Skier falling detection method and system based on hierarchical learning

Abstract

The invention belongs to the field of behavior detection, and provides a skier falling detection method and system based on hierarchical learning, and the method comprises the steps: collecting multi-source motion data and personnel information of a skier, and carrying out the processing of the data, and obtaining time sequence feature data and personnel feature data; segmenting the time sequence characteristic data by using a window with a preset length, and marking according to a fall identifier to divide the time sequence characteristic data into fall data and non-fall data; screening non-tumble data by adopting a multi-level threshold method based on speed and attitude variation, and retaining data which are judged as high-risk behaviors; merging the falling data, the high-risk behavior data and data generated through data enhancement to form a model training data set; training the hierarchical learning network model by using the model training data set; and the trained hierarchical learning network model is utilized to identify real-time motion data, and skier falling detection is realized. The problems that in the prior art, generalization performance is poor, and individuation characteristics are not achieved are solved.

Description

Technical Field

[0001] This invention belongs to the field of behavior detection, and in particular relates to a method and system for detecting skier falls based on hierarchical learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As a new trend in national fitness and leisure activities, ice and snow culture is gradually integrating into people's lives. At the same time, skiing safety faces significant challenges: on the one hand, it requires continuous improvement of ski resort construction and enhancement of skiers' safety awareness; on the other hand, accident detection is crucial, as it helps pinpoint the problems that lead to accidents, thus enabling precise optimization. Falls are the most common type of skiing accident.

[0004] Deep learning methods based on visual information have proven to be effective in various tasks. However, they have significant limitations in practical applications: on the one hand, due to the vast area of ​​ski resorts, a large number of cameras need to be deployed to achieve comprehensive monitoring, which significantly increases system costs; on the other hand, skiing is often conducted in a variable outdoor environment, and weather conditions (snowfall, fog, light changes, etc.) can easily interfere with the quality of visual data, further affecting recognition performance.

[0005] In contrast, Micro-Electro-Mechanical Systems (MEMS) have gradually attracted widespread attention from researchers due to their advantages such as small size, light weight, low cost, and ease of integration. Although MEMS have shown potential in motion monitoring, research on fall detection in skiing is still limited. Existing methods are mostly based on thresholding, relying on experience to select features and design corresponding thresholds. Their generalization performance is poor, making them difficult to apply to a wide range of complex scenarios, and they lack personalized features. Furthermore, their detection accuracy and robustness are insufficient to meet the needs of practical applications. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a skier fall detection method and system based on hierarchical learning. It integrates an inertial sensor, magnetometer, GPS, and processing unit into a wearable device, such as a ski suit, which is responsible for data collection and storage. The collected data is analyzed and processed on the terminal device to realize fall detection and other downstream tasks. At the same time, a hierarchical learning model is proposed to reduce the performance impact caused by class imbalance.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a skier fall detection method based on hierarchical learning, comprising: Constructing the model training dataset: Collect multi-source motion data and personnel information of skiers, and process them separately to obtain time-series feature data and personnel feature data; use a preset length window to segment the time-series feature data and label it according to fall indicators, dividing it into fall data and non-fall data; for non-fall data, use a multi-level threshold method based on speed and posture changes to filter and retain data judged as high-risk behaviors; merge the fall data, high-risk behavior data, and data augmented data to form the model training dataset; A hierarchical learning network model is constructed: the hierarchical learning network model includes a one-dimensional convolutional layer, an attention mechanism layer, a bidirectional long short-term memory network layer, and a fully connected layer; the hierarchical learning network model is trained using the model training dataset, and Focal Loss is used as the loss function; the trained hierarchical learning network model is used to identify real-time motion data to achieve skier fall detection.

[0008] As a further technical feature, the personnel information includes skiing experience, selected ski slope level, ski type, and gender; the personnel information is incorporated into the feature data using a two-digit encoding method.

[0009] As a further technical feature, during the ski season, skiers wear an integrated sensing device in the interscapular region for natural skiing, and manually trigger the device to record the current time as a fall marker when a fall occurs during skiing; the integrated sensing device integrates an inertial sensor, a magnetometer, GPS, and a processing unit.

[0010] As a further technical feature, the processing of the multi-source motion data also includes: Attitude information is obtained by using quaternion error Kalman filtering or complementary filtering; The acceleration in the sensor coordinate system is converted to the navigation reference system using the attitude information; The converted acceleration is integrated to obtain a velocity estimate, and the integrated velocity is corrected by fusing the velocity information output by GPS with complementary filtering to eliminate integral drift.

[0011] As a further technical feature, the multi-level threshold method includes: Calculate the resultant velocity in the navigation reference frame at the current moment and compare it with the set velocity threshold; If the combined speed is greater than the set speed threshold, proceed to the next step; otherwise, filter. If the cumulative time is greater than the set speed threshold for T consecutive non-fall data points, proceed to the next step; otherwise, filter. Calculate the attitude change in each direction within the window and compare it with the set attitude change threshold; If the change in attitude in each direction exceeds the set attitude change threshold, it is considered high-risk behavior data and is retained; otherwise, it is filtered out.

[0012] As a further technical feature, when a window of data is determined to be a high-risk behavior, the data from multiple adjacent windows before and after it are retained in the model training dataset to maintain data continuity.

[0013] As a further technical feature, the data augmentation method includes: adding noise to the fall data, scaling the amplitude, splicing the data before the fall data with the non-fall data, or modifying one or more of the personnel information features.

[0014] A second aspect of the present invention provides a skier fall detection system based on hierarchical learning, comprising: The dataset construction module is configured to: construct the model training dataset by collecting multi-source motion data and personnel information of skiers, processing them to obtain time-series feature data and personnel feature data; segmenting the time-series feature data using a preset-length window and labeling it according to fall indicators, dividing it into fall data and non-fall data; filtering the non-fall data using a multi-level threshold method based on speed and posture changes, retaining data judged as high-risk behaviors; merging the fall data, high-risk behavior data, and data augmented to form the model training dataset. The model building module is configured to: construct a hierarchical learning network model, which sequentially includes a one-dimensional convolutional layer, an attention mechanism layer, a bidirectional long short-term memory network layer, and a fully connected layer; train the hierarchical learning network model using the model training dataset and employ Focal Loss as the loss function; and use the trained hierarchical learning network model to identify real-time motion data to achieve skier fall detection.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a skier fall detection method based on hierarchical learning as described in the first aspect of the present invention.

[0016] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a skier fall detection method based on hierarchical learning as described in the first aspect of the present invention.

[0017] The above one or more technical solutions have the following beneficial effects: This invention focuses on the fall detection problem in skiing scenarios. It incorporates personnel information as feature data into the model, enriching the data features used in existing research. By integrating multi-dimensional personnel information of skiers, the model can more accurately capture individual differences among skiers, thereby effectively improving the accuracy and adaptability of fall detection. For example, beginners, due to their less-than-skilled skiing, may be more prone to falling than experienced skiers at the same speed and with the same movements. Incorporating personnel information allows the model to better identify this situation and make accurate judgments.

[0018] This invention addresses the class imbalance problem in this task by using a multi-level thresholding method. Through hierarchical structure design and data augmentation, it effectively improves model performance while significantly reducing computational load.

[0019] This invention provides a portable wearable device that is applicable to various scenarios, such as ski protection, ski resort safety optimization, and skier planning. It does not require visual information assistance and provides an event recording function, thus eliminating reliance on visual information.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a schematic diagram of the method flow of the first embodiment; Figure 2 This is a schematic diagram of the data processing flow of the first embodiment; Figure 3 This is a schematic diagram of the multi-level threshold method in the first embodiment; Figure 4 This is a schematic diagram of the fixed position of the data acquisition device in the first embodiment. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] The overall concept proposed in this invention is as follows: In data collected during skiing, non-fall data constitutes the vast majority, while fall data constitutes a small portion, indicating a class imbalance problem. This imbalance can cause the model to favor the class with more data during training, leading to overfitting. Simultaneously, the class with less data provides less feature information, preventing the model from truly learning its features and thus reducing its generalization performance. To address this issue, this invention proposes a hierarchical learning model to mitigate the performance impact of class imbalance.

[0027] Example 1 like Figure 1 As shown, this embodiment discloses a skier fall detection method based on hierarchical learning; including: S1: Construct the model training dataset: S1.1: Collect multi-source motion data and personnel information of skiers, and process them separately to obtain feature data; specifically including: In this embodiment, the skiers participating in the experiment wore a specially designed data acquisition device. Inertial sensors, magnetometers, GPS, and a processing unit were arranged on the same circuit and integrated into this data acquisition device, which was worn in the interscapular region of the body. Figure 4 As shown in the figure. The inertial sensor includes a three-axis accelerometer and a three-axis gyroscope.

[0028] Staff must inform skiers of the purpose of the experiment in advance and obtain their consent before conducting the experiment.

[0029] S1.1.1: Arrange for skiers to wear the device, perform skiing activities normally, and collect multi-source motion data until the skiers leave the ski resort and stop data collection.

[0030] Data collection spans one ski season, typically from mid-November to mid-March of the following year, totaling four months. Additionally, this data collection device provides a function key; pressing and holding this key records the current time as a fall marker. This function is required to assist skiers when they fall, except in special circumstances. Falls are categorized as follows: Type 1: Falling due to loss of balance caused by insufficient skill or experience; Type 2: Falling due to loss of balance caused by collision with other skiers; Type 3: Falling due to loss of balance caused by collision with ski slope facilities; Type 4: Falling due to loss of balance caused by improperly worn ski equipment, resulting in equipment becoming loose or falling off.

[0031] S1.1.2: Staff must record the skier's personal information, including skiing experience, selected ski slope level, ski type, and gender. Skiing experience includes beginner, intermediate, advanced, and above; selected ski slope level includes beginner, intermediate, and advanced; and ski type includes snowboard and ski.

[0032] It's important to know that the probability of a fall is significantly correlated with the skier's personal information; for example, beginners are much more likely to fall than experienced skiers. Therefore, incorporating personal information can enrich feature representations and improve task recognition performance.

[0033] S1.1.3: Export the multi-source motion data collected by the device, including three-axis acceleration, three-axis gyroscope, three-axis magnetic field, latitude and longitude, altitude, speed, and fall marker data. Among them, the fall marker data is used as supervisory information and is not feature data.

[0034] S1.1.4: Using triaxial acceleration, triaxial gyroscope, and triaxial magnetic field data, attitude calculation is performed using quaternion error Kalman filtering to obtain attitude information; quaternion error Kalman filtering can also be replaced by complementary filtering or other filtering methods.

[0035] S1.1.5: Utilize attitude information to convert acceleration from the sensor coordinate system to the navigation reference system; such as... Figure 2 As shown, specifically: The acceleration information in the sensor coordinate system is transformed into the navigation reference system using the calculated attitude information, so that the acceleration information in the sensor coordinate system and the velocity information output by GPS are mapped to the same coordinate system.

[0036] S1.1.6: The converted acceleration is integrated to obtain a velocity estimate, and complementary filtering is used to fuse the velocity information from the GPS output to correct the integrated velocity, thereby eliminating integration drift. Specifically: Velocity information is obtained by integrating the acceleration information from the navigation reference frame. Directly integrating the acceleration will accumulate and drift over time, while GPS output velocity information is absolutely accurate and does not drift over time. Therefore, complementary filtering is used to correct the integrated velocity characteristic using the GPS output velocity information. Complementary filtering can also be replaced by Kalman filtering or other filtering methods. The commonly used ENU (East-North-South) reference frame is selected as the reference coordinate system, but other commonly used navigation reference frames can also be chosen.

[0037] S1.1.7: By concatenating the multi-source motion data processed above, a temporal feature data with 18 dimensions can be obtained. , In addition, personnel information is also used as feature data. Skiing experience is coded using two digits: 01 for beginner, 10 for intermediate, and 11 for advanced and above. Similarly, ski slope level is coded using two digits: 01 for beginner, 10 for intermediate, and 11 for advanced. Ski type and gender are coded using binary numbers: 0 for snowboard and 1 for skis; 0 for female and 1 for male. Finally, the data is concatenated to obtain personnel feature data with a feature dimension of 6. , .

[0038] S1.2: The temporal feature data is segmented using a preset length window and labeled according to fall indicators, dividing it into fall data and non-fall data; specifically: S1.2.1: Segment the time-series feature data using a window of size 100 and a stride of 50, construct the dataset, and obtain each data point as... .

[0039] S1.2.2: Label the segmented data. Since the learning task is fall detection, the data can be divided into two categories: fallen and non-fall. The fall labels are derived from the fall marker data. The fall marker data recording time is synchronized with the data collection time, and kinematic information is combined to confirm the fall data labels for the corresponding data. The labels for other data are non-fall data.

[0040] S1.3: For non-fall data, a multi-level threshold method based on speed and posture changes is used for filtering, retaining data judged as high-risk behaviors; specifically: A serious problem exists in this task: because falling is a low-probability event, the constructed dataset will contain the vast majority of non-fall category data. This problem will significantly degrade model performance and waste more computational resources.

[0041] To address the aforementioned issues, after data segmentation, fall-related data is directly added to the dataset used for model training. Non-fall-related data is first filtered using a multi-level thresholding method to remove most of the data. The remaining data is then added to the model training dataset as high-risk behavior or fall-like data. For example... Figure 3 As shown, the multi-level threshold method includes the following steps: S1.3.1: Calculate the net velocity of the reference frame at time t.

[0042] In the formula, , , Represents velocity components in different directions.

[0043] S1.3.2: Compare the resultant velocity v with the set velocity threshold v*. If v > v*, proceed to the next step; otherwise, filter it out as non-fall data.

[0044] S1.3.3: Cumulative duration. If the sum of the velocities of T consecutive non-fall data points is greater than v*, proceed to the next step; otherwise, filter them out as non-fall data points.

[0045] S1.3.4: Calculate the attitude change in each direction within the window. :

[0046] In the formula, , These represent the attitude feature data at time i and time j, respectively. S1.3.5: Compare the attitude change in each direction with a set threshold. Compare, if greater than If it is considered a high-risk behavior, it is considered a non-fall data and filtered out.

[0047] It should be noted that when a behavior is judged to be high-risk, the data from the first 5 windows will be included in the dataset even if it is not considered a fall, in order to ensure data continuity; similarly, the data from the last 5 windows should also be retained.

[0048] S1.4: Combine fall data, high-risk behavior data, and data-augmented data to form the model training dataset. Specifically, The previous step can significantly reduce non-fall data, but it is still far more numerous than fall data. Further data augmentation methods can be used to mitigate this problem. Since this invention incorporates personnel information when constructing data features, when constructing new fall data, on the one hand, noise can be added to the collected data, the data can be slightly scaled, and data preceding fall data can be concatenated with other non-fall data. This mixed concatenation method enriches data features, simulates data under various conditions, and increases data diversity. On the other hand, one or more of the personnel information features, such as gender or ski slope, can be modified.

[0049] S2: Constructing a hierarchical learning network model: The hierarchical learning network model consists of a one-dimensional convolutional layer, an attention mechanism layer, a bidirectional long short-term memory network layer, and a fully connected layer. The hierarchical learning network model is trained using a model training dataset, and Focal Loss is used as the loss function. The trained hierarchical learning network model is then used to identify real-time motion data to achieve skier fall detection.

[0050] The specific process includes: S2.1: Preprocess the data in the model training dataset, mainly standardizing the raw data such as acceleration and gyroscope data. The Z-Score standardization method can be used to process the data of each channel:

[0051] in, The mean of the data. The standard deviation is denoted as .

[0052] S2.2: Divide the preprocessed dataset into training and testing sets in a 4:1 ratio. Further divide the training set into training and validation sets in a 9:1 ratio. The input data dimension of the hierarchical learning network model is... Where B is the batch size, L is the time length, and d is the feature dimension.

[0053] S2.3: Obtain the input data by dimensional transformation. The network uses a single 1D convolutional network layer with a kernel size of 1 and an output channel of 1. The output will be Then, perform another dimensional transformation to obtain a dimension of The output characteristics.

[0054] S2.4: Input the above output into the attention mechanism layer to learn different weights at different times. The higher the weight, the greater the contribution of that information to the task. Its output dimension remains the same. .

[0055] S2.5: Input the above output into the Bi-LSTM network layers, for a total of two layers, finally obtaining the dimension as follows: The data.

[0056] S2.6: Pass the above output through a fully connected layer to obtain the final output dimension B, which is the probability value of falling corresponding to the sample data.

[0057] S2.7: Calculate the loss by comparing the final output with the true labels. To further address the performance impact of data class imbalance, Focal Loss is used in the loss calculation, and its calculation method is as follows:

[0058] in, Predict probability values ​​for the model. As a balance factor, is the modulation factor, which is a hyperparameter.

[0059] S2.8: Through multiple iterations of training, the hierarchical learning network model is trained until it reaches the preset accuracy or the maximum number of iterations. During each iteration, the model automatically adjusts the network parameters based on the loss calculation results, continuously optimizing its ability to judge the skier's fall state, thereby gradually improving the model's performance in the skier fall detection task.

[0060] Example 2 This embodiment discloses a skier fall detection system based on hierarchical learning; including: The dataset construction module is configured to: construct the model training dataset by collecting multi-source motion data and personnel information of skiers, processing them to obtain temporal feature data and personnel feature data; segmenting the temporal feature data using windows and labeling it according to fall indicators, dividing it into fall data and non-fall data; filtering the non-fall data using a multi-level threshold method based on speed and posture changes, retaining data judged as high-risk behaviors; merging the fall data, high-risk behavior data, and data augmented to form the model training dataset. The model building module is configured to: construct a hierarchical learning network model, which sequentially includes a one-dimensional convolutional layer, an attention mechanism layer, a bidirectional long short-term memory network layer, and a fully connected layer; train the hierarchical learning network model using the model training dataset and employ Focal Loss as the loss function; and use the trained hierarchical learning network model to identify real-time motion data to achieve skier fall detection. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0061] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a skier fall detection method based on hierarchical learning as described in Embodiment 1 of this disclosure.

[0062] Example 4 The purpose of this embodiment is to provide an electronic device.

[0063] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a skier fall detection method based on hierarchical learning as described in Embodiment 1 of this disclosure.

[0064] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0065] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0066] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A skier fall detection method based on hierarchical learning, characterized in that, include: Constructing a model training dataset: Collect multi-source motion data and personnel information of skiers, and process them separately to obtain time-series feature data and personnel feature data; The temporal feature data is segmented using a window of preset length and labeled according to fall indicators, and divided into fall data and non-fall data. For non-fall data, a multi-level threshold method based on velocity and posture changes is used to filter and retain data that is judged as high-risk behavior. The fall data, high-risk behavior data and data-augmented data are merged to form the model training dataset. Constructing a hierarchical learning network model: The hierarchical learning network model includes, in sequence, a one-dimensional convolutional layer, an attention mechanism layer, a bidirectional long short-term memory network layer, and a fully connected layer; The hierarchical learning network model is trained using the model training dataset, with Focal Loss as the loss function; the trained hierarchical learning network model is then used to identify real-time motion data to achieve skier fall detection.

2. The skier fall detection method based on hierarchical learning as described in claim 1, characterized in that, The personnel information includes skiing experience, selected ski slope level, ski type, and gender; the personnel information is incorporated into the feature data using a two-digit code.

3. The skier fall detection method based on hierarchical learning as described in claim 1, characterized in that, During the ski season, skiers wear an integrated sensing device in the interscapular region for natural skiing, and manually trigger the device to record the current time as a fall marker when a fall occurs during skiing; the integrated sensing device integrates an inertial sensor, a magnetometer, GPS, and a processing unit.

4. The skier fall detection method based on hierarchical learning as described in claim 1, characterized in that, The processing of the multi-source motion data also includes: Attitude information is obtained by using quaternion error Kalman filtering or complementary filtering; The acceleration in the sensor coordinate system is converted to the navigation reference system using the attitude information; The converted acceleration is integrated to obtain a velocity estimate, and the integrated velocity is corrected by fusing the velocity information output by GPS with complementary filtering to eliminate integral drift.

5. The skier fall detection method based on hierarchical learning as described in claim 1, characterized in that, The multi-level threshold method includes: Calculate the resultant velocity in the navigation reference frame at the current moment and compare it with the set velocity threshold; If the combined speed is greater than the set speed threshold, proceed to the next step; otherwise, filter. If the cumulative time is greater than the set speed threshold for T consecutive non-fall data points, proceed to the next step; otherwise, filter. Calculate the attitude change in each direction within the window and compare it with the set attitude change threshold; If the change in attitude in each direction exceeds the set attitude change threshold, it is considered high-risk behavior data and is retained; otherwise, it is filtered out.

6. The skier fall detection method based on hierarchical learning as described in claim 1, characterized in that, When a window of data is identified as a high-risk behavior, the data from multiple adjacent windows before and after it are retained in the model training dataset to maintain data continuity.

7. The skier fall detection method based on hierarchical learning as described in claim 1, characterized in that, The data augmentation method includes: adding noise to the fall data, scaling the amplitude, splicing the data before the fall data with the non-fall data, or modifying one or more of the personnel information features.

8. A skier fall detection system based on hierarchical learning, characterized in that, include: The dataset construction module is configured to: construct the model training dataset by collecting multi-source motion data and personnel information of skiers, processing them to obtain temporal feature data and personnel feature data; segmenting the temporal feature data using windows and labeling it according to fall indicators, dividing it into fall data and non-fall data; filtering the non-fall data using a multi-level threshold method based on speed and posture changes, retaining data judged as high-risk behaviors; merging the fall data, high-risk behavior data, and data augmented to form the model training dataset. The model building module is configured to: build a hierarchical learning network model, wherein the hierarchical learning network model includes, in sequence, a one-dimensional convolutional layer, an attention mechanism layer, a bidirectional long short-term memory network layer, and a fully connected layer; The hierarchical learning network model is trained using the model training dataset, with Focal Loss as the loss function; the trained hierarchical learning network model is then used to identify real-time motion data to achieve skier fall detection.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps of a skier fall detection method based on hierarchical learning as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the skier fall detection method based on hierarchical learning as described in any one of claims 1-7.

Images