Computer-implemented method for training machine learning models in unexpected event assessment
By training machine learning models and using individual characteristics to divide subject groups, the problem of insufficient applicability of machine learning algorithms in the prior art is solved, and personalized fall detection and protection is achieved, especially real-time airbag triggering in motorcycle accidents.
Patent Information
- Application Number
- CN202480007284.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-02-08
- Filing Date
- 2024-02-07
- Publication Date
- 2025-08-22
AI Technical Summary
It is difficult to develop data-driven machine learning algorithms suitable for different types of people to provide efficient personalized protection in accident detection, especially in motorcycle accidents, where real-time triggering and fall detection of airbag systems are insufficient.
By training machine learning models, using sensor data and individual characteristics, divide subject groups and train personalized machine learning submodels, combining individual data fine-tuning models to improve the accuracy and speed of fall detection.
It realizes high accuracy and rapid fall detection for individual users in different scenarios, improves the efficiency of personalized protection, and is suitable for people walking and riding two-wheeled vehicles.
Smart Images
Figure CN120530441A_ABST
Abstract
Description
[0001] manual
[0002] The present disclosure relates to an improved computer-implemented method for training machine learning models in contingency assessment.
[0003] Accidents can lead to personal injury and injury-related deaths. Accident detection, prediction, prevention, and mitigation are becoming more practical when combined with machine learning methods.
[0004] Accidents include falls, such as those caused by environmental conditions like slippery sidewalks, or even when a person is prone to falls at home or while walking outdoors. Medical treatment associated with falls, particularly hip fractures, can result in significant medical expenses. Significant research has been conducted on detecting falls in real time and using this information to trigger emergency notification systems to notify hospital attendants or next of kin of the person involved in the incident.
[0005] There has also been some recent research work on personal protective equipment that provides active safety protection to the user when a fall is detected, such as triggering a portable airbag system to protect vulnerable parts of the body, such as the hips, from the impact of a fall. The algorithms running on such wearable devices must operate in "hard real-time conditions" to trigger the active safety protection with sufficient lead time for the safety system to be activated and fully protected.
[0006] Accidents can also occur when riding a bicycle, scooter, motorcycle, etc., due to skidding, sliding, loss of control of the vehicle, collision with another vehicle or object, etc. Falls associated with a moving vehicle are generally more sudden and require more rapid attention than falls of pedestrians, such as those caused by tripping or slipping.
[0007] Motorcycle (MC) accidents, in particular, can result in serious injury or be fatal, in part because the rider is more exposed than a driver in an automobile. Additionally, there are no restraints on motorcycles, such as seatbelts in automobiles. There are also other motion dynamics that make motorcycle riders more vulnerable in the event of an accident or crash. Motorcycles with integrated airbag systems are an emerging technology that can reduce the risk of serious injury or fatality in motorcycle accidents, particularly in head-on or slightly angled crashes, thereby helping to prevent the rider from being ejected from the motorcycle as a result of the crash.
[0008] If used effectively, this could save lives. However, detecting a collision within milliseconds and subsequently triggering a signal to inflate an airbag requires a real-time airbag triggering system. Similar technology could also be used in vests with one or more airbags or self-inflating vests, or smart helmets with one or more airbags, suitable for protecting a motorcyclist wearing at least one of the vest and helmet if the rider is thrown from the motorcycle after a collision. Such helmets could also be equipped with one or more devices capable of issuing one or more types of warnings. Other protective clothing components, such as belts, are also contemplated.
[0009] However, there is still room for improvement in this area. In particular, it would be desirable to provide training data for data-driven machine learning algorithms (ie, for data-driven machine learning computer programs) that is tailored to different types of people, both in terms of physical and behavioral characteristics.
[0010] This is achieved using a computer-implemented method for training a machine learning model in human incident assessment, including incident detection. The method includes implementing the machine learning model and preparing training data. Preparing the training data includes obtaining and automatically acquiring sensor data generated by sensors used to collect data from an initial group of subjects, and dividing the group of subjects into subgroups, wherein each subgroup is associated with certain characteristics unique to the subjects in the subgroup.
[0011] The method further includes transmitting information including prepared training data for machine learning model training, training the initial machine learning model using the training data including the training data prepared in fall assessments of subjects in the initial group, and training a plurality of machine learning sub-models using the training data including the training data prepared in fall assessments of subjects in corresponding sub-groups. In this way, one machine learning sub-model is obtained for each sub-group.
[0012] This means that an initial general machine learning model is trained first, and then multiple machine learning sub-models are trained. Subjects in a certain subgroup can utilize corresponding machine learning sub-models that are more suitable for the characteristics and needs of the subjects in the subgroup than the initial machine learning model.
[0013] If the machine learning sub-model is unavailable or unsuitable for some reason, it is always possible to revert to the initial machine learning model for a subject or group of subjects. Therefore, both the initial machine learning model and the machine learning sub-model can be deployed on the device or in the cloud to provide proactive safety protection or some emergency call functionality.
[0014] According to some aspects, the computer-implemented method further comprises preparing individual training data for individual subjects in the corresponding subgroup by obtaining and automatically acquiring sensor data generated by sensors used to collect data. The sensor data is generated during actions of the individual subjects in the corresponding subgroup, the actions comprising activities of daily living (ADLs) comprising real fall events and / or simulated falls. The method further comprises using the individual training data of the training data prepared in the fall assessment for each subject in the subjects in the corresponding subgroup, and training an individual machine learning model for the individual subject using certain features unique to the individual subject in the subgroup.
[0015] This means that new training data can be collected from subjects on an individual basis, enabling fine-tuning of individual machine learning models for the subject in question. The resulting more customized variants of each individual machine learning model improve the accuracy and speed of detecting dangerous situations such as falls or near falls in different scenarios of daily life activities. They can even introduce more fine-grained features and / or newer features for those individual users as part of the severity of falls and / or fall risk assessment in their typical daily routine activities.
[0016] According to some aspects, certain features include at least one of the following:
[0017] - the type of accident,
[0018] - environmental conditions,
[0019] - Subject age,
[0020] - Subject's height,
[0021] - Subject's weight,
[0022] - the subject's gait pattern,
[0023] - medical condition of the subject,
[0024] - driving characteristics, and
[0025] -Vehicle type.
[0026] This means that many types of data can be used to fine-tune individual machine learning models based on each subject's individual personal data.
[0027] 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. In other words, many types of well-known sensors can be used to obtain the required data.
[0028] According to some aspects, the subject is a person walking, and the accident is a fall caused by stumbling and / or slipping. According to some other aspects, the subject is a two-wheeled vehicle rider, and the accident is caused by skidding, sliding, losing control of the vehicle, and / or colliding with another vehicle or object.
[0029] This means that the present disclosure is applicable to both people on foot and people riding two-wheeled vehicles, and to different types of related accidents.
[0030] According to some aspects, preparing the training data includes obtaining and automatically acquiring current and past historical data, the historical data including user geolocation information and / or weather information from a positioning system.
[0031] This raises the possibility of performing personalization at an even more granular level.
[0032] The present disclosure also relates to a computer program, a control unit and a system associated with the above advantages.
[0033] Therefore, it should be understood that all computer-implemented methods, computer programs, and computer-readable media as described herein and according to the present disclosure can be implemented in all hardware such as control unit devices and apparatuses as described herein, as well as in systems as described herein. Then, all hardware such as control unit devices as described herein, as well as in systems as described herein, are arranged to execute the computer-implemented methods and computer programs, thereby obtaining the same advantages and effects as discussed for the computer-implemented methods herein. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] The present disclosure will now be described in more detail with reference to the accompanying drawings, in which:
[0035] Figure 1 A walking user wearing a sensor is schematically shown;
[0036] Figure 2 schematically illustrates a user riding a motorcycle and wearing sensors;
[0037] Figure 3 The subject groups are shown schematically;
[0038] Figure 4 A control unit is schematically shown;
[0039] Figure 5 An exemplary computer program product is shown; and
[0040] Figure 6 A flow chart of a method according to the present disclosure is shown.
[0041] Figure 7A flow chart of a method according to an example is shown. DETAILED DESCRIPTION
[0042] Various aspects of the present disclosure will now be described more fully below with reference to the accompanying drawings. However, the various devices, apparatuses, systems, computer programs, and methods disclosed herein may be implemented in many different forms and should not be construed as being limited to only the aspects described herein. Like reference numerals in the drawings represent like elements throughout the text.
[0043] The terms used herein are only used to describe various aspects of the present disclosure and are not intended to limit the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms as well.
[0044] refer to Figure 3 and Figure 6 The present disclosure relates to a computer-implemented method for training a machine learning model in human accident assessment including accident detection. The method comprises implementing S100 a machine learning model and preparing S200 training data. Figure 1 and Figure 2 , preparing S200 training data then includes obtaining and automatically acquiring S210 sensor data generated by sensors 2, 3; 6, 7 for collecting data from the initial group 200 of subjects 201a-201c.
[0045] According to some aspects, the contingency assessment further includes contingency characterization and / or calculation of contingency risk probability.
[0046] refer to Figure 1 , there is a subject 1 who is a user wearing at least one sensor device 2, 3, where the sensor device is a wrist sensor device 2 and a waist sensor device 3. In this case, the user 1 is walking.
[0047] refer to Figure 2 , there is a subject 1, user 1, wearing at least one sensor device, here a wrist sensor device 6 and a vest sensor device 7, the vest sensor device being included in a protective garment such as a protective vest 10. In this case, user 1 is riding a motorcycle 8. User 1 may wear sensors that are directly or indirectly attached to the garment worn by user 1.
[0048] The present disclosure is applicable to users who are traveling on foot or on or in any kind of vehicle, such as a bicycle, a motorcycle 8, a car, or the like.
[0049] Sensor data is typically, but not limited to, collected (i.e., obtained) from sensors 2, 3, 6, 7, which include, for example, motion sensors on (e.g., mounted on) a wearable device, such as a 3D sensor, such as an accelerometer for measuring 3-dimensional acceleration, and a gyroscope for measuring 3-dimensional rotational velocity from a person / subject 1 (i.e., a training subject) and / or user 1 wearing such wearable device 2, 3, 6, 7 according to the present disclosure.
[0050] During experiments and other means of collecting data, i.e., during the behavior of training subjects wearing wearable devices (e.g., 3D sensors and, where appropriate, edge devices) according to the present disclosure, data is generated by these 3D sensors in the form of multivariate time series at specific frequencies typically in the range of 25 Hz to 200 Hz, which can be obtained (i.e., collected and buffered at the edge device) before being moved to an offline storage device or uploaded to a cloud device for further processing.
[0051] The acquired data may include data obtained from sources other than sensors. The acquired data may involve current data and past historical data. This may include the user's geolocation information from a GPS system and weather information. According to some aspects, preparing S200 training data includes acquiring and automatically obtaining S211 current and past historical data, including the user's geolocation information and / or weather information from a positioning system. This increases the possibility of performing personalization at an even more granular level.
[0052] According to some aspects, such as Figure 2 As shown, user 1 is wearing a device 9 for providing the location of user 1 and transmitting said location. Such device 9 may include any type of suitable positioning system, such as, for example, GPS or GNSS (Global Navigation Satellite System), as well as any type of wireless communication system. The device may be included in a vest 10 or any suitable type of clothing or wearable device.
[0053] For example, incident detection using incident detection algorithms relies on detecting patterns of motion that are characteristic of incidents such as falls. However, due to real-world heterogeneity, fall patterns often vary depending on the type of fall, environmental conditions, and individual age, height, weight, gait patterns, and medical conditions, as well as the driver's driving characteristics and vehicle type. Therefore, it is impossible to develop a universal incident detection method that works optimally for every planner. Individual traits and differences contribute significantly to this heterogeneity, which in turn motivates the need to optimize incident detection based on personalized aspects.
[0054] Therefore, according to the present disclosure and with reference to Figure 3 and Figure 6 , the method further comprises dividing S220 the initial group 200 of subjects 201a-201c into subgroups 200a-200h, wherein each subgroup 200a-200h is associated with certain characteristics that are unique to the subjects 201a-201c in the subgroup; and transmitting S300 information comprising the prepared training data for training of the machine learning model. Figure 3 For clarity, only a few subjects 201a-201c in the first subgroup 200a are indicated. Of course, there are corresponding multiple subjects in all subgroups 200a-200h, only a few subjects are shown without indication in other subgroups 200b-200h, and any suitable number of subjects 201a-201c may be present in each subgroup 200a-200h. Certain features include certain characteristics.
[0055] For clarity, the initial group 200 is shown here as a circle, and the subgroups 200a-200h are shown as circular sectors that divide the circle into a plurality of sections, each section constituting a subgroup 200a-200h. In this example, there are eight equal sections or subgroups 200a-200h, but this is of course merely an example, and the number of subgroups can vary. The circular shape of the initial group 200 and the circular sector shapes defining the subgroups are chosen for illustrative purposes only and should not be considered limiting in any way.
[0056] The method also includes training S400 an initial machine learning model using training data of the training data prepared in the fall assessment of subjects 201a-201c included in the initial group 200, and training S500 multiple machine learning sub-models using training data of the training data prepared in the fall assessment of subjects 201a-201c included in the corresponding sub-groups 200a-200h, so that each sub-group 200a-200h obtains one machine learning sub-model.
[0057] This means first training the initial general machine learning model and then training multiple machine learning sub-models. Of course, this training can be performed in another order or at least partially in parallel.
[0058] It should be noted that some method steps may be performed in another order than that described above. For example, training S400 the initial machine learning model may be performed at any time after obtaining and automatically acquiring S210 sensor data generated by sensors 2, 3; 6, 7 for collecting data from the initial group 200 of subjects 201a-201c.
[0059] Then, subjects 201a-201c in a certain subgroup 200a can utilize a corresponding machine learning sub-model that is more suitable for the characteristics and needs of the subjects in that subgroup than the initial machine learning model. If the machine learning sub-model is unavailable or unsuitable for some reason, it is always possible to revert to the initial machine learning model for a certain subject or group of subjects. Therefore, both the initial machine learning model and the machine learning sub-model can be used for deployment on the device or in the cloud to provide active safety protection or some emergency call functionality.
[0060] According to some aspects, initial data is collected from subjects 201a-201c who have fallen naturally, or in a controlled environment (i.e., fall data), or from daily activity data of a representative group of subjects who may be performing various daily activities (such as elderly people from a nursing home). This data can then be used to design appropriate data augmentation for time series data from motion sensors, and this augmented data can then be used to develop an initial machine learning model, such development utilizing techniques such as contrastive learning.
[0061] Furthermore, it is also desirable to have a machine learning model that is even more suitable for the subjects 201a-201c. Therefore, according to some aspects, the computer-implemented method further comprises preparing S600 individual training data for the individual subjects 201a-201c in the corresponding subgroup 200a by obtaining and automatically acquiring S610 sensor data generated by the sensors 2,3; 6,7 used to collect data.
[0062] The sensor data is generated during the actions of individual subjects 201a-201c in corresponding subgroups 200a-200h, wherein the actions include activities of daily living (ADL), which include real fall events and / or simulated falls, which in turn include falls or collisions involving motorcycles or similar vehicles of which the subjects are drivers.
[0063] The method further includes using individual training data prepared in the fall assessment of each subject 201a-201c in the corresponding subgroup 200a-200h, and training S700 an individual machine learning model for the individual subject 201a-201c using certain features unique to the individual subject 201a-201c in the subgroup 200a-200h. As described above, the certain features include certain characteristics.
[0064] This means that new training data can be collected from subjects 201a-201c on an individual basis, enabling fine-tuning of the individual machine learning model for the subject 201a-201c in question. According to one example, this can be performed for certain individual subjects 201a-201c in the subgroup 201 (e.g., premium customers) based on data collected from those specific users or users in the premium group. In other words, not all subjects in the subgroup need to be used for the individual machine learning model, only certain subjects.
[0065] The resulting more customized variants of each individual machine learning model improve the accuracy and speed of detecting dangerous situations such as falls or near falls in different scenarios of daily life activities. They can even introduce more fine-grained features and / or newer features for those individual users as part of the severity of falls and / or fall risk assessment in their typical daily routine activities.
[0066] Here, fine-grained data such as personal body profile attributes weight, BMI, age, daily activities list, medical history are collected and according to some aspects the subject 201a-201c under consideration is asked to perform a set of predefined activities for a defined number of times.
[0067] Based on the user data analysis, the subjects 201a-201c can be mapped to the previously defined subgroups 200a-200h and the corresponding machine learning sub-models can be selected, which then form the basis for fine-tuning the individual machine learning models based on the individual personal data of each subject.
[0068] Note that the individual machine learning models defined above may also be updated over time as new data becomes available or if the subject's movement pattern set changes in a significant manner since the corresponding individual machine learning model was last developed / updated.
[0069] When a fully personalized individual machine learning model is not yet available for a subject 201a-201c, the subject 201a-201c can utilize the corresponding machine learning sub-model based on correlation. If this is not possible, the individual can resort to using the initial machine learning model.
[0070] According to some aspects, the present disclosure is therefore directed to a method for performing personalization at a more granular level. Starting with data from all subjects available for training, subgroups or clusters of data are created using sensor data similarities and metadata such as user height / age, weight, medical history, etc., which are then used to train machine learning submodels.
[0071] When a new subject enters the system, he or she does not have any personal data, such as personal data related to ADLs or falls, falls or collisions involving motorcycles or similar vehicles. The system will use metadata correlation to determine the most relevant subgroups 200a-200h and use the associated machine learning sub-models for that user until a sufficient amount of individual data is collected to enable training of an individual machine learning model for that subject.
[0072] It may take some time until a sufficient amount of individual data is extracted, during which time the subject will still receive a moderate level of personalization due to the use of the relevant machine learning sub-model. Only if no relevant machine learning sub-model can be found for the new subject should one resort to using the initial machine learning model.
[0073] All new data is collected for each new subject that enters the system and then propagated to the corresponding subgroup 200a-200h to which the subject belongs. This means that each subject's newly collected data also improves the performance of other subjects with similar characteristics in the same subgroup 200a-200h. This data can also be propagated upwards in the hierarchy to the general layer to improve the initial machine learning model. In this way, all shared data from the subjects is utilized, with the fully personalized individual machine learning model placing much more weight on samples belonging to the specific subject to which the individual machine learning model belongs during training.
[0074] According to some aspects, certain features include at least one of the following:
[0075] - the type of accident,
[0076] - environmental conditions,
[0077] - Subject age,
[0078] - Subject's height,
[0079] - Subject's weight,
[0080] - the subject's gait pattern,
[0081] - medical condition of the subject,
[0082] - driving characteristics, and
[0083] -Vehicle type.
[0084] This means that many types of data can be used to fine-tune individual machine learning models based on each subject's individual personal data.
[0085] According to some aspects, the sensors 2 , 3 ; 6 , 7 include at least one of a motion sensor, a 3D sensor, an accelerometer, a gyroscope, and / or a camera.
[0086] In other words, many types of well-known sensors can be used to obtain the required data.
[0087] According to some aspects, the subjects 1, 201a-201e are walking persons, and wherein the accidental event is a fall event caused by tripping and / or slipping.
[0088] According to some aspects, subjects 1, 201a-201e are two-wheeled vehicle riders, and wherein the accident occurs due to skidding, spinning, loss of control of the vehicle, and / or collision with another vehicle or object.
[0089] This means that the present disclosure is applicable to both people on foot and people riding two-wheeled vehicles, and to different types of related accidents.
[0090] Now refer to Figure 3 and Figure 7 A non-limiting example of describing a human accident assessment is a human fall assessment.
[0091] Following initial data collection 700, subject clustering 710 follows, where the subjects are divided into subgroups 200a-200h. Initial machine learning models are then trained 711 in parallel using domain-specific techniques through data augmentation 712. Data augmentation 712 is related to the model training aspect of the model, as indicated by the arrow pointing to initial training 711. The results are transferred 713 to subgroups 200a-200h. Machine learning submodels 714 are generated, along with the initial machine learning model 715, which are then output 716 and deployed 717.
[0092] For the individual machine learning model, subject data is input 720, and the subject performs a predefined movement 721, such as walking, sitting / standing, lying down, climbing stairs, jogging (if possible), etc. in a defined trajectory. The movement can be repeated at certain intervals (such as, for example, months or weeks). Alternatively or additionally, when the individual machine learning model is adjusted using regular daily activities, the movement can also be automatically detected.
[0093] The subject data is then analyzed 722, for example, using a device or in an application. The subject data is then mapped 723 to corresponding subgroups 200a-200h using the machine learning sub-models 714. Each individual machine learning model is then adjusted 724 to fit the individual subject, and the individual machine learning models are output 725 and deployed 717.
[0094] For a two-wheeled vehicle accident assessment, the steps discussed above are largely the same, with some modifications. Specifically, output model 716 is the initial vehicle accident detection machine learning model. Predefined motions 721 can be, for example, predefined vehicle maneuvers, such as the subject driving on different types of roads, driving in a straight line, driving around a roundabout, stopping at a traffic light, driving aggressively, and so on.
[0095] Figure 4 Components of the control unit 400 are schematically shown in terms of a number of functional units according to one embodiment.
[0096] The processing circuit 410 is provided using any combination of one or more of a suitable central processing unit (CPU), multiprocessor, microcontroller, digital signal processor (DSP), dedicated hardware accelerator, etc., capable of executing software instructions stored, for example, in a computer program product in the form of storage medium 430. The processing circuit 410 may also be provided as at least one application-specific integrated circuit (ASIC) or field-programmable gate array (FPGA).
[0097] In particular, processing circuitry 410 is configured to cause control unit 400 to perform a set of operations or steps. These operations or steps are discussed above in conjunction with various radar transceivers and methods. For example, storage medium 430 may store a set of operations, and processing circuitry 410 may be configured to retrieve the set of operations from storage medium 430 to cause control unit 400 to perform the set of operations. The set of operations may be provided as a set of executable instructions. Thus, processing circuitry 410 is configured to perform the methods and operations disclosed herein.
[0098] The storage medium 430 may also include permanent storage, such as any single memory or combination of magnetic memory, optical memory, solid-state memory, or even remotely mounted memory.
[0099] The control unit 400 may also include a communication interface 420 for communicating with at least one other unit. As such, the communication interface 420 may include one or more transmitters and receivers, including analog and digital components and an appropriate number of ports for wired or wireless communication.
[0100] The processing circuit 410 is adapted to control the general operation of the control unit 400, for example, by sending data and control signals to external units and a storage medium 430, by receiving data and reports from external units, and by retrieving data and instructions from the storage medium 430. Other components of the control unit 400 and related functions are omitted so as not to obscure the concepts presented herein.
[0101] The control unit 400 may be implemented by one single product or by several products that together form a control unit arrangement.The control unit may be implemented at least at a remote server, for example in a cloud service.
[0102] Figure 5 A computer program product 510 is shown comprising computer-executable instructions 520 disposed on a computer-readable medium 530 for performing any of the methods disclosed herein.
[0103] According to some aspects, the present disclosure also relates to a device for personnel accident assessment, for transmitting information and for implementing a computer-implemented method for personnel accident assessment, wherein the information includes information from personnel accident assessment, wherein the computer-implemented method for fall assessment is as described herein. The device includes a machine learning model, a sensor that collects data from a user, a device for obtaining data, a device for processing data, and a device for transmitting information. The device includes a computer program and / or a computer-readable medium, both of which are as described herein, and for applying the computer-implemented method for personnel accident assessment, and for transmitting information, wherein the information includes information from the personnel accident assessment as described herein.
[0104] According to some aspects, a device for personal accident assessment according to the present disclosure as described herein is disclosed, wherein the device is wearable by a user 1 and comprises sensors 2, 3, 6, 7 comprising one or more accelerometers and / or one or more gyroscopes.
[0105] The present disclosure also relates to a system for training a machine learning model in human incident assessment, for transmitting information, for enabling implementation of a machine learning model as described herein, and for enabling preparation of training data as described herein. Human incident assessment includes incident detection.
[0106] The system includes a machine learning model, a training subject 1, a sensor that collects data from the training subject, a device for obtaining the data, a device for processing the data, and a device for transmitting information. The system utilizes a computer program 510 as described herein and / or a computer readable medium 530 as described herein.
[0107] According to some aspects, the information includes information from a human contingency assessment as described herein and / or the computer-readable medium 530 includes a computer program 510 product comprising computer-readable instructions 520 for applying a computer-implemented method for human contingency assessment and for transmitting information, wherein the information includes information from a human contingency assessment as described herein.
[0108] According to some aspects, a system for personnel incident assessment and for transmitting information is disclosed. Information including information from a personnel incident assessment according to the present disclosure as described herein is disclosed. The system includes one or more devices for personnel incident assessment and for transmitting information, wherein the information includes information from a personnel incident assessment as described herein.
Claims
1. A computer-implemented method for training a machine learning model in human incident assessment including incident detection, wherein the method comprises implementing (S100) the machine learning model, and Prepare (S200) training data The preparation (S200) of training data includes obtaining and automatically acquiring (S210) sensor data generated by sensors (2, 3; 6, 7) for collecting data from an initial group (200) of subjects (201a-201c); as well as dividing (S220) the group (200) of the subjects (201a-201c) into subgroups (200a-200h), wherein each subgroup (200a-200h) is associated with certain characteristics that are unique to the subjects (201a-201c) in the subgroup (200a-200h); wherein the computer-implemented method further comprises transmitting ( S300 ) information including prepared training data for machine learning model training; training (S400) an initial machine learning model using training data prepared from training data included in fall assessments of the subjects in the initial group (200); A plurality of machine learning sub-models are trained (S500) using the training data prepared in the fall assessment of the subjects (201a-201c) included in the corresponding sub-groups (200a-200h) so that each sub-group (200a-200h) obtains one machine learning sub-model.
2. The computer-implemented method of claim 1 , further comprising: Individual training data for individual subjects (201a-201c) in the corresponding subgroup (200a-200h) are prepared (S600) by the following operations: obtaining and automatically acquiring (S610) sensor data generated by sensors (2, 3; 6, 7) for collecting data, wherein the sensor data is generated during actions of individual subjects (201a-201c) in the corresponding subgroup (200a-200h), the actions comprising activities of daily living (ADL), the activities of daily living comprising real fall events and / or simulated falls; and The individual training data of the training data prepared in the fall assessment of each subject in the subjects (201a-201c) in the corresponding subgroup (200a-200h) is used, and the individual machine learning model for the individual subject (201a-201c) in the subgroup (200a-200h) is trained (S700) using certain features unique to the individual subject (201a-201c) in the subgroup (200a-200h).
3. The computer-implemented method of any one of claims 1 or 2, wherein the certain characteristics include at least one of the following: - the type of accident, - environmental conditions, - Subject age, - Subject's height, - Subject's weight, - the subject's gait pattern, - medical condition of the subject, - driving characteristics, and -Vehicle type.
4. The computer-implemented method according to any of the preceding claims, wherein the sensor (2, 3; 6, 7) comprises at least one of a motion sensor, a 3D sensor, an accelerometer, a gyroscope and / or a camera.
5. A computer-implemented method according to any one of the preceding claims, wherein the subject (1; 201a-201c) is a walking person, and wherein the accidental event is a fall event caused by tripping and / or slipping.
6. A computer-implemented method according to any one of claims 1 to 4, wherein the subject (1; 201a-201c) is a two-wheeled vehicle rider, and wherein the accident is caused by skidding, sliding, loss of control of the vehicle and / or collision with another vehicle or object.
7. A computer-implemented method according to any one of the preceding claims, wherein preparing (S200) training data includes obtaining and automatically acquiring (S211) current and past historical data, wherein the current and past historical data include user geolocation information and / or weather information from a positioning system.
8. A computer program (510), and / or a computer-readable medium (530) comprising said computer program (510), said computer program comprising computer-readable instructions (520) for applying the computer-implemented method and / or the preparation of training data according to any one of claims 1 to 7.
9. A computer program (510), and / or a computer-readable medium (530) comprising the computer program (510), the computer program comprising computer-readable instructions (520) for a machine learning model according to any one of claims 1 to 7.
10. A control unit (400) for training a machine learning model, the control unit being adapted to at least control the implementation of the machine learning model according to any one of claims 1 to 7, and / or the preparation of training data according to any one of claims 1 to 7.
11. A system for training a machine learning model in personnel accident assessment, for transmitting information, for enabling the implementation of a machine learning model according to any one of claims 1 to 7, and for enabling the preparation of training data according to any one of claims 1 to 7; The personnel accident assessment includes accident detection; wherein the system comprises a machine learning model, a training subject (1), a sensor for collecting data from the training subject, a device for obtaining the data, a device for processing the data, and a device for transmitting information; Wherein the system utilizes a computer program (510) according to claim 8 or 9, and / or a computer readable medium (530) according to claim 8 or 9.