Procedures for classifying critical driving situations, selecting data similar to the critical driving situation, and retraining the automatic system.

Abstract

Methods for identifying critical driving situations when operating an autonomous vehicle (AF), selecting similar data and retraining an automated driving system (MLA) of the autonomous vehicle (AF) with the following steps: Identification of a critical driving situation (CC) when operating the autonomous vehicle (AF), Requesting and obtaining data from a backend server (BES) that is similar to the critical driving situation (CC) by an intelligent assistant (IA), and Using similar data to train the automated driving system (MLA) of the autonomous vehicle (AF), wherein The similar data stored in the backend server (BES) is based on classified sensor data from a large number of vehicles (AF) with a semantic structure, which is determined during the operation of the vehicles (AF) and tagged with a tagging device (TG) and provided with metadata before being transmitted to the backend server (BES) for storage, and the semantic structure is determined via an ontology (ON).

Description

[0001] The invention relates to a method for classifying traffic situations with malfunctions, selecting similar data and retraining the automatic system, as well as a corresponding device.

[0002] Both in the development and description of functionalities in automated driving, such as training a driving function, annotating the training data, describing the scope of functions and system behavior, and conceptually selecting training jobs for machine-learned models, there is a lack of a clear, logically consistent vocabulary that enables automatic and machine-understandable handling of individual modules or entire systems within the functionalities of automated driving, such as training, redundancy resolution, further development of existing approaches, and monitoring of complex systems consisting of various modules.

[0003] Current approaches to developing and describing functionalities in automated driving separate individual functionalities, such as sensors, perception, understanding, prediction, planning, and actuators, as well as tasks—that is, training, monitoring, further development, and optimization—into separate sub-disciplines. A holistic approach to automated driving as a whole does not currently exist with regard to formal description systems.

[0004] In the field of robotics, formal description systems and their application are used solely for describing tasks for the executing robot.

[0005] The current clear separation of tasks and application areas within sub-disciplines hinders collaboration across these boundaries and prevents uniform monitoring of the entire system. A lack of a descriptive language for scenario and context description leads to incompatible training datasets and significant manual effort in designing training jobs for machine learning modules.

[0006] Furthermore, a conceptual description of a module within an automated driving system, in terms of benign and difficult-to-handle situations, is either impossible or severely limited due to the lack of a standardized description language. Current approaches rely on manual, and therefore time-consuming, selection, analysis, and modeling, as well as the creation of "local solutions" for individual subtasks in automated driving.

[0007] In particular, this also leads to the unsustainable use of information available in a vehicle. Information is generated multiple times and / or not reused.

[0008] The German patent application DE 10 2016 009 655 A1 relates to a method for operating a vehicle in which the vehicle is controlled in a semi-autonomous or autonomous driving mode based on decisions made by an artificial neural network structure. The decisions are proposed by a static first neural network and a learning-capable second neural network, and if a decision proposed by the first neural network and a decision proposed by the second neural network coincide, this joint decision is used to control the vehicle.If a decision proposed by the first neural network and one proposed by the second neural network differ, and if the execution of the second decision fulfills predefined safety and / or ethical criteria to a greater degree than the decision proposed by the first neural network, then the decision proposed by the second neural network is used to control the vehicle. In all other cases, the decision proposed by the first neural network is used to control the vehicle.

[0009] The publication EP 2 881 829 A2 relates to a method for steering a vehicle from its current position to the vicinity of a target position. This method comprises determining the driving risk of multiple positions in the vehicle's vicinity at a given time and at several subsequent times, and determining a trajectory for the vehicle that connects or approximately connects the current position and the target position, taking into account calculated driving risks, such as those based on collision probabilities or traffic regulations, as well as vehicle dynamics and comfort parameters. The vehicle is then steered along the modified trajectory.

[0010] German patent application DE 10 2019 127 229 A1 relates to systems, devices, and methods for identifying objects and scenarios that have not been learned or that cannot be identified by vehicle perception sensors or driver assistance systems. Embodiments are designed to use a learned vehicle dataset to identify target objects in vehicle sensor data. In one embodiment, a process is provided that involves performing a scene detection operation on the vehicle to derive a vector of target object attributes from the vehicle sensor data and to generate a vector representation of the scene detection operation and the vehicle sensor data attributes. The vector representation is compared to a familiarity vector to demonstrate the effectiveness of the scene detection operation.Furthermore, the vector representation can be evaluated to identify one or more target objects or significant scenarios, including scenarios with unidentifiable objects and / or driving scenes, for the report.

[0011] Generally, when developing an intelligent agent for maneuver classification in self-driving vehicles, there are two distinct situations to consider. First, there are so-called prototypical situations where the vehicle behaves as expected. Second, there are so-called critical situations where the vehicle behaves unexpectedly. Unfortunately, the universe of possible situations in self-driving vehicles is virtually infinite.

[0012] Suppose a critical situation arises while driving an autonomous vehicle. To train the intelligent agent, i.e., the driver assistance system that drives the vehicle automatically, relevant training data is needed. The question then arises: how can such training data be generated to train the machine learning algorithm of the intelligent agent?

[0013] The invention is based on the objective of automatically annotating data generated in the vehicle during autonomous driving in order to produce a selection of training data for training the automatic driving system, with which the automatic driving system behaves appropriately in critical situations.

[0014] This problem is solved by a method having the features of claim 1 and by a corresponding device having the features of claim 7. Preferred embodiments of the invention are the subject of the dependent claims.

[0015] The inventive method for identifying critical driving situations when operating an autonomous vehicle, selecting similar data and retraining an automatic driving system of the autonomous vehicle comprises the following steps: - Identification of a critical driving situation when operating the autonomous vehicle, - Requesting and obtaining data from a backend server, similar to the critical driving situation, by an intelligent assistant, and - Using similar data to train the autonomous vehicle's automatic driving system.

[0016] In this way, it is easy to retrain an automatic driving system using similar data when a critical driving situation occurs.

[0017] The similar data stored in the backend server is based on classified sensor data from a large number of vehicles, which are semantically structured and collected during vehicle operation. This data is then tagged with metadata using a tagging device and transmitted to the backend server for storage, with the semantic structure being mediated via an ontology.

[0018] Preferably, the classification of the sensor data is based on a classification of use cases, in particular on a hierarchical classification of the use cases maneuver type, environment, road type, weather conditions and / or vehicle type.

[0019] Furthermore, the classification of use cases is preferably based on the evaluation of sensor data from a large number of vehicles. The more data available in the backend server, the higher the probability of finding data similar to critical driving situations for training the autonomous driving system.

[0020] Preferably, the ontology is based on a hierarchical classification of the use cases, in particular on the use cases already mentioned above such as maneuver type, environment, road type, weather conditions and / or vehicle type.

[0021] Preferably, in the case of a critical driving situation, the request and procurement of data similar to the critical driving situation is based on the semantic structure of the ontology, so that a search for similar data in the backend server is possible.

[0022] Furthermore, after training, the automated driving system is checked with the data stream supplied by the backend server to determine whether it has successfully handled the critical driving situation that triggered the retraining. In other words, it is checked whether the autonomous driving system now reacts appropriately to the critical driving situation that occurred.

[0023] The device according to the invention for identifying critical driving situations when operating an autonomous vehicle, selecting similar data and retraining an automatic driving system of the autonomous vehicle, wherein the device is set up and designed to carry out the method described above, comprises - a driving system arranged in the autonomous vehicle, - a large number of sensors arranged in the autonomous vehicle to detect the vehicle's surroundings and vehicle data, including position data, - a device for identifying critical driving situations, - an intelligent assistant for requesting and obtaining data similar to critical driving situations based on the semantic structure of an ontology, and - a facility for retraining the automatic driving system.

[0024] A preferred embodiment of the invention is explained below with reference to the drawings. The drawings show... Fig. 1. The autonomous vehicle in the universe of possible situations, Fig. 2 the conventional approach, Fig. 3. An overview of the proposed approach, Fig. 4 the proposed approach in greater detail, Fig. 5 the individual procedural steps for implementing the ontology, Fig. 6. the identification and classification of maneuver types, and Fig. 7. The architecture of the Tagging and Search Software.

[0025] Fig. Figure 1 schematically depicts an autonomous vehicle AF and its interaction with its real-world environment RU. The autonomous vehicle AF is equipped with a multitude of sensors SE, whereby the measurement results of the sensors SE can influence the autonomous vehicle AF, and conversely, the autonomous vehicle AF can influence the sensors SE. The real-world environment RU comprises a universe of possible driving situations UPS with respect to the behavior of the autonomous vehicle AF, the number of which is unknown. Within the universe UPS of possible driving situations, a distinction is made between two types: prototypical situations PS and critical situations CS, which together constitute the set of possible driving situations S with respect to the autonomous vehicle AF.In other words, the autonomous vehicle AF operates within its set of possible driving situations S, where prototypical situations are those with which the machine learning algorithm MLA has been trained and is therefore familiar. In contrast, critical situations CS are those driving situations that are unknown to the machine learning algorithm MLA and to which it cannot react adequately. This can potentially lead to malfunctions in the machine learning algorithm MLA, causing the autonomous vehicle AF, which is controlled by the machine learning algorithm MLA, to react incorrectly to the critical driving situation. In other words, the autonomous vehicle AF relies on the machine learning algorithm MLA and is controlled by it.

[0026] Based on the requirements of the autonomous vehicle (AF), the sensors (SE) measure data (D), which can be stored in a database (DB) on a backend server (not shown). In other words, the database (DB) comprises a set of diverse data (D), which can be used to train the machine learning algorithm (MLA). Conversely, the MLA requires such data (D). Furthermore, the data (D) describe the real environment (RU) perceived by the sensors (SE). The interaction between the real environment (RU) and the universe of possible situations (UPS) consists of the fact that the set of situations (S) affecting the autonomous vehicle (AF) is, on the one hand, a component of the real environment (RU), and on the other hand, the real environment (RU) contains a multitude of such situations (S) affecting the autonomous vehicle (AF).

[0027] In Fig. Figure 2 schematically depicts the conventional procedure when a critical situation arises. In step A1, a critical situation is identified based on its occurrence. Following the identification of a critical situation in step A1, the subsequent step A2 attempts to gather data regarding the identified critical situation. This data from step A2 is fed into the corresponding networks (not shown) in the subsequent step A3, and step A4 expresses the expectation that the data from step A2 has captured the critical situation. To the right of this sequence of steps, it is shown that the gathered data D is fed into the machine learning algorithm MLA for training purposes, so that the operation of the autonomous vehicle AF can be improved based on the newly trained critical situation.

[0028] The advantages of the conventional approach when a critical situation arises are: - easy access to the problem, - no need to view the data in a conceptual context, - Many unforeseen situations can be addressed with minimal effort and reasonable accuracy, and - The more data is available, the better the network performs.

[0029] The disadvantages can be listed as follows: - extensive computational requirements, - the required amount of data is large and difficult to classify, - Functionality cannot be guaranteed for every possible situation, and - there is a large storage requirement.

[0030] Fig. Figure 3 shows the procedure chosen here. On the left side of the Fig. Figure 3 shows the schematic process flow. In the first step, B1, a critical situation is identified. In the subsequent step, B2, semantically similar data regarding the critical situation are determined, which are then used in step B3 to train the machine learning algorithm.

[0031] The following is a brief example of a critical or disadvantageous situation, which looks like this: A vehicle is traveling in the middle lane of a three-lane road, while a vehicle is stationary in the right lane. People, such as occupants, are outside the stationary vehicle in the right lane. Since the left lane of the vehicle traveling in the middle lane is empty, a driver acting with foresight would move into that left lane to maintain sufficient distance from the people in the right lane.

[0032] It has now emerged that an autonomous vehicle typically does not change to the left lane, but stubbornly remains in its middle lane, meaning that endangering people in the right lane cannot be ruled out. This can be considered a so-called corner case, since a proactive driver would change lanes, which is not usually the case with an autonomous vehicle.

[0033] In general, corner cases in the context of autonomous driving can be defined as a situation in which a module within the automatic driving system experiences an unknown situation with a functional deficiency.

[0034] The right side of the Fig. Figure 3 illustrates the procedure described above in greater detail. An autonomous vehicle (AF) identifies an adverse situation (USI), as described in the example above. An adverse situation (USI) is not necessarily a critical situation, but it can be. Due to the adverse situation (USI), a request for similar data is sent to an intelligent agent (IA). In other words, because of the adverse situation (USI), an intelligent agent (IA) is tasked with finding and retrieving the required data from a backend server (BES). The backend server (BES) contains data (D) that has a semantic structure (SMS).Such data D retrieved by the intelligent assistant IA are fed to the machine learning algorithm MLA for training with data similar to the disadvantageous situation, which in turn influences the driving style of the autonomous vehicle AF, so that the functionality of the machine learning algorithm MLA, extended by the retraining, can be checked.

[0035] The advantages of this approach are as follows: - the data have a semantic structure that allows for inferences, - Relevant data can be derived from the database based on these borderline cases or unfavorable situations, - expanded concept with the possibility of understanding generic situations, - computationally simple data searches, provided the data tagging was carried out effectively, and - Data comes from various sensors, which enables robust reasoning and inference.

[0036] The disadvantages of the procedure are: - complicated approach regarding extension and generalization, - the efficiency of the procedure depends on the depth of knowledge in the ontology and the specified relations, - Many classifiers are difficult to classify as a function of the use case, and - Creating a flawless ontology that reasones in analogy to humans in all possible situations is difficult.

[0037] Fig. 4 shows this in Fig. 3. Proposed approach for finding relevant data to manage unforeseen or critical situations. An autonomous vehicle AF is depicted, moving within a real vehicle environment RU, for example, on a roadway, where the real vehicle environment RU is an integral part of the environment UW. During the operation of the autonomous vehicle AF, unforeseen behavior UV occurs, which in Fig. Figure 4 is schematically represented by the arrow. This unforeseen behavior UV of the autonomous vehicle AF represents a vulnerability CC of the trained model TM operating the autonomous vehicle AF and is realized in the example above by maintaining the middle lane, thereby endangering pedestrians standing in the right lane. In other words, vulnerability CC is a so-called corner case and, in extreme cases, a critical situation. As a consequence of the vulnerability CC, the next step NSD initiates a search for similar data; that is, the intelligent agent IA is tasked by step NSD with finding similar data corresponding to the unforeseen behavior UV.

[0038] To this end, the intelligent assistant IA searches a backend server BES for data similar to that of the encountered corner case CC. The data to be found does not necessarily have to describe the identical situation, but should be sufficiently similar to the encountered corner case CC. Sensor data from autonomous vehicles AF, along with assigned metadata MD, are transmitted to the database of the backend server BES. Furthermore, annotations and classifiers, which convey semantic meaning, are added to the raw data in the database of the backend server BES by an annotation instance AA. This means that the classifiers exhibit a kind of hierarchy and relationships among themselves, which encode human knowledge in a suitable form, thus enabling inferences. The raw data of the autonomous vehicles AF, thus enriched with metadata, are therefore...Tags were added, allowing for the evaluation or classification of the data and enabling inferences. Consequently, the database of the BES backend server contains a wealth of specific annotated driving situations.

[0039] Based on the annotated data stored in the backend server's database, the intelligent assistant IA can search for data similar to the encountered corner case CC. This relevant data (RD) is added to the trained model TM for training purposes, in order to further train and improve the already trained model TM, which corresponds to the machine learning algorithm. Using the improved trained model TM, the driving behavior of the autonomous vehicle AF can be adapted to the detected corner case CC.

[0040] The advantages of this approach are as follows: - the annotated data have a semantic structure that allows for interference, - Relevant data regarding a corner case can be determined, - The concept can be extended by adding further ontologies to gain an understanding of a semantic scene, - the approach requires little computing power as long as the tagging of the raw data is carried out effectively, and - Data from a variety of different sensors are used, which are fed in via appropriate interfaces, so that robust conclusions and inferences are possible within the available data.

[0041] The disadvantages of this approach are as follows: - Complicated approach regarding extension and generalization to further cases, - The efficiency of the approach is a function of the depth of the ontology, - Some classifiers have difficulties classifying as a function of the use case, - Creating a flawless ontology that handles use cases and inferences in a way comparable to a human is difficult.

[0042] In the Fig. Figure 5 shows the essential steps C1 to C8, which make it possible to train the machine learning algorithm, i.e. the trained model for controlling an autonomous vehicle, using data similar to a Corner Case, as explained above.

[0043] In step C1, a use case is identified, and in the subsequent step C2, an initial ontology is created. Then, in step C3, classifiers are identified, allowing for the development of a concept / architectural design in step C4. Following this, in step C5, initial script interfaces or a framework structure can be created, enabling the implementation of a machine learning algorithm in step 6. This necessitates the implementation of the final ontology in step S7, allowing the resulting code to be tested and evaluated for feasibility in step C8.

[0044] Fig. Figure 6 shows in schematic form the steps in step C1 of the Fig. 5. The identification of possible use cases mentioned above is shown schematically, focusing here on the use case of maneuver type MNT. The following is... Fig. Section 6 should not be considered an exhaustive list of all use cases for maneuver type MNT, but is merely exemplary. Further use cases may relate to the vehicle environment, weather conditions, vehicle type, and road type, each of which will be classified.

[0045] The use case of maneuver type MNT can be divided into basic maneuvers (BAS) and tactical maneuvers (TAC), each with subclassifications. Regarding basic maneuvers (BAS), a distinction is made between the Acceleration maneuver (AC) with the subclassifications Normal Acceleration (NAC) and Sudden Acceleration (SAC). Furthermore, the Turning maneuver (T) with the subclassifications Left Turn (LT) and Right Turn (RT) falls under the category of basic maneuvers (BAS). Other basic maneuvers (BAS) include the Lane Change (LC) with the distinctions Left Lane Change (LLC) and Right Lane Change (RLC). Then there is the category Relaxed Driving (CR) and Acceleration (SP). Similarly, there is the Deceleration category (DC) with the subclassifications Deceleration with Brake (WB) and Deceleration without Brake (WOB), whereby the with-brake category (WB) is further distinguished into Gentle Braking (SB) and Emergency Braking (EB).Finally, the classification "Reversing RV" also falls under the application case "basic maneuver BAS".

[0046] Regarding the area of ​​tactical maneuvers (TAC), a distinction is made between the maneuver "Stopping" (ST) with the specifications "Parking" (P), "Stopping at a Traffic Light" (STL), "Giving Way to Pedestrians" (YP), and "Standing Still for Other Reasons" (SOR). Further tactical maneuvers include "Crossing an Intersection" (DTI), "Overtaking" (OT), and "Turning the Vehicle" (UT). As already mentioned, this list is not exhaustive. Fig. The 6 described maneuvers are presented with their subclassifications.

[0047] Fig. Figure 7 shows a schematic representation of the software architecture, which, on the one hand, uses tagging software (TG) to classify, tag, and structure incoming, real-time sensor data so that it is accessible in a backend server (BES) for queries by data-querying software (DFS). As can be seen from the Fig. As can be seen in section 7, the two structures, tagging software TG and data querying software DFS, are arranged separately on top of each other for the sake of clarity.

[0048] The tagging software TG comprises a machine learning algorithm MLA, which is trained using training data TD. The training data TD consists of pre-recorded, labeled sensor data SDL, which was labeled either manually or via an automated process. In other words, the recorded, labeled sensor data SDL is stored in a saved format, for example, on a suitable storage medium. To train the machine learning algorithm MLA, it is provided with the hierarchical structure of the corresponding ontology ON, so that the machine learning algorithm MLA can be trained according to the given ontology O.Autonomous vehicles (AF) moving within their environment, particularly in road traffic, transmit sensor data (D) in real time to the tagging software (TG). This real-time data (D) can be supplemented by pre-recorded, unlabeled sensor data (SDUL), which represents recorded but not directly processed sensor data from autonomous vehicles (AF). This data (D) and SDUL are aggregated in a real-time sensor data unit (SDRT), for example, into a CSV file for further processing, and fed to a trained machine learning algorithm (TMLA). TMLA has adopted the learned parameters from the machine learning algorithm (MLA) and applies appropriate tags to the data. The data tagged by the trained machine learning algorithm (TMLA) is then fed to the metadata instance (MD), where it is supplemented with map data and other classifiers.The tagged data of the metadata instance MD are summarized in the next instance into a structured data stream SDS, to which the original real-time sensor data SDRT are also added, so that this data package is stored in a backend server BES.

[0049] The lower part of the Fig.The data query software DFS, as depicted in Figure 7, includes an ontology ON, which is required by a knowledge base KB. Furthermore, an input query IQ is posed, for example, when an autonomous vehicle requests data regarding an overtaking maneuver. This query is directed to an intelligent agent IA, which accesses both the knowledge base KB and the data of the backend server BES. The backend server BES sends a structured data set SDS to the intelligent agent IA, which then creates an output data stream OUT that fulfills the requested semantics, in this case, the query concerning an overtaking maneuver. Reference symbol list AF autonomous vehicle MLA machine learning algorithm SE Sensors RU Real Vehicle Environment UPS Universe of possible situations D data DB database PS prototypical situation CS critical situation S situation A1 Identification of critical situation(s) A2 Providing additional data A3 feeding the additional data into the networks A4 Hope for determining the critical situation B1 Identification of critical situation(s) B2 Request semantically specific data B3 Using this data for training SMS semantic structure IA Intelligent Agent USI Identification of an unusual situation UW Environment TM Trained Model UV unexpected behavior CC Corner Case / Limit MD Meta Data BES Backend Server NSD needs to find similar data AA Addendum Annotations RD Sending relevant data C1 Identification of a use case C2 Providing an initial ontology C3 Identification of classifiers C4 Concept / Architectural Design C5 Generation of initial interfaces / framework C6 Implementation of the machine learning algorithm C7 Implementation of the Ontology C8 Testing and evaluating the code with regard to feasibility. MNT maneuver type BAS Basic Maneuvers TAC Tactical Maneuvers AC Acceleration NAC Normal Acceleration SAC Sudden Acceleration T Turn LT Turn left RT Turn right LC Lane Change LLC lane change to the left RLC lane change to the right CR smooth driving (cruising) SP Acceleration (Speeding) DC Decelerating WB with brake SB soft braking EB emergency brakes WOB without brakes RV Wenden (Reversing) ST Stop P Parking STL Stop at Light Signal YP Right of way over pedestrians SOR standstill for other reasons DTI Driving across an intersection Overtaking on the off-road UT U-turn TG Tagging Software SDL recorded labeled sensor data SDUL recorded unlabeled sensor data ON Ontology IQ Input Query TD Training Data HS Hierarchical Structure SDRT sensor data in real time MLA machine learning algorithm TMLA-trained machine learning algorithm CL cards and other classifiers MD Metadata SDS structured data stream DFS Data Retrieval Software KB Knowledge Base IA Intelligent Agent OUT Output of the found data

Claims

[1] Method for identifying critical driving situations when operating an autonomous vehicle (AF), selecting similar data and retraining an automated driving system (MLA) of the autonomous vehicle (AF) comprising the steps: Identification of a critical driving situation (CC) when operating the autonomous vehicle (AF), Requesting and obtaining data from a backend server (BES) that is similar to the critical driving situation (CC) by an intelligent assistant (IA), and Using similar data to train the automated driving system (MLA) of the autonomous vehicle (AF), wherein The similar data stored in the backend server (BES) is based on classified sensor data from a large number of vehicles (AF) with a semantic structure, which is determined during the operation of the vehicles (AF) and tagged with a tagging device (TG) and provided with metadata before being transmitted to the backend server (BES) for storage, and the semantic structure is determined via an ontology (ON). [2] Method according to claim 1, characterized by that the classification of sensor data is based on a classification of use cases (MNT), in particular on a hierarchical classification of use cases such as maneuver type, environment, road type, weather conditions and / or vehicle type. [3] Method according to claim 2, characterized by that the classification of use cases (MNT) is based on the evaluation of sensor data from the multitude of vehicles (AF). [4] Method according to any one of claims 1 to 3, characterized by , that the ontology (O) is based on a hierarchical classification of use cases (MNT). [5] Method according to any one of claims 1 to 4, characterized by , that in the case of a critical driving situation (CC), the request and procurement of data similar to the critical driving situation is based on the semantic structure of the ontology (O), so that a search for similar data in the backend server (BES) is possible. [6] Method according to claim 5, characterized by , that the automatic driving system, after training, is checked with the data stream supplied by the backend server (BES) to determine whether the critical driving situation that triggered the retraining has been successfully handled. [7] Device for identifying critical driving situations when operating an autonomous vehicle (AF), selecting similar data and retraining an automated driving system (MLA) of the autonomous vehicle (AM), wherein the device is set up and designed to carry out the method according to one of the preceding claims, with a driving system arranged in the autonomous vehicle (AF), a large number of sensors (SE) arranged in the autonomous vehicle (AF) for detecting the vehicle environment and vehicle data, including position data, a facility (USI) for identifying critical driving situations, an Intelligent Assistant (IA) for requesting and obtaining data similar to critical driving situations based on the semantic structure of an ontology (ON), and a facility for retraining the automatic driving system.

Images