function) and / or which classification task(s) and / or object detection task(s) and / or (semantic) segmentation task(s) are to be fulfilled from which input variable(s).

[0018] According to the invention, the training process of the vehicle environment data evaluation model to be trained is carried out on the basis of training data which is at least partially based on external environment data which is generated by at least one external vehicle and whose evaluation by the external vehicle (or the evaluation carried out by the external vehicle) was classified as erroneous and / or critical with regard to the validity of an evaluation result (the classification can be carried out by the external vehicle and / or by a user of the external vehicle and / or by an (external) operator) and was transmitted (by the external vehicle) to at least one external storage device for use in the training process.

[0019] Preferably, the training process parameterizes the parameterizable processing chain of the vehicle environment data evaluation model (to be trained) or sets the parameters of the set of parameters to (fixed) values.

[0020] In other words, the vehicle environment data evaluation model to be trained is trained (in particular, among other things) with training data generated on the basis of external environment data, which have (already) proven to be erroneous and / or critical in an evaluation by an external vehicle with regard to the validity of the evaluation result.

[0021] This allows the experience of third-party vehicles already in the field to be used to advantage.

[0022] This allows a vehicle class in autonomously driving vehicles to be trained in such a way that errors due to consciously or unconsciously introduced image or object changes are detected and the image recognition then delivers the correct result.

[0023] Preferably, a first set of training data is used as the training data set in the training process, in which the training data is provided with correct annotation data, and a second set of training data is used, in which the training data is provided with incorrect and / or critical annotation data and with data indicating the ground truth.

[0024] In particular, an image recognition system can be trained by presenting unmodified and falsified images or objects and naming the correct object in each case and thus learning it - (for example, the stop sign falsified with a sticker should also be recognized as a stop sign).

[0025] The other vehicle is preferably equipped with a pre-trained machine learning vehicle environment data evaluation model for processing the external environment data acquired and / or generated by an environment detection device (of the other vehicle). The machine learning vehicle environment data evaluation model can be based on the same machine learning vehicle environment data evaluation model to be trained as the one for the vehicle. It is also conceivable that it is a different machine learning vehicle environment data evaluation model, which is (only) suitable and intended for performing the same evaluation function or perception function (such as image classification, semantic segmentation and / or object recognition and the like).

[0026] The method relates in particular to deep learning and / or the isolation of relevant image and / or environmental data information.

[0027] In an additional or alternative method, the training process is based on training data that is at least partially based on external environment data, which can only be mapped to an evaluation result by at least one machine learning vehicle environment data evaluation model with a probability that is lower than a predetermined and / or predeterminable probability value. This offers the advantage that external environment data is taken into account in the training process, for example, to a greater extent, where existing machine learning vehicle environment data evaluation models exhibit difficulties or indicate difficulties with regard to a valid evaluation. Therefore, increased training can be carried out in this critical area of ​​training data.

[0028] The machine learning vehicle environment data analysis model can be a vehicle environment data analysis model of a third-party vehicle. However, it is also conceivable that it is a vehicle environment data analysis model that is being tested and / or developed and has been trained for testing and / or development purposes.

[0029] It is conceivable that the probability value can be specified by a user (of the vehicle environment data evaluation model) of the other vehicle and / or a test person or a developer of the machine learning vehicle environment data evaluation model. However, it is also conceivable that the probability value is (uniformly) fixed.

[0030] Preferably, if a probability lower than the probability value of an evaluation result is determined, the machine learning vehicle environment data evaluation model is suitable and determined to submit a request to a user (preferably via a user interface) to review and / or correct the evaluation result. Preferably, annotation data, such as at least one label, is generated (preferably automatically upon user input) from the review and / or correction (by the user) for the external environment data to be reviewed. This external environment data is preferably used together with the annotation data as training data in the training process or added to the training data set to be used in the training process.

[0031] This offers the advantage of a safety loop. If the AI ​​or the machine learning vehicle environment data analysis model is unable to identify a specific image or object, it reports, labels the raw data, and sends it to a human instance. This instance adds the image information to the entire set of objects to be learned and also adds corrupted variants for the learning process. If deviations are detected, the human user can also inform or train the AI.

[0032] The training process preferably includes a so-called "adversarial patch training," in which the vehicle environment data evaluation model is trained to specifically mask out non-relevant image and / or object components and to disregard the non-relevant image and / or object components when deriving an evaluation result. A distinction between non-relevant image and / or object components can be made, for example, depending on a color space and / or a surface texture and / or a gloss level and / or an outline contour of an image and / or object component of the same color space and / or sensor data values ​​from the same range.

[0033] In particular, the AI ​​learns to selectively mask out irrelevant parts of the image or object and focus on the essential features that define the image or object (e.g., the shape and color of the sign and the inscription "STOP"). Using color as an example: If, for example, more than 98% of the sign's surface is red, the remaining "colored" 2% should be ignored. The same applies to the white color of the "STOP" lettering.

[0034] Preferably, the training process of the vehicle environment data evaluation model of machine learning comprises changing predefined training data, for example by changing the values ​​of individual or multiple data points of the training data.

[0035] The modification of the training data is preferably based on at least one randomly generated variable. It is also conceivable that the values ​​of the training data points are modified in such a way that the changes are characteristic of noise in the training data. It is also conceivable that the training data are modified in such a way that the changes include noise information.

[0036] When learning a single element (e.g. stop sign), the AI ​​calculates millions of stop signs from the one target image to be learned (and labeled as “stop sign”), essentially variants with a wide variety of errors or interference information, e.g. the addition of noise.

[0037] This has the advantage that the learning of the objects can be done completely virtually (disturbance information is generated and inserted using an algorithm with random aspects) and therefore does not require expensive additional data labeled by humans.

[0038] Additionally, patterns from the real world (similar to a virus signature in virus scanners) can be imported as needed, allowing the AI ​​of the entire fleet to be improved virtually instantly. (For example, people with special interests have designed a new type of sticker with a disturbing image that they specifically stick on stop signs.)

[0039] Preferably, the training process comprises the iterative execution of training steps, in which a change to the training data occurs in a first training step. In a subsequent training step, it is checked whether the changed training data yields the same evaluation result or the same output variable(s) of the vehicle environment data evaluation model as the training data before their (last and / or first) changes. For example, it is conceivable that the output variable(s) is / are compared with the annotation data (labels) of the original training data.

[0040] Preferably, the first and second training steps are repeated iteratively until the output variable(s) changes and, most preferably, no longer matches the annotation data of the original training data.

[0041] In a preferred method, the at least one external storage device is automatically queried before the start of the training process, and external environmental data transmitted to the at least one external storage device (in particular from external vehicles) is taken into account when creating the training data intended for the training process. This offers the advantage that all critical or problematic external environmental data or interference information already detected by an external vehicle in the vehicle swarm at that time can be taken into account in the training process. This enables a high level of up-to-dateness.

[0042] In a further preferred method, it is assessed whether the training process and / or a further training process is initiated based on the external environment data transmitted by the at least one external vehicle. This offers the advantage that external environment data with, for example, a particularly high degree of criticality or which is particularly likely to be incorrectly evaluated by the vehicle environment data evaluation model (in particular of the respective external vehicle and / or a central vehicle environment data evaluation model that uploads and / or retrieves the external environment data from the at least one external storage device) can initiate or trigger a training process, so that this external environment data can be taken into account particularly promptly.

[0043] This can increase road safety. The high probability is preferably greater than a predefined and / or (by a central operator) predeterminable hazard value, which is preferably 20%, preferably at least 30%, preferably at least 50%, preferably at least 75%, and particularly preferably at least 80% (probability of incorrect evaluation).

[0044] It is also conceivable that an (automatic) initiation or triggering of a training process occurs (only) when a predetermined amount of external environment data and / or external environment data from a predetermined number of different external vehicles are available, the probability of which exceeds the predetermined and / or predeterminable risk value. This can advantageously enable an (automatic) training process to be initiated only when a certain level of relevance is reached.

[0045] An (automatic) initiation or triggering can be achieved, for example, by issuing a warning message and / or a user prompt (via an output device) to perform a training process and / or to check whether a training process should be performed. The output can be provided visually and / or acoustically (preferably depending on the probability of incorrect evaluation) and preferably at different warning levels.

[0046] It is preferred, particularly when taking into account external environmental data with a probability of incorrect evaluation by the vehicle environment data evaluation model (of the respective external vehicle and / or the central vehicle environment data evaluation model), which exceeds the specified and / or specifiable risk value, that after completion of the training process of the vehicle environment data evaluation model (triggered thereby), the vehicle environment data evaluation model is made available for updating and / or downloading the vehicle environment data evaluation models on external vehicles - preferably automatically.

[0047] Preferably, the (automatic) update and / or the (automatic) download must be enabled (in particular by a (central) operator) before a (in particular any) third-party vehicle is able to download and / or update the trained vehicle environment data evaluation model.

[0048] Preferably, an (automatic) update and / or an (automatic) download is carried out wirelessly (in particular via an over-the-air update; “software over the air”).

[0049] In a further preferred method, the at least one external storage device is checked repeatedly and / or at regular intervals for the presence of external environmental data transmitted by the at least one other vehicle. This enables the at least one external storage device to be checked (automatically) at regular intervals for the presence of problematic or critical external environmental data detected by other vehicles. This achieves regular updating. The time intervals are preferably shorter than 1 year, preferably shorter than 6 months, preferably shorter than 3 months, preferably shorter than 4 weeks, and particularly preferably shorter than 2 weeks.

[0050] In a further preferred method, a plurality of third-party vehicles transmits external environment data generated by the respective third-party vehicles, the evaluation of which was classified as erroneous and / or critical by the respective third-party vehicle, to the at least one external storage device. This ensures that a swarm of vehicles collects such critical or erroneous external environment data. Preferably, the plurality of third-party vehicles comprises more than 10, preferably more than 100, and particularly preferably more than 1000 vehicles. The integration of so many third-party vehicles enables rapid detection of problematic external environment data, if necessary with regard to their evaluation.

[0051] In a further preferred method, the at least one external vehicle (preferably the plurality of external vehicles) transmits annotation data determined during the evaluation of the external environment data and (respectively) associated with the external environment data to the at least one external storage device. The annotation data can be, for example, labels.

[0052] The annotation data may include annotation data that is erroneous and / or highly likely to be erroneous (with respect to the respective external environment data). The annotation data may also include annotation data that has been corrected and / or entered and / or assessed as correct by an operator.

[0053] In a further preferred method, if external environmental data is classified as faulty and / or critical with regard to the validity of an evaluation result following evaluation thereof, the at least one external vehicle stores the external environmental data on an internal storage device and / or this external environmental data is transmitted to the external storage device as part of a repair and / or inspection operation and / or service and / or maintenance operation of the vehicle, for example carried out by a workshop. The workshop can therefore, for example, read a (vehicle-bound) internal memory of the vehicle for the presence of evaluation results (obtained by the vehicle environmental data evaluation model) that are faulty and / or critical with regard to the validity of the evaluation result and, if necessary, trigger and / or initiate and / or carry out a transmission to the at least one external storage device.

[0054] It is also conceivable that the workshop can carry out an automatic triggering and, if necessary, automatic transmission to at least one external storage device.

[0055] However, it is also conceivable that - for example at regular intervals and / or repeatedly and / or when a predetermined storage capacity and / or stored data volume of external environment data is reached - the external vehicle independently transmits data to the at least one external storage device.

[0056] Preferably, an external server accessing the external storage device first checks the external environment data stored on the external storage device for harmful environment data before checking whether these should be taken into account in the training process.

[0057] The training process preferably includes a so-called “adversarial patch training”, in which locally limited disturbances are preferably eliminated (location-optimized).

[0058] Preferably, the training process comprises a so-called “adversarial perturbation”, by means of which a so-called “clean training” is carried out, in particular a cleaning by eliminating disturbing background noise.

[0059] The training process preferably considers error messages from the field or from other vehicles (or their users). In particular, the AI ​​or the vehicle environment data evaluation model can report independently in this case. Error images are preferably labeled using a multi-eye principle. Particularly preferred are images or (other) environment data or training data that are not correctly recognized, which are forwarded to a human jury.

[0060] Preferably, error messages from the field or from other vehicles, especially from their users, are taken into account in the training process. Preferably, a user (of a other vehicle) specifies a correct or expected result. Preferably, the AI ​​or the vehicle environment data evaluation model learns this or adapts the algorithm accordingly. For this purpose, a control panel and / or an actuating element and / or an input option can be provided in the other vehicle and / or a linked or assigned input device (e.g., a user's end device, such as a smartphone).

[0061] In a further preferred method, the environment detection device is a sensor device for detecting spatially resolved (in particular 2D and / or 3D) sensor data, in particular image data (preferably from a vehicle environment of the respective vehicle).

[0062] The vehicle function is preferably a driving function of the vehicle, for example, a longitudinal and / or lateral guidance function of the vehicle. The vehicle function is preferably a vehicle function that is (essentially) executed and / or can be executed in real time, or a vehicle function that is executed during operation and / or movement of the vehicle and / or in real time.

[0063] Preferably, the vehicle function is or is performed based on the (recorded) environmental data. To perform the vehicle function, the vehicle sensor data is processed using the trained machine learning vehicle environment data evaluation model, particularly in real time and preferably while the vehicle is moving (in a computer-implemented method step).

[0064] Preferably, the (external) environment data are data collected in relation to an external (external) vehicle environment (by the environment detection device).

[0065] The vehicle function is preferably a function of a vehicle component of the vehicle. The vehicle component is preferably selected from a group comprising an environment perception system (for example, a camera image analysis system and / or object recognition system for recognizing objects in environmental data acquired by environment detection devices), a vehicle safety system (for example, a user warning system and / or a braking system), a system for guiding the vehicle, in particular an automatic vehicle lateral control system, a lane keeping system, a lane change assistance system, a navigation system, a vehicle locking system, an infotainment system, an entertainment system, and / or a comfort system for increasing the driving comfort of an occupant, and combinations thereof.

[0066] It is also conceivable that the vehicle function is a function for observing and / or monitoring and / or detecting an operating action of (at least) one user of the vehicle (in particular, the driver). For example, it could be hands-off detection, which is intended to detect when the driver is not (properly) holding their hands on the steering wheel. In this case, the (external) surroundings data preferably comprises data collected (by the environment detection device) relating to the (external) vehicle interior.

[0067] Preferably, the environment detection device is a detection device for detecting a vehicle environment (of the vehicle) and is preferably selected from a group of sensor devices which comprises a (color) camera, a front camera, a rear camera, an infrared camera, LIDAR (abbreviation for light detection and ranging or light imaging, detection and ranging), radar, ultrasonic sensors and the like, as well as combinations thereof.

[0068] The at least one external storage device is preferably a server, in particular a cloud-based and / or an external server. An external storage device or an external server is understood to mean, in particular, a server that is external to the vehicle and / or to the at least one third-party vehicle or the plurality of third-party vehicles, in particular a backend server. The external server is, for example, a backend of a vehicle manufacturer or a service provider, which is configured to evaluate, manage and / or store environmental data and / or data derived therefrom (preferably in each case relating to traffic situations). The functions of the backend or the external server can be carried out on (external) server farms. The (external) server can be a distributed system. The external server and / or the backend can be cloud-based.

[0069] The present invention is further directed to a, in particular processor-based, environmental data evaluation device for a vehicle having at least one environmental detection device for detecting environmental data for carrying out an at least partially automatic vehicle function, preferably a perception function, of the vehicle on the basis of the detected environmental data.

[0070] The environment data evaluation device is suitable and intended to receive the environment data detected by the environment detection device of the vehicle and to process it using a trained vehicle environment data evaluation model of machine learning.

[0071] The trained vehicle environment data evaluation model comprises a set of parameters which are set to values ​​learned as a result of a training process.

[0072] According to the invention, the trained vehicle environment data evaluation model is generated according to one of the above-described embodiments of a method for generating a trained vehicle environment data evaluation model using machine learning.

[0073] It is therefore also proposed within the scope of the environmental data evaluation device according to the invention that external environmental data are taken into account in the training process of the vehicle environmental data evaluation model, which external environmental data were generated by at least one external vehicle and are or are assessed as erroneous and / or critical with regard to a validity of an evaluation result (obtained by means of a vehicle environmental data evaluation model and / or data processing and / or data evaluation of the external environmental data by a processor device of the external vehicle).

[0074] The present invention is further directed to a vehicle, in particular a motor vehicle, comprising an environmental data evaluation device for a vehicle as described above, according to one embodiment. The vehicle can, in particular, be a (motorized) road vehicle.

[0075] A vehicle can be a motor vehicle, in particular a motor vehicle controlled by the driver himself (“driver only”), a semi-autonomous, autonomous (for example, autonomy level 3, 4 or 5 (of the SAE J3016 standard)) or self-driving motor vehicle. Autonomy level 5 refers to fully automated vehicles. The vehicle can also be a driverless transport system. The vehicle can be controlled by a driver or drive autonomously. In addition to a road vehicle, the vehicle can also be an air taxi, an aircraft and another means of transport or another type of vehicle, for example an aircraft, watercraft or rail vehicle.

[0076] The present invention is further directed to a computer program or computer program product comprising program means, in particular a program code, which represents or encodes at least some and preferably all of the method steps of the method according to the invention and preferably one of the described preferred embodiments and is designed to be executed by a processor device.

[0077] The present invention is further directed to a data memory on which at least one embodiment of the computer program according to the invention or a preferred embodiment of the computer program is stored.

[0078] The present invention is further directed to a method for generating a training data set for training a machine learning vehicle environment data evaluation model for processing environmental data acquired using at least one environmental detection device for a vehicle in order to perform an at least partially automatic vehicle function, preferably a perception function, of the vehicle based on the acquired environmental data. The vehicle environment data evaluation model to be trained comprises a set of parameters whose values ​​can be set as a result of a training process.

[0079] According to the invention, training data (for carrying out the training process) are generated or determined which are based at least partially on external environmental data which are generated by at least one external vehicle and whose evaluation by the external vehicle was classified as erroneous and / or critical with regard to the validity of an evaluation result and which were transmitted to at least one external storage device for use in the training process.

[0080] In particular, the method can comprise all of the method steps described above in connection with the method for generating a trained vehicle environment data evaluation model (in particular the determination and / or generation of the training data) individually or in combination with one another.

[0081] The present invention has been described with reference to an environment detection device for a vehicle. The present invention is also applicable to methods and systems generally used in machine image processing of environmental data recorded by environment detection devices (e.g., in infrastructure). The applicant reserves the right to also claim related subject matter.

[0082] Further advantages and embodiments can be seen from the attached drawings: Showing: Fig. 1 a schematic representation of a vehicle with an environmental data evaluation device according to the invention according to an embodiment; Fig. 2 a schematic illustration of a so-called “adversarial patch training”; Fig. 3, Fig. 4 a schematic illustration to illustrate a so-called “adversarial training process”; Fig. 5 an illustration of the present invention using a traffic sign; and Fig. 6 a flowchart of a method according to the invention according to an embodiment.

[0083] Fig. Figure 1 shows a schematic representation of a vehicle 1 with an environmental data evaluation device 10 according to the invention according to one embodiment. Reference numeral 20 denotes an environmental detection device of the vehicle 1, which is suitable and intended for detecting environmental data, preferably camera images.

[0084] Based on the evaluation of these recorded environmental data by means of the, in particular processor-based, environmental data evaluation device 10 using a trained vehicle environmental data evaluation model of machine learning, a vehicle function, such as a driving function (such as lateral and / or longitudinal guidance), of the vehicle is preferably carried out.

[0085] The reference numerals 12 identify external vehicles, for example of a swarm of vehicles, which preferably each record external environment data (for example camera images) with their own environment recording devices and process and evaluate these, in particular by means of a vehicle environment data evaluation model of machine learning.

[0086] Reference numeral 14 denotes an external storage device or an external server. If, for example, the external vehicles 12 determine that an evaluation result obtained using the vehicle environment data evaluation model is incorrect, such as an incorrect category being assigned or an object to be recognized not being identified or not being identified correctly, the external vehicle 12 transmits the associated external environment data to the external storage device 14, preferably together with the incorrectly assigned annotation data (label, category, or the like) and particularly preferably together with corrected annotation data (label, category, or the like).

[0087] Such transmission of the external environment data to the external storage device for potential use as training data for the vehicle environment data evaluation model of the vehicle is marked with an arrow P.

[0088] When creating the training data set for training the machine learning vehicle environment data evaluation model to be used for the vehicle 1, these external environment data are preferably retrieved from the external storage device 14 and taken into account, for example used as training data, preferably together with the incorrectly assigned annotation data (label, category or the like) and particularly preferably together with corrected annotation data.

[0089] Fig. 2 shows a schematic illustration of a so-called “adversarial patch” training process using the example of an image classification network 52, which recognizes images of a dog (“Dog”), bird (“Bird”), cat (“Cat”) or deer (“deer”) depicted in image data and can determine and / or output the respective confidence (shown in diagram 54).

[0090] Reference numeral 40 denotes exemplary image data 40, which are present in unaltered form. The processor device 44 can, for example, determine a relative arrangement of a pattern (“patch”) 46 (which can also be determined by the processor device 44) depending on the relative arrangement 42 of the object depicted in the image data 40 (here, a dog) and modify the unaltered image data 40 accordingly, so that modified image data 50 is obtained in which the pattern is integrated. This is fed to the (artificial) neural network (for example, the vehicle environment data evaluation model) as an input variable, and the output variable(s) are determined (for example, the respective confidence values ​​for dog (“Dog”), bird (“Bird”), cat (“Cat”), and deer (“Deer”)). The image data can be modified until the confidence values ​​have changed in a predetermined manner.The changes to this image data can then be checked and compared with the original and / or unchanged image data for patterns or features, so that it can be preferentially determined which feature or patterns have led to a change in the evaluation result (here confidence values).

[0091] By using adversarial training, attacks on image classification networks can be mitigated by adversarial patches. Unlike imperceptible attacks, adversarial patches are clearly visible, connected perturbations that cover a small area of ​​an image and cause the network to misclassify the image. They are easier to handle in practice because the patch can be printed as a sticker and attached to an object or camera lens, causing subsequent automatic analysis of the image to fail. One example where such an attack could be critical is the classification of stop signs by autonomous vehicles.

[0092] Adversarial training improves a network's robustness against adverse examples by training it to classify them correctly. The network's training process then preferably consists of alternating between training the network to classify against strong adversarial examples and generating strong adversarial examples that can fool the current iteration of the network. Preferably, a stronger attack is used to generate adversarial examples, which would result in a more robust network.

[0093] To defend against adversary patches, patches that cause a strong attack are preferentially identified (e.g., by determining and comparing the evaluation result) and selected for the training process, and then adversary training is carried out against them.

[0094] Image-specific: For each image, a separate patch is preferably generated, which is more effective against the respective image than a universal patch that is suitable for all images of this class.

[0095] Untargeted: The patch only needs to cause the network to misclassify the image, but does not require it to classify the image as a specific alternative class. This is done by optimizing patch values ​​that maximize the network's classification loss for each image.

[0096] Location optimization: The patch values ​​and relative position are optimized together. Prioritizing the relative position of the changes or patches is used to find the optimal image position for placing the patch that will result in the strongest attack.

[0097] Evaluations of experiments have shown: Varying the location of the changes or patch makes the attack more powerful.

[0098] Training against adversarial patches improves the robustness of the network. The robustness of the network is further improved by training against site-optimized adversarial patches.

[0099] The accuracy of the network does not decrease compared to non-adversarial examples.

[0100] The visualization of successful attacks suggests that adversarial training reduces the area on an image where an adversarial patch attack can succeed.

[0101] The Fig. 3 shows a schematic illustration of a change in original image data 56 by adding noisy data 58 with adversarial disturbance (with weighting of the data to each other, factor 0.01) and the obtained modified image data 60. By adding this noisy data 58, an object recognition of a "panda" with 81.97% confidence can change to the "Papillon dog" recognized in the modified image data 60 with 99.56% confidence.

[0102] Fig. Figure 4 shows a ResNet feature visualization before and after adversarial training. For example, adversarial training ensures that in the case of the “horse + mixtures” depicted in data 62 (see arrow P1), only a pure “horse” is recognized in the data after training (see arrow P1). Furthermore, adversarial training ensures that in the case of the “car + mixtures” depicted in data 62 (see arrow P2), only a pure “car” is recognized in the data after training (see arrow P2).

[0103] Fig. 5 shows an illustration of the present invention using a traffic sign, here a stop sign.

[0104] Is: In AI-based methods for recognizing traffic signs, for example, errors occur when objects from the real world – which are barely or not at all perceptible to the human eye – are distorted.

[0105] Example: If a small sticker with colorful pixels is applied to a stop sign, the vehicle driver will no longer recognize this sign correctly and may therefore make a serious driving error, even leading to an accident with property damage or personal injury.

[0106] Idea or target: Vehicle AI in autonomous vehicles should be trained in such a way that errors due to consciously or unconsciously introduced image or object changes are detected and the image recognition also delivers the correct result

[0107] Implementation: Training the image recognition class by presenting unmodified and falsified images or objects and naming the correct object in each case and thus learning it - (e.g. the stop sign falsified with a sticker should also be recognized as a stop sign).

[0108] Adversarial Patch Training: - The AI ​​learns to specifically hide irrelevant parts of the image or object and to focus on the essential features that define the image or object (e.g. the shape and color of the sign and the inscription STOP). - Color: If more than 98% of the sign surface is red, the remaining "colored" 2% should be ignored. The same applies to the white color of the STOP lettering.

[0109] When learning a single element (e.g. stop sign), the AI ​​calculates millions of stop signs from the one target image to be learned (and labeled as “stop sign”), essentially variants with a wide variety of errors or interference information, e.g. the addition of noise.

[0110] Safety loop: If the AI ​​is unable to identify a specific image or object, it reports the error, labels the raw data, and sends it to a human. This human then adds the image information to the entire set of objects to be learned, as well as adding corrupted variants for the learning process. If deviations are detected, the human user can also inform or train the AI.

[0111] Advantages: The learning of the objects can be done completely virtually (disturbance information is generated and inserted using an algorithm with random aspects) and thus does not require expensive additional data labeled by humans.

[0112] In addition, real-world patterns can be imported as needed, similar to a virus signature in virus scanners, thus improving the AI ​​of the entire fleet almost instantly. (For example, people with special interests have designed a new type of sticker with a disturbing image that they specifically stick on stop signs.)

[0113] Fig. 6 shows a flowchart of a method according to the invention according to one embodiment. Reference numeral 21 denotes the beginning of the method, in which deep learning is initiated. Reference numeral 22 indicates that a labeled image is present. Reference numeral 24 denotes the method step in which the algorithm (for example, the machine learning vehicle environment data evaluation model) learns the specific image and modifies the input image.

[0114] In step 26, a check is made to determine whether the image is still correctly recognized. If this is the case ("yes"), the next step, preferably the next, checks whether all current interference information has been considered.

[0115] If this is the case ("yes"), a check is carried out in a step 30, preferably the next step, to determine whether the error messages from the field are present. These are, for example, error messages detected by other vehicles. If this is the case ("yes"), the new error messages are taken into account in step 40 (algorithm training process).

[0116] If this is not the case (“no”) in step 30, i.e. there are no error messages from the field, the newly acquired test data is saved (step 32).

[0117] If the image is not recognized correctly in step 26 (“no” in step 26), a so-called “backpropagation of error” takes place in step 36.

[0118] If not all current noise information was considered in step 28 (“no” in step 28), new noise information is applied, in particular in training step 24.

[0119] Reference number 34 marks the end of deep learning.

[0120] The applicant reserves the right to claim all features disclosed in the application documents as essential to the invention, provided they are novel, individually or in combination, over the prior art. It is further noted that the individual figures also describe features that may be advantageous in and of themselves. The skilled person will immediately recognize that a specific feature described in a figure may be advantageous even without adopting further features from that figure. Furthermore, the skilled person will recognize that advantages may also arise from a combination of several features shown in individual or different figures. List of reference symbols 1 vehicle 10 Environmental data evaluation device 12 third-party vehicle 14 external storage device 16 Communication device 20 Environment detection device 21-34 procedural steps 40 Unaltered image data 42 Relative position of the imaged object 46 patterns or patches 44 Processor setup 48 Determined relative position 50 Modified image data 52 Artificial neural network 54 Confidence diagram 56 Original image data 58 noise 60 Modified image data 62, 64 image data 66 traffic signs P, P1, P2 arrows QUOTES CONTAINED IN THE DESCRIPTION

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

[0000] US 2021 300 433 A1

[0007] US 2021 397 896 A1

[0008] Cited non-patent literature

[0000] https: / / deepai.org / machine-learning-glossary-and-terms / adversarial-machine-learning; accessed on August 17, 2023

[0005] heise.de / hintergrund / Neuronale-Netze-Wie-sieangegriffen-werden-und-wie-man-sie-verteidigt-6132752.html?seite=2; accessed on August 17, 2023

[0006]

Claims

[1] Method for generating a trained vehicle environment data evaluation model of machine learning for processing environmental data acquired by means of at least one environmental detection device (20) for a vehicle (1) for carrying out an at least partially automatic vehicle function, preferably a perception function, of the vehicle (1) on the basis of the acquired environmental data, wherein the vehicle environment data evaluation model to be trained comprises a set of parameters which can be set to values ​​as a result of a training process, characterized bythat the training process of the vehicle environment data evaluation model to be trained is carried out on the basis of training data which is at least partially based on external environment data which is generated by at least one external vehicle (12) and whose evaluation by the external vehicle (12) was classified as faulty and / or critical with regard to the validity of an evaluation result and was transmitted to at least one external storage device (14) for use in the training process. [2] Method according to claim 1, characterized by that before the start of the training process, the at least one external storage device is automatically queried and external environment data transmitted to the at least one external storage device are taken into account when creating the training data intended for the training process. [3] Method according to one of the preceding claims, characterized bythat on the basis of the external environment data transmitted by at least one external vehicle, it is assessed whether the training process and / or a further training process is initiated. [4] Method according to one of the preceding claims, characterized by that the at least one external storage device is checked repeatedly and / or at regular intervals for the presence of external environment data transmitted by the at least one external vehicle. [5] Method according to one of the preceding claims, characterized by that a plurality of third-party vehicles transmits external environment data generated by the respective third-party vehicles, the evaluation of which was classified as faulty and / or critical by the respective third-party vehicle, to the at least one external storage device. [6] Method according to one of the preceding claims, characterized bythat the at least one external vehicle transmits annotation data associated with the external environment data, which were determined during the evaluation of the external environment data, to the at least one external storage device (14). [7] Method according to one of the preceding claims, characterized by that in the event of a classification of external environment data as a result of their evaluation as faulty and / or critical with regard to the validity of an evaluation result, the at least one external vehicle stores the external environment data on an internal storage device and these are transmitted to the external storage device as part of a repair and / or inspection operation of the vehicle, for example carried out by a workshop. [8] Method according to one of the preceding claims, characterized bythat the training process is carried out on the basis of training data which is at least partially based on external environment data which can be mapped to an evaluation result by at least one vehicle environment data evaluation model of machine learning only with a probability which is lower than a predetermined probability value. [9] Environmental data evaluation device (10) for a vehicle (1) with at least one environmental detection device for detecting environmental data for carrying out an at least partially automatic vehicle function, preferably a perception function, of the vehicle on the basis of the detected environmental data, wherein the environmental data evaluation device (10) is suitable and intended to receive the environmental data detected by the environmental detection device (20) of the vehicle (1) and to process them with the aid of a trained vehicle environmental data evaluation model of machine learning, wherein the trained vehicle environmental data evaluation model comprises a set of parameters which are set to values ​​which were learned as a result of a training process, characterized by that the trained vehicle environment data evaluation model is generated according to one of the preceding claims. [10] Vehicle (1), in particular motor vehicle, comprising an environmental data evaluation device according to the preceding claim.

Images