Automatic detection and assessment of low-speed crashes

Abstract

Method for detecting a vehicle accident in which a vehicle (1) and an object (10) collide, wherein a motion variable associated with the impact is so low that at least one active occupant protection system (39) of the vehicle (1) designed for accidents is not activated, and in which the impact event - signals and / or data generated by sensors (3) of the vehicle (1) are processed in such a way that - the signals and / or data are filtered, - feature data is generated based on the filtered signals and / or data, - using the feature data, the impact event is assigned to a class index of a classification and -- where feature vectors are formed from the feature data, -- wherein a parameterizable classification unit (8) is used which comprises several classification processors (18a, 18b, 18c) which each execute different classification procedures learned in different machine learning methods (ANN, HMM, SVM), -- wherein the feature vectors are processed according to a parameterization of the classification unit (8) based on the respective classification method (HMM, SVM, ANN) in at least one of the classification processors (18a, 18b, 18c) and the results of the respective outputs by the classification processors (18a, 18b, 18c) are statistically examined to form an overall result, -- where, for the overall result, a class index is assigned based on the most probable event class for the underlying event as assigned by the respective classification procedures and / or a class probability of a currently evaluated feature vector is determined and -- wherein control and / or regulation data belonging to the respective classification methods (HMM, SVM, ANN) are used for parameterization, which are stored in a classification database (9).

Description

[0001] The invention relates to a method, a system, and a computer program for the automatic detection and evaluation of low-speed crashes. The invention relates in particular to the detection of vehicle accidents at relatively low speeds, and also to an evaluation of the accident, especially to determine the resulting damage to a vehicle. The invention also relates to a correspondingly equipped vehicle.

[0002] Modern motor vehicles are equipped with a multitude of sensors that can detect accidents and, if necessary, automatically trigger active or passive safety systems such as airbags, seatbelt pretensioners, emergency call systems, etc. However, these detection systems are generally designed to only recognize accidents below a predetermined lower detection threshold, for example, a certain minimum vehicle speed and / or minimum acceleration.

[0003] For example, DE 41 17 811 A1 discloses a method for crash detection in which the change in vehicle speed over a limited period is continuously measured and evaluated as a so-called partial speed difference. Acceleration values ​​can also be calculated and evaluated using threshold values ​​to distinguish between high-speed and low-speed crashes.

[0004] In DE 10 2007 027 649 A1, a control unit and a method for controlling personal protective equipment are proposed, in which a feature vector with at least two features from at least one signal is used to form an accident sensor. The feature vector is classified by a Support Vector Machine (SVM).

[0005] In German patent application DE 10 2008 003 081 A1, a method and a device for crash classification of a vehicle safety system were described, in which a currently determined driving situation is classified, based on a machine learning process, according to whether a personal protection device should be activated. For this purpose, crash classes are defined by predefined crash characteristics.

[0006] From EP 2 854 112 A1 a detection system for recognizing accidents at low speeds or relatively low impact is known, in which acceleration values ​​recorded along several axes are each filtered by means of bandpass filters and the filtered values ​​are used to determine the severity of the impact.

[0007] WO 90 / 11207 A1 describes a control system for occupant restraint and protection systems in vehicles, in which signals from an acceleration sensor are subjected to several different evaluations. The signal progression is evaluated according to different criteria and with several differently defined thresholds. Based on the temporal progression and the respective threshold, two different types of accidents, such as head-on and oblique impacts, can be distinguished. Differently defined progressions and thresholds for the severity of the accident can also be provided.

[0008] US Patent 5,684,701 A describes a sensor system for detecting accidents and triggering occupant protection systems in vehicles, in which the sensor signals are evaluated using self-learning methods based on artificial neural networks. To learn a sensor algorithm, sensor data from accidents obtained in artificially induced crash tests is used, along with corresponding sensor data not derived from accidents.

[0009] From US patent 7,113,079 B2, a method and a system for detecting vehicle collisions are known, in which a Hidden Markov Model (HMM) is used to evaluate sensor parameters.

[0010] German patent DE 10 2014 013 472 A1 describes a method for detecting and classifying damage or contact events on motor vehicles, in which various forms of damage or contact events are to be detected, evaluated, and classified by recording and evaluating at least one structure-borne sound signal. An amplitude envelope of the structure-borne sound signal is determined, the structure-borne sound signal is classified based on the temporal profile of the amplitude envelope, and respective damage or contact events are assigned to different types of temporal profiles of the amplitude envelope.

[0011] German patent DE 10 2006 049 262 A1 describes a method for controlling personal protective equipment such as adaptive airbags or seatbelt pretensioners, in which a sequence of feature vectors is generated, each containing at least two features from at least one signal from an accident sensor. The sequence is classified according to the probability with which the at least one sequence belongs to a particular class, and the personal protective equipment is controlled accordingly. From stored sequences, the sequence that most likely corresponds to the generated feature vector sequence is selected. This sequence then determines the classification.

[0012] The contents of the above-mentioned publications are hereby incorporated into this description by reference.

[0013] As mentioned earlier, the systems described above are designed to activate active occupant protection systems. However, it would also be desirable for a vehicle to be able to autonomously detect accident events for damage assessment purposes, even when a collision occurs at relatively low speed, with relatively weak external impact (impact energy), and / or with relatively low acceleration. Such events can occur, in particular, during parking maneuvers as so-called parking bumps.

[0014] The object of the invention is to contribute to the safest possible detection of a collision between a vehicle and an object where the magnitude of the impact is relatively low.

[0015] This problem is solved by the invention specified in the independent claims. Advantageous embodiments of the invention are specified in the dependent claims.

[0016] According to the invention, to detect a vehicle accident in which a vehicle and an object collide and in which the magnitude of the impact is so low that at least one active occupant protection system of the vehicle designed for accidents is not activated, it is provided that signals and / or data generated by sensors of the vehicle relating to the impact event are processed in such a way that the signals and / or data are filtered, feature data are generated on the basis of the filtered signals and / or data, and the impact event is assigned to a classification using the feature data, whereby a classification database is used.

[0017] For the assignment, at least one classification method can be used. The classification method can, in particular, include at least one classification rule.

[0018] The movement parameter is a particularly relevant parameter for activating the vehicle's active occupant protection system. It is specifically used to control the activation process and remains a crucial parameter.

[0019] The motion parameter can be, in particular, the impact velocity, the change in velocity (acceleration) associated with the impact, especially the maximum and / or average change in velocity and / or acceleration of the vehicle (or a sensor installed in the vehicle), and / or an energy parameter associated with the impact, especially impact energy. It can also be a parameter derived from one or more of the aforementioned parameters. The impact velocity can, in particular, be the relative velocity between the vehicle and the object at the beginning of the impact, the actual velocity of the vehicle, or the actual velocity of the object, especially when the vehicle is stationary.

[0020] The sensors can generate sensor signals or data about the motion quantity, for example, the vehicle's speed, accelerations in one or more different directions, such as longitudinal and / or lateral acceleration. When detecting a one-dimensional motion quantity in one direction, it can be specifically designed so that longitudinal vibrations occurring along the axis of that direction are not detected, but only the motion quantity with respect to that one direction.

[0021] The sensors can include at least one microphone, one camera, one ultrasound sensor, one radar sensor, one lidar sensor, and / or one pressure or force sensor. Each sensor can detect a movement along a specific direction and / or derive it from the respective sensor signals. The sensor, particularly a pressure sensor, or its signals, can be used for two purposes: detecting a low-speed crash and controlling the active occupant protection system. A pressure sensor is particularly suitable for detecting a one-dimensional movement in a specific direction in such a way that vibrations occurring along the axis of that direction are not detected. In contrast, a microphone (or sound sensor) is particularly suitable for detecting such vibrations (or sound waves).

[0022] An impact severity determination module can include a crash severity determination unit, which uses predefined rules to determine and output a value for the severity of the impact. It can also include a crash direction determination unit, which uses predefined rules to determine and output a value for the impact direction relative to the vehicle's axes. Finally, it can include a crash localization determination unit, which uses predefined rules to determine the location of the impact on the vehicle.

[0023] The invention is based on the realization that classifying impact events between a vehicle and an object at low speeds, distinguishing between an accident in which the vehicle is damaged by the impact of the foreign object and a non-critical event such as driving over a pothole or emergency braking (a so-called misuse event), presents a difficult technical challenge. This is because events of these two classes are very similar or cannot be clearly distinguished using the signals typically emitted by vehicle sensors (e.g., acceleration sensors) and used, for example, to trigger occupant protection systems or automatic emergency call devices such as the eCall system, which is planned for introduction in Europe.The invention is based in particular on the realization that the application of the aforementioned accident detection systems for higher speeds with predominantly acceleration threshold-based methods for detecting less serious accidents at low relative speeds between the vehicle and the colliding object is insufficient, since the differentiation between damage events and non-invasive misuse events is only relatively unreliable.

[0024] Within the scope of the invention, a configurable signal processing system can be specified in particular for the vehicle-autonomous, automatic detection of accidents involving relatively small movements (so-called low-speed crashes). The system can be designed as a real-time system. For this purpose, a relatively low computational complexity in data processing can be provided.

[0025] The invention recognizes that by selectively filtering the respective signals or data, generating feature data associated with the impact event, representing a predetermined feature of the impact event, particularly from multiple sensor signals or data, and by processing the feature data, and especially by using a classification method, the assignment of the impact event to a classification, particularly to the two classification levels "Misuse Event" or "Low Speed ​​Crash", can be carried out with higher accuracy than with conventional methods.

[0026] In advantageous embodiments of the invention, feature data are generated from the signals or data emitted by several sensors. In particular, vector data can be generated from several feature data for an impact event, and furthermore, so-called feature vectors can be generated that represent an observed variable, especially a multidimensional one, for the impact event.

[0027] The vehicle sensors may, in particular, include sensors that are not primarily or directly intended for the purpose of recording accident data, and especially not for controlling the activation of an active occupant protection system. According to the invention, it has been recognized that at least partial features for accident detection can be implicitly derived and / or calculated from signals and / or data recorded, generated, and / or output by such sensors. Examples of such sensors are microphones for sound detection, cameras for image detection, ultrasonic, radar, lidar, or pressure / force sensors.

[0028] The invention offers the particular advantage that low-speed crash detection can be achieved using sensors already present in a vehicle, i.e., whose sensor signals are used for at least one other function in the vehicle that is independent of low-speed crash detection. Therefore, no additional or separate sensor is required for low-speed crash detection, and in particular, no separate structure-borne sound sensor is needed. Low-speed crash detection can then rely exclusively on signals already used elsewhere in the vehicle.

[0029] The signals and / or data generated by the vehicle sensors, and especially those transmitted via an asynchronous communication bus, can be advantageously synchronized in time for filtering purposes, particularly before filtering. This can be achieved, for example, by means of a uniform system time provided in the vehicle, and especially via a predefined group of electrical control units, and / or by means of timestamps. During the filtering of the signals and / or data generated by the sensors, predetermined selection signals and / or selection data can be selected and / or derived from the signals and / or data previously generated by the sensors using predefined time window data.

[0030] The invention advantageously enables automatic, vehicle-autonomous detection of an accident event at relatively low speed or with relatively weak external influences. In particular, the invention provides a system and a method for the vehicle-autonomous, automatic detection and / or evaluation of low-speed and / or low-impact crashes or vehicle collisions with low-speed crash characteristics.

[0031] The invention can be used regardless of whether the vehicle is in motion and collides with or touches a stationary obstacle, or whether the vehicle is stationary and, for example, is struck or touched by another vehicle. Such events can occur particularly in parking lot collisions. The invention allows such an event to be detected, the corresponding data to be stored in the vehicle for documentation purposes, and / or transferred to an external storage device. This allows a low-speed crash event, for example, for rental or leased vehicles, to be advantageously documented for evidentiary purposes, optionally with the additional recording and documentation of the corresponding time and location.

[0032] Based on information about a detected low-speed crash event, other control systems of the vehicle can also be advantageously controlled in a targeted manner, for example a pedestrian accident protection system with appropriate protective devices for a pedestrian hit by the vehicle at low speed.

[0033] Within the scope of the invention, it can be provided that the corresponding vehicle sensors and / or at least one control unit involved in accident detection are active regardless of whether the vehicle's ignition is switched on. In particular, it can be provided that a control unit has a bus wake-up device that wakes the vehicle data bus and / or specific sensors from a sleep state in order to perform accident detection. The wake-up device can, in particular, have at least one sensor for directly and / or indirectly detecting a movement parameter of the vehicle and a corresponding vehicle bus interface. This enables, in particular, automatic, vehicle-autonomous accident detection, for example, when the vehicle is parked.

[0034] The invention also advantageously enables the use of sensors, control units, and / or data connections, particularly data buses, already commonly provided in vehicles, and especially those already installed, to implement a method according to the invention without major modifications and, in particular, without substantial hardware changes. The invention can thus be implemented, in particular, by purely program-related, software-related adaptations of the corresponding vehicle components. For example, it may be sufficient to simply extend the non-volatile and / or volatile memory in individual vehicle components to store additional control program code, and / or to provide additional control program code or software components, extend or supplement individual interfaces, and / or make other such adaptations.A corresponding vehicle component, in particular an electronic control unit, must be equipped with appropriate resources, including appropriate computing power, to implement the invention, whereby the computing power requirements are not too high, since the processing of data for low-speed crash detection – unlike systems that activate occupant protection systems when relatively high impact-related motion quantities, in particular speeds and / or accelerations, are present – ​​does not necessarily have to be performed in real time.

[0035] The vehicle sensors are preferably located on and / or inside the vehicle. The object can be, for example, another stationary or moving object, in particular a second vehicle, a pedestrian, or an obstacle such as a fence or tree.

[0036] According to an advantageous embodiment of the invention, the at least one classification method and / or the classification database is formed using a machine learning (ML) method, in particular using a method based on an artificial neural network (ANN), a support vector machine (SVM), a hidden Markov model (HMM) and / or a decision tree.

[0037] Within the framework of the machine learning process, a database is created, particularly using so-called training datasets. These training datasets can be provided and / or obtained, especially through data recordings and / or transcripts of predefined, known accident events and / or based on the execution and corresponding data recording of, in particular, representative driving tests.

[0038] Based on the sensor data generated and recorded during the events or driving tests and / or other data generated in the vehicle and provided in particular via a vehicle data bus (vehicle communication bus) (input data) and the known classification assignment of the respective events with regard to a low-speed crash event versus a misuse event (output data), corresponding input patterns can be trained and the data of the database generated.

[0039] The recognition accuracy of a system executing the method according to the invention can be iteratively optimized or improved according to an advantageous embodiment, wherein, in particular, successive system variations can be made with regard to parameterization and implementation, for example, with regard to input signals used, calculated, in particular derived and / or extracted, features or feature vectors formed therefrom, and / or algorithms for data or signal processing, etc. Based on this, system results and / or system performance data can be iteratively evaluated, and, if necessary, the system can be retrained, with cross-validations also being possible.In particular, based on this, iterative analytical system models and / or hypotheses for optimization can be established and / or planned system variations can be carried out, which can then in turn lead to a new parameterization and / or implementation.

[0040] Within the scope of the invention, it is advantageous to determine damage to a vehicle following an accident, in particular by performing the method steps described in this document for detecting a low-speed crash, and furthermore by determining additional accident data based on the established assignment of the impact event, with which the extent of the damage is determined, in particular estimated. For this purpose, data for estimating the impact direction, the point of impact on the vehicle, and / or the impact intensity (impact energy, impact speed, and other dimensions, distances, speeds, changes in speed and / or acceleration, etc.) can be acquired, for example, by means of vehicle sensors such as at least one speed sensor, acceleration sensor, distance sensor, and / or a camera.

[0041] According to a further advantageous embodiment of the invention, the data generated by the vehicle's sensors, data derived therefrom, and / or further data generated by the vehicle's control units are transmitted via a vehicle data bus. The transmission takes place, in particular, between sensor assemblies and / or control units, and / or from sensor assemblies via potentially directly connected control units, via the respective connected data bus(s) to a respective control unit that wholly or partially implements a method according to the invention or wholly or partially comprises a system according to the invention. Received data can be processed by an evaluation unit to perform at least some of the process steps according to the invention.

[0042] In principle, any type of electronic data bus can be used as the data bus in the context of this description, in particular a data bus commonly used in the automotive sector, such as a Local Interconnect Network (LIN) data bus, a Controller Area Network (CAN) data bus, an Ethernet data bus, a FlexRay data bus, or a Media-Oriented Systems Transport (MOST) data bus. Data transmission between a sensor and the evaluation unit can also take place across multiple data buses, i.e., via a data bus system that includes, in particular, several identical, similar, or different data buses.

[0043] The invention is particularly implementable as a method. It can also be specified in the form of a system for detecting a vehicle accident in which a vehicle and an object collide, wherein a movement magnitude is so low that at least one active occupant protection system of the vehicle designed for accidents is not activated, wherein the system comprises: a) sensors intended for the vehicle which generate signals and / or data relating to the impact event and b) a data processing facility comprising a classification database, designed to process the signals and / or data such that b1) the signals and / or data are filtered, b2) feature data are formed on the basis of the filtered signals and / or data, and b3) the impact event is assigned to a classification using the characteristic data, with the classification database being used.

[0044] The signals or data can be generated, provided, and / or transmitted, particularly in the form of messages. The assignment can be carried out, in particular, using at least one classification method. The classification method can comprise at least one classification rule.

[0045] The system may have further features, in particular components that are set up to carry out process steps described in this document in connection with process steps according to the invention.

[0046] The system can be specifically designed for implementation on an electronic control unit (ECU) installed in a motor vehicle, which provides access to at least one vehicle communication bus system (CAN, Flexray, etc.).

[0047] The system features read access to the connected communication bus at its input interface. This allows for selection or restriction to message types that contain signals or data relevant for feature calculation. If the signals, data, or messages required for processing are encoded, the system requires, stores, and / or provides the corresponding information for decoding them.

[0048] On the output side, the system features, in particular, write access to the communication bus. Preferably, a separate message type can also be provided with which the detection result, especially at runtime, is output to the communication bus at regular intervals in the form of a corresponding message. It is particularly advantageous to output a corresponding message, e.g., in the form of "Low-speed crash detected," only in the case of a positively detected accident event. This makes it possible, for example, for arbitrary systems that also have access to the communication process to use the message about a detected low-speed crash output by a system according to the invention for their respective applications, to process it further, and / or to forward it to connected downstream systems.For example, it would be conceivable to display the detection result message on a user interface (Human Machine Interface, HMI), particularly a graphical one, such as a screen in the vehicle for the driver. The message, and any other data recorded in the vehicle before, during, or after the impact, can be transmitted fully automatically or, for example, initiated by the driver via a control command, to internal and / or external devices. The data can be transmitted, for example, to a mobile device such as a smartphone and / or to a so-called backend server, which is configured for data exchange with a large number of predefined vehicles. A backend server can provide so-called telematics services for vehicles, where information, for example, about a vehicle, can be transmitted to and from the vehicle.Data is collected, processed, stored, provided, and / or forwarded. In particular, components or functions of a vehicle can also be remotely controlled via a telematics service.

[0049] Furthermore, the system can have a system data interface through which system configurations can be made and / or information about the current system status can be queried. The system data interface can be connected via an internal vehicle communication bus and / or comprise external interface components independent of the vehicle's communication buses.

[0050] The invention can also include a vehicle comprising a system according to the invention, as well as a computer program that, when loaded and executed on a computer, performs process steps according to the invention.

[0051] Further embodiments of the invention are explained in more detail below with reference to the figures. Where identical, equivalent, or similar elements are shown in the figures, they may be designated with the same reference numerals across all figures. The figures show: Fig. 1 a vehicle with a system for detecting low-speed crash events and the process steps carried out with it, Fig. 2 feature spaces for low-speed crash events, Fig. 3 a system of Fig. 1 corresponding system with a learning mode, Fig. 4 a damage detection system for a vehicle and Fig. 5 a vehicle with a sensor system.

[0052] The in Fig. 1. Vehicle 1, represented only symbolically, includes a detection system 2 for detecting low-speed crash events on the vehicle 1, for example, caused by an impact of a schematically shown object 10 on the vehicle 1, at a relatively low relative speed between vehicle 1 and object 10. Furthermore, the Fig. Figure 1 shows a schematic representation of various sensors 3 located in the vehicle 1, an electronic control unit (ECU) 3a, and a controller area network (CAN) data bus 4 of the vehicle 1, through which signals and / or data can be exchanged between the sensors 3 and / or via electronic control units 3a of the vehicle 1. The sensors 3 generate corresponding sensor signals or data for measured variables, such as the vehicle 1's speed, accelerations in different directions (longitudinal / lateral acceleration), etc. The respective data or signals generated by the sensors 3 are packaged into corresponding data bus messages, possibly via directly connected, especially wired, control units, in possibly pre-filtered and / or unfiltered and especially coded form, and made available via the CAN data bus 4 of the vehicle 1.They can be read online, particularly in real time, by the detection system 2, which is also connected to the CAN data bus 4 and is implemented on a dedicated control unit, and / or offline, particularly for training purposes, from a trace memory 5 of the vehicle 1 and / or from an externally connected, corresponding trace database. If the vehicle 1 has an optional trace memory, all or selected data or messages output via the CAN data bus 4 during operation can be stored for a predefined period, making them available for later training and / or analysis purposes. The trace database is a database in which corresponding bus messages are stored in the form of so-called log traces, which were recorded beforehand, typically using appropriate data loggers, e.g.during the execution of corresponding driving or crash tests and / or under real accident conditions. The respective data may also be pre-processed and stored in a form advantageous for a machine learning (ML)-based training process used by the recognition system.

[0053] The recognition system 2 comprises a signal preprocessing unit 6 (SPU), a feature computation unit 7 (FCU), a classification unit 8 that operates on the basis of a machine learning process (Machine Learning Classification Engine, MLCE), and a classification database 9 that contains data that is generated and / or used in a machine learning process.

[0054] To detect a vehicle accident or a related impact event of an object 10 on the vehicle 1 with a low motion magnitude, such as a low impact speed or a low change in speed or acceleration associated with the impact (low-speed crash event), the signals and / or data generated by the sensors 3 of the vehicle 1 are processed in the signal preprocessing unit 6 using predefined rules in such a way that they are at least selectively filtered.

[0055] The impact velocity is, in particular, the relative velocity between the object and the vehicle, for example, a relative velocity greater than 0 and less than or equal to 20 km / h before the impact. However, the impact velocity can also be defined as the difference in the vehicle's velocity relative to its initial velocity at the time of impact. The respective velocity values ​​can be referenced to a specific point in time related to the impact event, for example, 300 ms before the point at which the acceleration maximum associated with the impact occurs. Alternatively, the profile of the corresponding velocity and / or acceleration values ​​can be used as the motion parameter.

[0056] The selectively filtered signals and / or data generated in the signal preprocessing unit 6 are then output to a feature calculation unit 7. In feature calculation unit 7, feature data is generated based on the previously selectively filtered signals and / or data using predefined rules. The feature data represents, in particular, an observational variable for the impact event. From a multitude of sensor signals or data generated by one or more sensors before, during, or after the impact event, a multitude of feature data, and especially multidimensional feature vectors, can be generated using predefined rules.

[0057] The signal preprocessing unit 6 is used in particular to filter relevant signals or data from the message traffic running via the communication bus (CAN data bus 4) connected to its input and to convert them into a suitable form in order to prepare them for the subsequent process of creating feature data.

[0058] The feature data generated in feature calculation unit 7, in particular feature vectors, and / or data derived therefrom, are output by feature calculation unit 7 to classification unit 8 as observation data, in particular as multidimensional observation feature vector data. There, the feature data are processed using at least one classification method and the classification database 9 in such a way that the impact event underlying the signals or data, or the feature data derived therefrom, is assigned to an event class of the classification. This assignment can be carried out using a classification model that comprises several classification methods.For classification purposes, two event classes are specifically provided. The first event class represents a low-speed crash, while the second event class represents the absence of such a crash (misuse case, which includes not only misuse events but also other non-critical events such as normal driving). Classification unit 8 then outputs a corresponding signal or at least a corresponding message, which can be read and used by other systems, particularly vehicle systems, for further control tasks and / or evaluations, for example, to notify a vehicle-independent device such as the vehicle owner's smartphone.for further use within offboard telematics platforms and services and / or for the fully or partially automatic determination of the damage caused to the vehicle by the accident.

[0059] The sequence of process steps for generating, providing, transmitting, processing, input and / or outputting signals and / or data is described in Fig. 1. Each is represented by arrows. The following section describes in more detail the system components and the process steps carried out within them for the signal preprocessing unit 6, the feature calculation unit 7, and the classification unit 8. Signal preprocessing unit

[0060] In the signal preprocessing unit 6, the sensor signals and / or data received from the CAN data bus 4 are first processed by a signal decoder 11 (FIL). The decoder decodes the signals or data, especially when they are encoded as messages, using a decoding rule stored, for example, in a vehicle database containing the corresponding message definitions of the respective communication bus protocol, and filters them according to predefined rules. Only selectively predefined signals and / or data are output to a synchronization unit 12 of the signal preprocessing unit 6, which contribute to the detection of an accident event. Although in Fig. 1 If only one decoding unit 11 is shown, which performs both decoding and filtering, two separate units may also be provided for these two functions.

[0061] When the respective signals or data transmitted via the communication bus, here CAN data bus 4, are propagated by means of asynchronous or "quasi-asynchronous" (with regard to their contained useful signals, e.g., despite a synchronous bus clock) message or data transmission, the previously decoded and filtered individual signals or data are synchronized in the synchronization unit 12 (SYN). Through appropriate sample rate conversion or resampling, the signals or data can also be brought to a common sampling frequency for subsequent processing. The signals and / or data brought to a common sampling frequency are fed to a signal windowing unit 13 (WIN) of the signal processing unit 6, where they undergo a process for the creation of a sliding time window.

[0062] In this process, for each predefined time window, the synchronized signals and / or data are logically assigned to one another and, in particular, combined into a data set. Each time window is defined using adjustable and preset parameters, specifically a window size parameter (duration of the time interval of a window or corresponding number of samples), an overlap parameter for the temporal overlap of adjacent signal windows (shift in sliding windowing / windowing), and / or a parameter for applying one or more predefined window functions, such as Hamming, Von-Hann (Hanning or Raised-Cosine), Blackman, or Kaiser. The parameters of a window function can also be set before its application, if necessary.In a “trivial” setting, for example, the window function would be a rectangular window (signal values ​​within the window are adopted unchanged) and an overlap factor of 0% (no overlap or sliding window moves by a full window length each time).

[0063] The respective signal samples are preferably temporarily stored in a memory according to the set window width and / or the set overlap factor. This ensures that signals or data required together for subsequent feature calculations can be reliably provided even if the signal samples arrive individually at runtime in the signal processing unit 6. Feature calculation unit

[0064] In feature calculation unit 7, a feature vector, particularly a multidimensional one, is calculated from the relevant sub-signals previously generated using the synchronous signal windows or from the corresponding data. Individual sub-components of a feature vector can be formed from individual features, each of which is calculated within correspondingly dedicated sub-feature calculation units 15. Furthermore, a single sub-feature calculation unit 15 can, where appropriate, calculate multiple sub-components of the feature vector or multiple features.

[0065] The core of the feature calculation unit 7 consists of several sub-feature calculation units 15 (FC1...FCn), which, using predefined rules, each calculate feature data that are assigned to either several, one, or, in particular, exactly one specific feature.

[0066] The feature calculation unit 7 optionally has a preprocessing unit 14 connected to the input side, or upstream of the sub-feature calculation units 15. The signals, data, partial or basic signals, and / or data segments previously generated by the signal windowing unit 13 of the signal preprocessing unit 6 are fed into this preprocessing unit. Here, these can optionally be preprocessed in such a way that operations required multiple times for the subsequent calculation of predefined feature data or their associated features, particularly for multiple feature data or features and / or sub-feature vectors, are already performed in at least one corresponding preprocessing step. This allows the respective results to be used directly in the subsequent steps for generating the feature data or features and / or sub-feature vectors. This avoids redundancy in the calculation and thus optimizes computing power.This is particularly advantageous for making the recognition system 2 capable of real-time operation.

[0067] Optionally, a post-processing unit 16 (POST) can be provided at the output side of the feature calculation unit 7, particularly downstream of the sub-feature calculation units 15. This POST unit processes the previously calculated feature data or features, especially sub-feature vectors, individually or in groups, and / or the resulting entire feature vector (which, within the scope of the invention, is also referred to as the observation feature vector) according to predefined rules. For example, vector quantization could be applied to an originally calculated feature vector and / or the observation feature vector. The data of the observation feature vector, formed from the individual features / sub-feature vectors, are stored in a register-memory 17.The data stored in predefined register cells of register memory 17 represent individually calculated features.

[0068] For each aggregated signal window block applied to the input side, feature calculation unit 7 outputs a resulting multidimensional observation feature vector. Depending on the settings or parameterization of feature calculation unit 7, this vector has at least two dimensions, or at least two calculated features as vector components. The output frequency is analogous to, or dependent on, the corresponding windowing parameters such as the sampling rate of the windowed samples, window width, and / or overlap factor.

[0069] The observation feature vector is fed to the classification unit 8 for further signal or data processing in order to classify the underlying physical event (which may also be non-critical) that may have been indirectly observed via the previously described process steps. Classification unit

[0070] In classification unit 8, a decision is made regarding the belonging to one of the previously defined event classes based on the feature signals or data supplied to it, summarized by the observation feature vector. For this purpose, the respective signals or data are first processed in at least one of the classification processors 18a, 18b, or 18c according to the settings or parameterization of classification unit 8, using the classification methods learned by the respective machine learning procedure. 18a represents a classification processor with an implementation according to a Hidden Markov Model (HMM) method, classification processor 18b with an implementation according to a Support Vector Machine (SVM) method, and 18c the classification processor with an implementation according to a method based on an artificial neural network (ANN). Corresponding parameters for the respective methods are also defined.Control and / or rule data belonging to classification models are stored in classification database 9. Training feature data may also be stored in classification database 9. The respective implementation of the aforementioned machine learning methods for applying the classification method comprises at least one, and in particular a multitude of, rules such as assignment rules, comparison rules, and / or calculation rules. The implementation may, in particular, take place in a computer program.

[0071] The Hidden Markov Model (HMM) is a stochastic model in which a system is modeled by a Markov chain with unobserved states, named after the Russian mathematician Andrey Andreevich Markov. Modeling as a Markov chain means that the system transitions randomly from one state to another, with the transition probabilities depending only on the current state and not on the preceding states. Furthermore, it is assumed that the transition probabilities are constant over time. In an HMM, however, these states themselves are not observed externally. They are hidden. Instead, each of these internal states is associated with observable output symbols, called emissions, which occur with certain probabilities depending on the state.The task usually involves deriving probabilistic statements about the hidden states from the observed sequence of emissions. A HMM can be considered a special case of a dynamic Bayesian network.

[0072] A Support Vector Machine (SVM) divides a set of objects into classes such that a wide margin around each class remains free of objects; it is a so-called large-margin classifier. The starting point for building a Support Vector Machine is a set of training objects, each with a known class. Each object is represented by a vector in an N-dimensional vector space. The Support Vector Machine's task is to fit a hyperplane of dimension N-1 into this space, which acts as a dividing surface and separates the training objects into two classes. The distance between the vectors closest to the hyperplane is maximized. This wide, empty margin ensures that even objects that do not exactly match the training objects are classified as reliably as possible.

[0073] When defining the hyperplane, it is not necessary to consider all training vectors. Vectors that are further away from the hyperplane and, in a sense, "hidden" behind a front of other vectors, do not affect the location and position of the separating plane. The hyperplane depends only on the vectors closest to it—and only these are needed to describe the plane mathematically precisely. These nearest vectors are called support vectors, based on their function.

[0074] A hyperplane cannot be "bent," so a clean separation using a hyperplane is only possible if the objects are linearly separable. This is generally not the case in real-world applications.

[0075] Other classification processors based on different classification methods, such as decision trees or other dynamic Bayesian networks, can also be used.

[0076] Based on the results of the processing in at least one classification processor 18a, 18b, or 18c, the class decision unit 19 then decides which event class the underlying or observed event is assigned to. Optionally, the results of the respective outputs from the different classification processors 18a, 18b, and 18c can be statistically analyzed to generate an overall result. The event classes can be defined, for example, in binary as the class label for the event class "Misuse" (or no accident / critical event) with the class index zero, and as the class label for the event class "Low Speed ​​Crash" with the class index one. Accordingly, the class decision unit 19 or the classification unit 8 outputs the class index as a result via an interface 20 (IF).The most probable event class for the underlying event is selected according to the respective classification procedure. If supported by the selected classification procedure and / or the class decision unit 19, the class probability of the currently evaluated feature vector can also be determined and output, either additionally or alternatively.

[0077] The acquisition and processing of signals, data, and / or messages, and in particular the classification procedure, is continuously performed during operation (while the vehicle is in motion, when the ignition is on, and / or possibly in a corresponding standby mode), even if no critical event is present, in order to detect a critical event promptly. As a rule, a classification of the event "no accident" is therefore constantly issued.

[0078] Fig. Figure 2 shows two feature spaces 24 and 25 for representative impact events and, in particular, borderline misuse events between a vehicle and an object. The feature data corresponding to a low-speed crash event are represented by circles 26, and the feature data corresponding to a misuse event are represented by crosses 27.

[0079] In feature space 24, the respective initial velocities of the vehicle are specified in the x-axis direction, and in the y-axis direction, the magnitude of the respective maximum acceleration of an acceleration sensor of the vehicle occurring during the impact event.

[0080] How Fig. As can be seen in Figure 2, the low-speed crash events are not clearly distinguished from the misuse events in feature space 24, especially in the overlap area 28 and, for example, for the event symbolized by the cross 27a. Fig. Figure 2 thus shows, on the one hand, that it is hardly possible to correctly assign or recognize the respective event type (low-speed crash vs. misuse) from the respective feature data of feature space 24 (or the aggregated motion variables "initial velocity" and "acceleration value" represented therein) if the event type is not known in advance. On the other hand, it shows Fig. 2, that such an assignment or recognition can be carried out much more accurately using the data of feature space 25, as evidenced by the clearer separation of the two areas 25a and 25b. Feature space 25 is a more advantageous feature space for event recognition. The derived features represented on the respective axes u and v of feature space 25 each correspond to one or more features that can be assigned using the data in Fig. The partial feature calculation units 15 (FC1...FCn) shown in Figure 1 are calculated, whereby the underlying features are determined differently than the features of feature space 24. In particular, these can be determined on more complex data and / or data processing steps, e.g., also on the basis of different data types, for example, on different motion variables and / or on other variables such as acoustic and / or optical variables. The clearer separation of areas 25a and 25b advantageously allows for a more precise automatic assignment of the respective data of an event to one of the two event classifications (low-speed crash event vs. misuse event).

[0081] In Fig. 3 is the one in Fig. 1 shown recognition system 2 shown, where, among other things, system components are shown that are intended for training the classification unit 8 as part of the machine learning process.

[0082] The detection system 2 can be described in at least two operating states: firstly, a training mode, in which features, in particular parameters and / or rules for the classification unit 8, are determined based on a machine learning process; and secondly, a productive operating state, or so-called "run mode," in which a decision is actually made about whether a low-speed crash has occurred, based on the data acquired by sensors in a vehicle, especially online at runtime and / or in real time. In connection with Fig. 1 was already described above using the system components shown there [mainly] the “Run-Mode”.

[0083] The System Data Unit (SDU) 21, also referred to as the System Data Module, serves as the central storage location for system-relevant data, which essentially includes the system configuration or parameterization, as well as the data stored in the classification database 9, which are generated and stored particularly during the machine learning process.

[0084] In classification unit 8, a logical classification selector switch 18d or a corresponding selection determines which of the respective classification processors 18a...18c is to be trained or used. This selection of one or more classification processors can also be made in "run mode," so that the corresponding feature signals or feature data to be processed pass through one or more classification processors in parallel or serially.

[0085] Both system configuration data and data for controlling read and / or write access to the classification database 9 are made available to the respective system components of the recognition system 2, in particular the classification unit 8, via an internal data interface 22 (IIF). For this purpose, corresponding, dedicated input and / or output ports 22a can be provided for the signal processing unit 6 (SPU IO), for the feature calculation unit 7 (FCU IO), and for the classification unit (MLCE IO).

[0086] System configurations can still be made externally via the external interface 23, and / or information on the current status of the detection system 2 can be queried. Configuration data (parameters) is stored in a configuration data store 21a.

[0087] Although that in Fig. Since the recognition system 2 shown in Figure 3 has the two operating modes "training" and "productive operation" (Run), it is possible to provide a corresponding recognition system 2 that only has the operating state "Run" and is pre-populated with control data, in particular classification data, classification rules, and / or classification models, which were previously generated or trained in another system, in particular a reference system, especially on the basis of a machine learning process, the so-called supervised learning. Such a system then has, in particular, the features shown in Figure 3. Fig. 1 system components shown.

[0088] In "Training" operating mode, two subordinate learning modes can be provided. In the first learning mode, "Train Features," the recognition system 2 performs signal preprocessing and feature calculation, storing the corresponding calculated feature vector for each generated signal time window [block], particularly in the classification database 9. In the second learning mode, "Train Model," data, rules, and / or models are initially specified manually. Specifically, the respective underlying event is manually assigned to each measurement data and / or derived data—a process known as labeling.In this process, the event classes to be assigned (Class 0 "Misuse" and Class 1 "Low Speed ​​Crash") are associated by a domain expert with the previously defined time windows and / or the previously calculated feature vectors and permanently stored within the training database 9. The actual machine learning classification model can then be trained. The resulting trained classification model, whose specific implementation can vary depending on the chosen machine learning method, is also permanently stored in the classification database 9. It is then used, if necessary, during run operation by the corresponding machine learning classification method to classify the respective feature data or feature vectors, or the observed feature vector, to be evaluated.

[0089] To which based Fig. 3. Various methods can be used to train the artificial neural network (ANN) as described above, either alone or in combination, in particular so-called supervised learning, reinforced learning, unsupervised learning and stochastic learning.

[0090] In supervised learning, an input pattern is given to the ANN (Active Neural Network), and the output it produces in its current state is compared to the value it is supposed to output. By comparing the expected and actual output, conclusions can be drawn about which changes to the network configuration need to be made.

[0091] In reinforcement learning, an agent determines the utility of sequences of actions in a given world. For this purpose, reinforcement learning primarily utilizes the theory of Markov decision processes (MDPs). The underlying approach is to distribute rewards given to an agent across previous actions in such a way that the agent understands and can exploit the utility of each action.

[0092] Unsupervised learning occurs primarily through the input of patterns to be learned. The artificial neural network then modifies itself according to these input patterns.

[0093] In stochastic learning, a learning algorithm attempts to find a hypothesis that makes predictions with the greatest possible accuracy. A hypothesis is understood as a mapping or rule that assigns the predicted output value to each input value. To achieve this, the algorithm modifies the free parameters of the chosen hypothesis class. Often, the hypothesis class is defined as the set of all hypotheses that can be modeled by a specific artificial neural network. In this case, the freely selectable parameters are the weights w of the neurons.

[0094] With regard to Fig. 1 and Fig. In summary, it can be stated that the described detection system 2 has three main system components or signal processing stages with regard to the signal processing chain, namely - a signal preprocessing unit or stage, - a feature calculation unit or level and - a classification unit or level.

[0095] Furthermore, the respective system 2 may preferably include a system data module configured to provide storage space for required system data, such as for system configuration, system parameterization, and / or a machine learning training database containing training data for a machine learning process. The system data module may also provide at least one internal and / or external interface for input and / or output data access.

[0096] In Fig. Figure 4 shows an automated damage estimation system 29 (ADE) for a vehicle, which can automatically decide after an impact event whether it is a low-speed crash event and, if applicable, automatically determine and / or assess the extent of any damage that may have occurred.

[0097] The damage detection system 29 comprises a crash detection system 37 (CD) and a damage estimation module 30 (DEM). The crash detection system 37, in turn, comprises two subsystems that evaluate data provided by vehicle sensors as accident data, essentially independently and, in particular, based on different algorithms. The data and / or algorithms to be evaluated can be at least partially the same for both subsystems.

[0098] The first subsystem is specialized in detecting low-speed crash events (Low Speed ​​Crash Detection, LSCD), and the second subsystem is specialized in detecting serious accident events, so that, for example, an active occupant protection system in the vehicle is activated. The first subsystem is essentially designed according to detection system 2 and, in particular, according to one of the [references to be added]. Fig. The detection systems described in 1 or 3 are designed for the corresponding classification of the impact event.

[0099] The second subsystem 38 is a crash detection module that detects serious accidents at higher relative speeds between the vehicle and the impacting object (High Speed ​​Crash Detection System, HSCDS), in particular based on sensors, signal and data processing methods, and control units for a safety restraint system (SRS), so that at least one active occupant protection system 39 can be activated, for example, an airbag. This system 38 can also receive input data from sensors 3, control units 3a, a communication bus, and / or vehicle-internal databases such as an accident database 40.

[0100] The signals and data output by the detection system 2 and the accident detection module 38 are transmitted via a common interface 20. The information provided via interface 20 regarding the event class assigned to the impact event by the detection system 2 and / or the information provided by the accident detection system 37, in particular the respective class labels or class indices, is transmitted to a damage estimation module 30 (DEM).

[0101] Furthermore, data and / or signals are fed to the damage assessment module 30 via the crash data path 31. These data and / or signals are provided by the vehicle's sensors 3, other sensors, and / or control units 3a, and are generated by these components in particular before, during, and / or after the impact event. This data may include, in particular, data or signals indicating accelerations, deformations of vehicle parts, and / or other mechanical and / or electrical defects of the vehicle. The data may also include, in particular, image data or image signals generated by a camera in the vehicle. The respective data may be stored, in particular, in a volatile data storage device and / or in a non-volatile data storage device, in particular in the vehicle's accident database 39.

[0102] In the damage assessment module 30, the signals or data supplied to it are processed and, in particular, evaluated such that an impact estimation module 30a, which assesses the impact (Impact Estimation, IE), outputs at least one damage feature value, in particular an impact feature vector. For this purpose, several data processing units are provided in the damage assessment module 30. A corresponding impact severity determination module 30 comprises a crash severity determination unit 32 (CSE), which determines and outputs a value for the severity of the impact based on predefined rules, and a crash direction determination unit 33 (IMP).A Crash Direction Unit (CDU), which determines and outputs a value for the impact direction relative to the vehicle axes based on predefined rules; a Crash Localization Determination Unit 34 (CLE), which determines and outputs the location of the impact on the vehicle based on predefined rules; and / or a Vector Processing Unit 35, which calculates the damage characteristic vector from the values ​​output by units 32, 33, and 34 based on predefined rules. This vector can then be output to an Expert System 36 (ES), which in turn includes a corresponding database 36a, for a more precise determination of any damage caused to the vehicle during the impact based on predefined rules and data from database 36a.

[0103] Where this document refers to predefined rules, such rules, in particular parameters of the rules, may be adjustable.

[0104] The accident damage detection system 29 can be provided entirely in the vehicle or partly in the vehicle (in particular with regard to the accident detection system 37 or its subsystems 2, 38) and partly in a device, in particular comprising a computer, outside the vehicle, in particular with regard to the damage assessment system 30 and its respective components.

[0105] Although the accident damage detection system 29 has been described with two subsystems, namely the low-speed crash detection system 2 and the high-speed crash detection system 38, in order to issue a trigger to activate the damage detection module 30, it can also be based on only one of these subsystems.

[0106] At the in Fig. In the vehicle 1 shown in Figure 5, a crash sensor system (CS) 41 is provided in its front area 1a, by means of which at least one accident-relevant movement parameter can be detected in a direction a shown by arrow 44. One or more sensor values, in particular at least one corresponding acceleration value, are processed in an electronic control unit 42 by means of a signal processor 42. If at least a first threshold value is exceeded or fallen below, a corresponding signal is output to an active occupant protection system (IS) 45 to control its actuators, which trigger its units such as airbag, seat belt tensioner, etc. A movement parameter value detected by the crash sensor system 41 is also output to the low-speed crash detection unit (LS) 2 described above, by means of which a low-speed crash event is detected.

[0107] The crash sensor system 41 comprises two acceleration sensors (BS) 46a, 46 and a force sensor (KS) 47, which can also be configured as a pressure sensor. The respective sensors can detect corresponding motion quantities, with the force sensor 47 capable of detecting quasi-static acceleration. The respective sensor signals are fed to the electronic control unit 42 for processing in the processor 43. The vehicle 1 can also utilize other, already integrated Fig. 1 to 4 components are included.

[0108] It was stated above that the described methods and systems can automatically detect a low-speed crash event with high reliability, and the corresponding data can be stored in the vehicle for documentation purposes and / or transferred to an external storage device. Furthermore, it can be provided, for example, that the vehicle user and / or owner is automatically notified via a mobile communication system in the vehicle, such as a Short Message Service (SMS), if the vehicle is struck while parked. It can also be provided that one or more other systems in the vehicle are activated in response to the collision, e.g., the horn to inform the other driver that the vehicle has been damaged, or a camera to document the damage.

[0109] The described devices and system components are controlled, in particular, by computer programs and may therefore include further elements of computers and digital control devices that are known per se, such as a microprocessor, volatile and non-volatile memory, interfaces, etc. The invention can therefore also be realized wholly or partially in the form of a computer program product that, when loaded and executed on a computer, wholly or partially effects a process according to the invention. It can, for example, be provided in the form of a data carrier such as a CD / DVD or in the form of one or more files on a server from which the computer program can be downloaded. Reference symbol list 1 vehicle 1a Front area 2 Recognition system 3 sensors 3a Control unit 4 CAN bus 5 Trace memory 6 Signal processing unit 7 Feature calculation unit 8 Classification unit 9 Classification database 10 objects 11 Signal Decoder 12 Synchronization unit 13 Signal window unit 14 Pre-processing unit 15 sub-feature calculation unit 16 Post-processing unit 17 Vector Processing Unit 18a...18c Classification Processor 19-class decision unit 20 interface 21 System data module 21a Configuration-Data Storage 22 internal interface 23 external interface 24 Feature space before processing 25 Feature space after processing 26 Low Speed ​​Crash Feature Data 27 Misuse of Feature Data 28 Overlap area 29 Damage detection system 30 Damage Assessment Module 30a Impact Assessment Module 31 Crash data path 32 Crash severity detection unit 33 Crash direction detection unit 34 Crash Localization Unit 35 Vector Processing Unit 36 Expert system 36a Database 37 Accident detection system 38 Accident Detection Module 39 occupant protection system 40 Accident Database 41 Crash sensor system 42 Control unit 43 processor 44 Acceleration direction 45 Active occupant protection system 46a, 46b Accelerometer 47 Pressure sensor

Claims

[1] Method for detecting a vehicle accident in which a vehicle (1) and an object (10) collide, wherein a movement quantity associated with the impact is so low that at least one active occupant protection system (39) of the vehicle (1) provided for accidents is not activated, and in which the impact event - signals and / or data generated by sensors (3) of the vehicle (1) are processed in such a way that - the signals and / or data are filtered, - feature data is created on the basis of the filtered signals and / or data, - the feature data is used to assign the impact event to a class index of a classification and -- where feature vectors are formed from the feature data, -- wherein a parameterizable classification unit (8) is used, which comprises a plurality of classification processors (18a, 18b, 18c), each of which executes different classification methods learned in different machine learning methods (ANN, HMM, SVM), -- wherein the feature vectors are processed in accordance with a parameterization of the classification unit (8) using the respective classification method (HMM, SVM, ANN) in at least one of the classification processors (18a, 18b, 18c), and the results of the respective results output by the classification processors (18a, 18b, 18c) are statistically analyzed to form an overall result, -- where for the overall result a class index is assigned to the most probable event class for the underlying event based on the respective classification method and / or a class probability of a currently evaluated feature vector is determined and -- wherein control and / or regulation data belonging to the respective classification methods (HMM, SVM, ANN) are used for the parameterization, which control and / or regulation data are stored in a classification database (9). [2] Method, in particular according to claim 1, for detecting a vehicle accident in which a vehicle (1) and an object (10) collide, wherein a movement quantity associated with the impact is so low that at least one active occupant protection system (39) of the vehicle (1) provided for accidents is not activated, and in which the impact event - signals and / or data generated by sensors (3) of the vehicle (1) are processed in a signal pre-processing unit (6) in such a way that - the signals and / or data are filtered, whereby during filtering by means of predetermined time window data predetermined selection signals and / or selection data are selected and / or derived from the signals and / or data formed by means of the sensors, - feature data is formed on the basis of the filtered signals and / or data and - the feature data are used to assign the impact event to a class index of a classification, using a classification database (9) and - wherein, if the signals and / or data are propagated by means of an asynchronous or quasi-asynchronous message transmission with regard to the useful signals they contain, the previously decoded and filtered individual signals and / or data are synchronized in a synchronization unit (12) of the signal preprocessing unit (6) and are brought to a common sampling frequency for subsequent further processing by means of a sampling rate conversion, and - the signals and / or data brought to a common sampling frequency are fed to a signal windowing unit (13) of the signal preprocessing unit (6), in which they undergo a process for sliding temporal signal window formation. [3] Method according to claim 2, wherein a window function, in particular Hanning, Hamming, Blackman, or Kaiser, is used for the sliding signal window formation. [4] Method according to one of the preceding claims, characterized bythat for the assignment at least one classification method is formed by means of a machine learning method, in particular by means of a method based on an artificial neural network, support vector machines, a hidden Markov model or decision trees. [5] Method according to one of the preceding claims, characterized by that the signals and / or data formed by means of the sensors (3) of the vehicle (1), signals and / or data derived therefrom and / or further signals and / or data formed by control units (3a) of the vehicle (1) are transmitted via a data bus (4) of the vehicle (1). [6] Method according to one of the preceding claims, characterized by that the movement quantity associated with the impact is the impact speed, the change in speed and / or acceleration of the vehicle associated with the impact and / or an energy quantity associated with the impact. [7] Method for determining damage to a vehicle (1) after an accident, comprising a method according to one of the preceding claims and further comprising the steps of determining further accident data by means of the assignment of the impact event, by means of which the extent of damage to the vehicle (1) is determined. [8] System for detecting a vehicle accident, arranged to carry out a method according to one of claims 1 to 7. [9] Damage detection system for a vehicle comprising a system according to claim 8. [10] A vehicle comprising a system according to claim 8 or 9. [11] A computer program product which, when loaded and executed on a computer, effects a method according to any one of claims 1 to 7.

Images