Computer implementation method for training machine learning models in vehicle accident event evaluation

Abstract

The present disclosure relates to a computer-implemented method for training a machine learning model in vehicle accident event assessment. The method includes implementing a machine learning model (S100) and preparing training data (S200) by acquiring and automatically learning (S210) sensor data generated from vehicle sensors (2, 3; 4, 5). The method further includes applying data augmentation to the collected data to artificially increase the amount of collected data learned from each sensor, such that an augmented data set is learned for each sensor. For each sensor (2, 3; 4, 5), applying data augmentation (S220) includes transforming the collected data and applying it to multiple different sensor mounting positions and / or multiple different sensor mounting orientations within a specific sensor mounting region (6, 7) (S221) so that an additional data set is acquired, the augmented data set including the collected data and the additional data set. The method further includes training a machine learning model using the augmented data set for each sensor (2, 3; 4, 5) (S300).

Description

Technical Field

[0001] The present disclosure relates to an improved computer-implemented method for training a machine learning model in vehicle accident event assessment.

Background Art

[0002] Vehicle accident events occur, for example, when riding a bicycle, scooter, motorcycle (MC), etc. due to, for example, skidding, slipping, loss of vehicle control, collision with another vehicle or object, etc. Falls associated with a moving vehicle often occur more suddenly and require a quicker response than falls that occur to a walking person, such as falling or slipping.

[0003] In particular, MCs can be involved in many collisions, many of which can result in severe injuries or death. Part of the reason is that riders on MCs are in a more vulnerable state compared to drivers inside automobiles. Also, there is no restraint such as a seat belt on an MC like in an automobile. There are also other dynamics that make MC riders much more vulnerable in case of an accident or collision. An MC integrated with an airbag system can reduce the risk of severe injury or death in case of an accident, especially in case of a frontal collision or a slightly angled collision, and helps resist the rider being thrown out of the MC as a result of the collision. This can be life-saving when used effectively.

[0004] Furthermore, an MC driver may be equipped with one or more airbags, or wear a self-inflating vest, or a smart helmet having one or more airbags, which are adapted to provide protection to an MC rider wearing at least one of the vest and the helmet when the MC rider is thrown out of the MC after a collision. Such a helmet may also be provided with one or more devices capable of emitting one or more types of warnings. For example, other protective clothing such as a belt is also conceivable.

[0005] However, a real-time airbag trigger system needs to detect a collision in milliseconds and then trigger a signal to inflate the airbags.

[0006] Most approaches typically involve observing changes in velocity or acceleration over time to predict vehicle collisions or other types of accidents. While this approach works relatively well, it is limited in terms of flexibility, particularly with respect to various circumstances such as velocity, angle, and other possible modes of collision.

[0007] Therefore, it is desirable to provide a precise real-time airbag trigger system for motorcycles and other two-wheeled or three-wheeled vehicles that mitigates the aforementioned drawbacks. [Overview of the Initiative]

[0008] This is achieved by a computer implementation method for training a machine learning model in vehicle accident event assessment, including vehicle accident event detection for two-wheeled or three-wheeled vehicles. The method includes implementing a machine learning model and preparing training data. Preparing training data includes acquiring and automatically learning sensor data generated from sensors located on the vehicle and used to collect data, and applying data augmentation to the collected data in order to artificially increase the amount of collected data acquired from each sensor. In this way, an augmented dataset is acquired for each sensor.

[0009] For each sensor, applying data augmentation involves transforming the collected data and applying it to multiple different sensor mounting positions and / or multiple different sensor mounting orientations within a particular sensor mounting area, so that an additional dataset is obtained. The augmented dataset includes the collected data and the additional dataset. The method further includes training a machine learning model using the augmented dataset for each sensor.

[0010] This means that data augmentation is applied, artificially increasing the collected data, the training data, to be much larger than the original training data, and adding new, artificially generated data that wasn't there initially but is produced to provide a more robust machine learning model.

[0011] For example, sensors may not be the same and / or have slightly different configurations, or collisions may have occurred at slightly faster speeds. Machine learning models can be trained using augmented data, making them far more robust to real-world irregularities.

[0012] In some aspects, applying data augmentation is - Applying appropriate rotational extension to motion sensors, and applying a small orientation along a specific axis to the sensor data to represent changes in the sensor mounting orientation on the vehicle. - Applying appropriate noise enhancement, such as Gaussian noise, to represent noisy data from a particular sensor, and -Includes at least one of the following: applying a first amplitude scaling data augmentation to represent the perceived intensity change of motion acceleration at different mounting positions of sensors on the MC that are in close proximity to each other.

[0013] Therefore, extensions can be obtained in many different ways, and the above extensions are suitable, for example, for handling multiple different sensor mounting positions and / or multiple different sensor mounting orientations within a particular sensor mounting area.

[0014] In some embodiments, at least two sensors are positioned within a sensor mounting area, and the machine learning models for the sensors are trained using the same augmented dataset.

[0015] This allows for the acquisition of a fully trained, robust backbone machine learning model for each sensor mounting area, which can then be used for one or more sensors located in adjacent positions, allowing for slight variations in these positions within the sensor mounting area.

[0016] According to some embodiments, if sensor data acquired from at least two sensors located within the same sensor mounting area indicates the occurrence of an accident within a certain predetermined period, at least one safety measure is triggered; otherwise, no safety measure is triggered.

[0017] This minimizes the risk of unintended triggering of safety measures, such as active safety systems, using multi-sensor devices. This is particularly useful when it is difficult to detect scenarios that look very obvious, such as collisions or other types of vehicle accidents, but are not actually accidents.

[0018] In some aspects, applying data augmentation is - To represent various velocities in vehicle collision dynamics, apply a second amplitude scaling data extension. - Applying time-compressed / uncompressed data augmentation to represent the velocity at which collision dynamics unfold, and - To represent noise or unwanted data in the signal due to errors, apply segments of appropriate length and position within the signal, which have been scrambled with appropriate intensity. This includes at least one of these.

[0019] Therefore, extensions can be obtained in many different ways, and the above extensions are suitable, for example, when a vehicle accident must be detected within a limited period of time in which it is considered to have been verified and registered by at least two of the sensors.

[0020] According to some aspects, the sensor includes at least one of a motion sensor, a 3D sensor, an accelerometer, a gyroscope, and / or a camera. This means that many types of available and well-known sensors can be used.

[0021] According to some aspects, additional datasets are used to cover possible deployment scenarios where the initially collected dataset lacks data.

[0022] According to some aspects, additional datasets are used to enhance robustness against various changes in vehicle speed, vehicle weight, and signal errors due to temporary or permanent malfunction of sensors during a collision scenario.

[0023] This diverse type of training data is the key to training robust and flexible machine learning models in the future.

[0024] The present disclosure also relates to computer programs, control units, and systems associated with the above advantages.

[0025] Thus, it is understood that all computer-implemented methods, computer programs, and computer-readable media described herein and according to the present disclosure may all be implemented in hardware such as the control unit configurations and devices described herein, as well as in the systems described herein. Hardware such as the control unit configurations described herein, as well as the systems described herein, are then configured to execute the computer-implemented method and computer program, thereby obtaining the same advantages and effects as described herein for the computer-implemented method.

Brief Description of the Drawings

[0026] Here, the present disclosure will be described in more detail with reference to the accompanying drawings. [Figure 1] FIG. 1 schematically shows a user riding a bike with an integrated sensor. [Figure 2] Figure 2 schematically shows a control unit. [Figure 3] Figure 3 shows an exemplary computer program product. [Figure 4A] Figure 4A shows a flowchart of a method according to the present disclosure. [Figure 4B] Figure 4B shows a flowchart of a method according to the present disclosure. **DETAILED DESCRIPTION OF THE INVENTION**

[0027] Aspects of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings. However, the different configurations, devices, systems, computer programs, and methods disclosed herein can be implemented in many different forms and should not be construed as limited to the aspects described herein. Like numbers in the drawings refer to like elements throughout.

[0028] The terms used herein are for the purpose of describing only aspects of the present disclosure and are not intended to limit the invention. As used herein, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0029] Modern data-driven approaches based on ML (Machine Learning) and DL (Deep Learning) provide greater flexibility to automatically learn from the provided collision data, driving data from various roads, and other stress tests. Also, these methods can adapt to newer collision scenarios, vehicles, and variables as more and more data becomes available. In essence, they can be retrained and learn new collision scenarios when newer data is provided.

[0030] Typically, ML-based methods are specific to the various aspects of the systems they are trained on. For example, if a model is trained on data collected from a specific sensor location on a vehicle, such as a particular orientation, it may not perform well at a different sensor location, even on the same vehicle with the same sensor type.

[0031] Referring to Figures 1, 4A, and 4B, the present invention relates to a computer implementation method for training a machine learning model in vehicle accident event evaluation, including vehicle accident event detection for a two-wheeled or three-wheeled vehicle 1. The method includes implementing a machine learning model S100 and preparing training data S200. Preparing training data S200 includes acquiring and automatically learning sensor data generated from sensors 2, 3; 4, 5 that are positioned on the vehicle 1 and used to collect data S210.

[0032] According to some embodiments, vehicle accident event evaluation also includes vehicle accident event detection, vehicle accident event characterization, and / or calculation of accident event risk probabilities.

[0033] Referring to Figure 1, there is a subject 8 in the form of a user 8 riding a motorcycle 8, and the user 8 is wearing a smart protective helmet 9 and a protective vest 10. According to some embodiments, at least one of the helmet 9 and vest 10 is equipped with one or more airbags 12, 13 adapted to provide protection to the user 8 wearing at least one of the vest 10 and helmet 9 when the user 8 is thrown from the motorcycle after a collision. Such a helmet 9 may also be equipped with one or more devices that can emit one or more types of warnings. Other protective clothing such as a belt is also conceivable, for example. One or more airbags 14 may also be integrated with the vehicle 1.

[0034] This disclosure is applicable to users traveling on or inside any type of two-wheeled or three-wheeled vehicle.

[0035] Sensor data is typically collected, i.e., acquired, from, sensors 2, 3, 4, 5, for example, mounted motion sensors included in vehicle 1, 3D sensors such as accelerometers that measure three-dimensional acceleration, and gyroscopes that measure three-dimensional rotational speed, but is not limited to these. According to some embodiments, sensors 2, 3; 4, 5 include at least one of motion sensors, 3D sensors, accelerometers, gyroscopes, and / or cameras.

[0036] This means that many types of readily available and well-known sensors can be used.

[0037] In some embodiments, as shown in Figure 2, the subject 8 wears means 11 for providing and communicating the subject 8's location. Such means 11 may comprise any type of suitable positioning system, such as GPS or GNSS (Global Navigation Satellite System), and any type of wireless communication system. Such GPS information may also be used to estimate the vehicle's speed, which can provide additional input for collision detection algorithms. For example, an airbag may be triggered only when a speed limit, such as 20 km / h, is passed. Such means may be provided in a vest 10 or any suitable type of clothing or wearable device.

[0038] For example, vehicle accident event detection achieved by a vehicle accident event detection algorithm relies on detecting motion patterns belonging to a vehicle accident, such as a collision, in which vehicle 1 moves suddenly and unexpectedly.

[0039] Continuing to refer to Figures 4A and 4B, preparing the training data S200 further includes applying data augmentation to the collected data in order to artificially increase the amount of collected data acquired from each sensor 2, 3; 4, 5 so that an augmented dataset is acquired for each sensor 2, 3; 4, 5. For each sensor 2, 3; 4, 5, applying data augmentation S220 includes transforming the collected data and applying it to multiple different sensor mounting positions and / or multiple different sensor mounting orientations within a particular sensor mounting area 6, 7 so that an additional dataset is acquired, the augmented dataset includes the collected data and the additional dataset.

[0040] The configuration parameters used in S221 to transform the collected data are, in some aspects, driven by heterogeneity requirements in the assumed target deployment scenario.

[0041] The method further includes training a machine learning model using an augmented dataset for each sensor 2, 3; 4, and 5 of the S300.

[0042] This means that data augmentation is applied, artificially increasing the collected data, the training data, to be much larger than the original training data, and adding new, artificially generated data that was not there initially, but is generated to provide a more robust machine learning model, enabling better prediction of collision / non-collision scenarios when real-world heterogeneity occurs. For the development of a high-quality machine learning model, a lot of relevant training is important, and collecting vehicle collision data such as MC is very expensive.

[0043] In some aspects, augmented data allows for the artificial creation of data for scenarios where the initially collected data was not applicable. For example, sensors may have different and / or slightly different configurations, or collisions may have occurred at slightly faster speeds. Machine learning models can be trained using augmented data, making them far more robust to real-world irregularities.

[0044] For example, as shown in Figure 1, if the orientation variation in the deployment of a particular sensor position is known or expected to be less than ±20% along the x-axis, then rotational expansion will only be generated within this range. Furthermore, if a particular sensor position has a known vibration pattern, this can be incorporated into the expanded data.

[0045] According to several embodiments, at least two sensors 2, 3; 4, 5 are located within sensor mounting areas 6, 7, and machine learning models for sensors 2, 3; 4, 5 are trained using the same extended dataset. This disclosure provides a fully trained, robust backbone machine learning model for each sensor mounting area 6, 7. This machine learning model can be used for one or more sensors located in adjacent positions, where slight variations in these positions within the sensor mounting areas 6, 7 are permitted. If more than two sensors and algorithms are involved, alternative methods such as majority voting may be used.

[0046] In some embodiments, the usefulness of such machine learning models lies in their robustness against various possibilities in each sensor mounting area 6, 7, such as small variations in sensor placement, e.g., small rotations, different types of sensor hardware. This is particularly useful in scenarios where data for algorithm development is collected from a specific sensor and sensor location, but during manufacturing, for example, due to various constraints and changes in sensor orientation, it is found that the sensor cannot be placed in the exact same location, but a nearby location is possible. Even the sensors used during manufacturing may differ from those used for data collection during development. Here, the same carefully trained machine learning model can be deployed on a single sensor or multiple sensors, where they are in the vicinity of their intended location, i.e., within the sensor mounting areas 6, 7. The machine learning model is robust against small variations in position, orientation, or noise profile from different types of sensors, or small amplitude differences due to changes in the type of sensor deployed, for their nearby locations.

[0047] Different sensor mounting areas 6 and 7 require different corresponding robust backbone machine learning models. This flexibility can be very beneficial for vehicle manufacturers during system integration, as they must balance various constraints when deploying system electronics in production vehicles.

[0048] Mounting areas 6 and 7 are not precisely specified, but may have a certain size due to numerous parameters related to the collected data, sensor type, etc. For example, a machine learning model generated for a mounting area at the footrest position may not be applicable to the headlamp position for collision prediction. This is because the mounting areas are not sufficiently close together.

[0049] According to some embodiments, when sensor data acquired from at least two of the sensors 2, 3; 4, 5 located within the same sensor mounting areas 6, 7 indicates the occurrence of an accident within a specific predetermined period, at least one safety measure is triggered; otherwise, no safety measure is triggered.

[0050] Essentially, a collision or other type of vehicle accident must be detected by at least two of sensors 2, 3; 4, and 5, and within a limited timeframe, in order to be verified and registered. This minimizes the risk of unintended triggering of safety measures, such as active safety systems, using multi-sensor devices. This is particularly useful when it is difficult to detect scenarios that look very clear, like a collision or other type of vehicle accident, but are not.

[0051] According to some embodiments, referring anew to Figures 4A and 4B, applying data extension S220 is, -Applying appropriate rotational extension to the motion sensor S222, and applying a small orientation along a certain axis to the sensor data in order to represent a change in the sensor mounting orientation on the vehicle, - To represent noisy data from a particular sensor, apply appropriate noise enhancement such as Gaussian noise (S223). -Includes at least one of the following: applying a first amplitude scaling data augmentation to represent the perceived intensity change of motion acceleration at different mounting positions of sensors on adjacent MCs, S224.

[0052] According to some embodiments, the above extension is suitable for handling multiple different sensor mounting positions and / or multiple different sensor mounting orientations within a particular sensor mounting area 6, 7.

[0053] In some further embodiments, continuing to refer to Figures 4A and 4B, applying data extension S220 is: - In order to represent the various velocities of the vehicle collision dynamics, apply a second amplitude scaling data extension in S225. - To represent the velocity at which collision dynamics unfold, apply time compression / uncompressed data augmentation S226, and - To represent noise or unwanted data in the signal due to errors, apply segments of appropriate length and position within the signal, which have been scrambled with appropriate intensity, to at least one of the above (S227).

[0054] In some embodiments, the above extension is suitable when a vehicle accident needs to be detected by at least two of sensors 2, 3; 4, and 5, and within a limited time period, in order to be verified and registered.

[0055] Depending on the deployment requirements, two or more of the augmentations may be used. Data augmentation corresponds to a carefully selected data transformation that is applied directly to the initially collected data, based on the requirements of the deployment scenario.

[0056] This disclosure enables the generation of machine learning models that are more resilient to small variations in the position and orientation of vehicle sensors 2, 3; 4, and 5. Data augmentation during the development of the machine learning models first allows for the training of backbone machine learning models using techniques such as, for example, controlled learning. These machine learning models can then be fine-tuned to relevant training data such as collision data, abuse data, and road driving data.

[0057] In some cases, additional datasets resulting from data augmentation are used to cover possible deployment scenarios where the initially collected dataset lacks data. This lack of data may stem, for example, from cost, time, or technical limitations in conducting those experiments. This kind of diverse training data is key to training robust and flexible machine learning models in the future.

[0058] In some embodiments, the additional datasets resulting from data augmentation can also be used to enhance robustness regarding various changes in vehicle speed, vehicle weight, and signal errors due to transient or permanent sensor malfunctions during collision scenarios.

[0059] Figure 2 schematically shows the components of a control unit 200 according to one embodiment, with respect to several functional units.

[0060] The processing circuit 210 is provided using one or more combinations of suitable central processing units (CPUs), multiprocessors, microcontrollers, digital signal processors (DSPs), dedicated hardware accelerators, etc., which can execute software instructions stored in a computer program product in the form of a storage medium 230, for example. The processing circuit 210 may further be provided as at least one application-specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0061] In particular, the processing circuit 210 is configured to cause the control unit 200 to execute a set of operations or steps. These operations or steps have been described above in relation to various radar transceivers and methods. For example, the storage medium 230 may store a set of operations, and the processing circuit 210 may be configured to retrieve the set of operations from the storage medium 230 and cause the control unit 200 to execute the set of operations. The set of operations may be provided as a set of executable instructions. Thus, the processing circuit 210 is configured to execute the methods and operations disclosed herein.

[0062] The storage medium 230 may also include a persistent storage device, which may be any one or a combination of, for example, magnetic memory, optical memory, solid-state memory, or further remotely mounted memory.

[0063] The control unit 200 may further include a communication interface 220 for communicating with at least one other unit. Thus, the communication interface 220 may include one or more transmitters and receivers having analog and digital components and an appropriate number of ports for wired or wireless communication.

[0064] The processing circuit 210 is adapted to control the general operation of the control unit 200, for example, by transmitting data and control signals to an external unit and a storage medium 230, by receiving data and reports from an external unit, and by retrieving data and instructions from the storage medium 230. Other components and related functions of the control unit 200 are omitted in order not to obscure the concepts presented herein.

[0065] Figure 3 shows a computer program product 310 which includes computer executable instructions 320 placed on a computer-readable medium 330 to perform one of the methods disclosed herein.

[0066] In some aspects, the disclosure also relates to a device for evaluating vehicle accident events, a device for communicating information including information from vehicle accident event evaluations, and a device for enabling a computer implementation method for vehicle accident event evaluations, the computer implementation method for rollover evaluation being as described herein. The device comprises a machine learning model, a sensor for collecting data from a user, means for acquiring data, means for processing data, and means for communicating information. The device comprises a computer program and / or computer-readable medium as described herein, for applying the computer implementation method for vehicle accident event evaluations, and for communicating information, the information including information from vehicle accident event evaluations as described herein.

[0067] In some embodiments, a device for evaluating vehicle accident events as described herein is disclosed, the device being provided in a vehicle 1 and comprising sensors 2, 3, 6, 7 having accelerometers and / or gyroscopes.

[0068] In some aspects, the disclosure also relates to a system for evaluating vehicle accident events, a system for communicating information including information from vehicle accident event evaluations, and a system for enabling a computer implementation method for vehicle accident event evaluations. A computer implementation method for vehicle accident event evaluations is described herein.

[0069] The system comprises a machine learning model, sensors 2, 3, 4, and 5, means for collecting data from sensors 2, 3, 4, and 5, means for acquiring the data, means 200 for processing the data, and means for communicating information. The system also comprises a computer program product 310 according to the above.

[0070] The information includes information from vehicle accident event evaluations as described herein, and / or the computer-readable medium 330 comprises a computer program 310 product including computer-readable instructions 320 for applying a computer implementation method for vehicle accident event evaluation and for communicating information, and the information includes information from vehicle accident event evaluations as described herein.

[0071] In some embodiments, a system for vehicle accident event evaluation and information communication is disclosed. The information includes information from vehicle accident event evaluation as disclosed herein. The system includes one or more devices for vehicle accident event evaluation and information communication, and the information includes information from vehicle accident event evaluation as described herein.

[0072] This disclosure relates to a system for training a machine learning model in vehicle accident event evaluation, for communicating information, for enabling the implementation of a machine learning model as described herein, and for preparing training data as described herein. Vehicle accident event evaluation includes vehicle accident event detection, and the system includes a machine learning model, sensors for collecting data, means for acquiring data, means for processing data, and means for communicating information. The system utilizes a computer program 310 as described herein and / or a computer-readable medium 330 as described herein.

Claims

1. A computer implementation method for training a machine learning model in vehicle accident event evaluation including vehicle accident event detection for two-wheeled or three-wheeled vehicles (1), wherein the method is Implementing the aforementioned machine learning model (S100), This includes preparing training data (S200), Preparing training data (S200) The system acquires and automatically learns sensor data generated from sensors (2, 3; 4, 5) located on the vehicle (1) and used to collect data (S210), The process includes applying data augmentation to the collected data in order to artificially increase the amount of data acquired from each sensor (2, 3; 4, 5) so that an augmented dataset is acquired for each sensor (2, 3; 4, 5) (S220), Applying data expansion (S220) is performed for each sensor (2, 3; 4, 5), The method includes transforming the collected data and applying it to a plurality of different sensor mounting positions and / or a plurality of different sensor mounting orientations within a particular sensor mounting area (6, 7) so that a further dataset is obtained (S221), wherein the expanded dataset includes the collected data and the further dataset, and the method is A computer implementation method further comprising training the machine learning model using the extended dataset for each sensor (2, 3; 4, 5) (S300).

2. Applying data extension (S220) - Applying appropriate rotational extension to the motion sensor (S222), which involves applying a small orientation along a specific axis to the sensor data in order to represent a change in the sensor mounting orientation on the vehicle (S222), - Applying appropriate noise enhancement such as Gaussian noise to represent noisy data from a particular sensor (S223), and The computer implementation method according to claim 1, comprising at least one of the following: - Applying a first amplitude scaling data augmentation to represent a change in the perceived intensity of the motion acceleration at different mounting positions of the sensors on the MC that are in close proximity to each other (S224).

3. The computer implementation method according to claim 1 or 2, wherein at least two sensors (2, 3; 4, 5) are positioned within a sensor mounting area (6, 7), and the machine learning model for the sensors (2, 3; 4, 5) is trained using the same extended dataset.

4. The computer implementation method according to claim 3, wherein at least one safety measure is triggered when the sensor data acquired from at least two of the sensors (2, 3; 4, 5) located within the same sensor mounting area (6, 7) indicates the occurrence of an accident within a specific predetermined period, and no safety measure is triggered otherwise.

5. Applying data extension (S220) - To represent various velocities in the collision dynamics of the vehicle, apply a second amplitude scaling data extension (S225). - Applying time compression / uncompressed data augmentation to represent the speed at which collision dynamics unfold (S226), and The computer implementation method according to claim 4, comprising at least one of the following: applying a segment of appropriate length and position within the signal, scrambled with appropriate intensity, to represent noise or unwanted data in the signal due to an error (S227).

6. The computer implementation method according to claim 1, wherein the sensors (2, 3; 4, 5) include at least one of a motion sensor, a 3D sensor, an accelerometer, a gyroscope, and / or a camera.

7. The computer implementation method according to claim 1, wherein the additional dataset is used to cover possible deployment scenarios where the initially collected dataset lacks data.

8. The computer implementation method according to claim 1, wherein the additional dataset is used to enhance robustness with respect to various changes in vehicle speed, vehicle weight, and signal errors due to transient or permanent sensor malfunctions during a collision scenario.

9. A computer program (310) including computer-readable instructions (320) for applying the computer implementation method and / or preparation of training data described in claim 1, and / or a computer-readable medium (330) including the computer program (310).

10. A computer program (310) comprising computer-readable instructions (320) for a machine learning model according to claim 1, and / or a computer-readable medium (330) comprising the computer program (310).

11. A control unit (200) for training a machine learning model, wherein the control unit (200) is adapted to control at least the implementation of the machine learning model described in claim 1 and / or the preparation of the training data described in claim 1.

12. A system for training a machine learning model in vehicle accident event evaluation, for information communication, for enabling the implementation of the machine learning model described in claim 1, and for enabling the preparation of training data described in claim 1, Vehicle accident event evaluation includes vehicle accident event detection, The system includes a machine learning model, a sensor for collecting data, means for acquiring data, means for processing data, and means for communicating information. The system utilizes the computer program (310) described in claim 9 or 10, and / or the computer-readable medium (330) described in claim 9 or 10.

Images