CN-121996923-A - Skier falling detection method and system based on hierarchical learning

CN121996923ACN 121996923 ACN121996923 ACN 121996923ACN-121996923-A

Abstract

The invention belongs to the field of behavior detection and provides a skier falling detection method and system based on layered learning, wherein the method comprises the steps of collecting multisource motion data and personnel information of a skier, and respectively processing the multisource motion data and the personnel information to obtain time sequence characteristic data and personnel characteristic data; the method comprises the steps of dividing time sequence characteristic data by using a window with a preset length and marking according to falling marks, dividing the time sequence characteristic data into falling data and non-falling data, screening the non-falling data by adopting a multi-level threshold method based on speed and attitude variation, reserving data judged to be high-risk behaviors, combining the falling data, the high-risk behavior data and data generated by data enhancement to form a model training data set, training a layering learning network model by using the model training data set, and identifying real-time movement data by using the trained layering learning network model to realize falling detection of skiers. Solves the problems of poor generalization performance, no individuation characteristic and the like in the prior art.

Inventors

MA LELE
Sun Canhang
WANG WENRUI
PENG CHENG
LIU HAO
CHEN WEIKANG
WANG ZHICHUANG

Assignees

观云(山东)智能科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251222

Claims (10)

1. A skier fall detection method based on hierarchical learning, comprising: The method comprises the steps of establishing a model training data set, namely collecting multisource movement data and personnel information of skiers, respectively processing the multisource movement data and the personnel information to obtain time sequence characteristic data and personnel characteristic data, dividing the time sequence characteristic data by using a window with preset length, marking the time sequence characteristic data according to a falling mark, and dividing the time sequence characteristic data into falling data and non-falling data; The method comprises the steps of constructing a layered learning network model, training the layered learning network model by using a model training data set, adopting Focal Loss as a Loss function, and identifying real-time motion data by using the trained layered learning network model to realize the falling detection of skiers.
2. The method for detecting the falling of a skier based on layered learning according to claim 1, wherein the personnel information comprises skiing experience, selected skiing grade, skiing type and sex, and wherein the personnel information is integrated into the characteristic data by adopting a two-digit coding mode.
3. A method of fall detection for skiers based on hierarchical learning as claimed in claim 1, wherein during the skiing season, the skiers wear integrated sensing means integrating inertial sensors, magnetometers, GPS and processing unit to perform natural skiing in the inter-scapular region and manual triggering means to record the current time as a fall identifier when a fall occurs during skiing.
4. A method of detecting a fall of a skier based on hierarchical learning as claimed in claim 1, wherein the processing of the multi-source movement data further comprises: Carrying out gesture calculation by adopting quaternion error Kalman filtering or complementary filtering to obtain gesture information; converting acceleration under a sensor coordinate system into a navigation reference system by utilizing the attitude information; And integrating the converted acceleration to obtain speed estimation, and adopting complementary filtering to fuse the speed information output by the GPS to correct the integrated speed so as to eliminate integrated drift.
5. A skier fall detection method based on hierarchical learning as claimed in claim 1, wherein the multi-level thresholding method comprises: Calculating the combined speed under the navigation reference system at the current moment, and comparing the combined speed with a set speed threshold; if the combined speed is greater than the set speed threshold, entering the next step, otherwise, filtering; Accumulating duration, if the combined speed of the continuous T non-falling data is greater than a set speed threshold, entering the next step, otherwise, filtering; calculating the attitude change quantity of each direction in the window, and comparing the attitude change quantity with a set attitude change threshold; If the posture change quantity of each direction is larger than the set posture change threshold value, the high-risk behavior data are judged to be reserved, otherwise, the high-risk behavior data are filtered.
6. A skier fall detection method based on hierarchical learning as claimed in claim 1, wherein when certain window data is determined to be high risk, a plurality of window data adjacent to each other in front of and behind the window data are retained together in the model training data set to maintain data continuity.
7. A method of detecting a fall of a skier based on hierarchical learning as claimed in claim 1, wherein the method of data augmentation comprises one or more of adding noise to the fall data, amplitude scaling, stitching data before the fall data with non-fall data, or modifying the personal information characteristics.
8. A skier fall detection system based on hierarchical learning, comprising: The data set construction module is configured to construct a model training data set, namely collecting multisource movement data and personnel information of skiers, respectively processing the multisource movement data and the personnel information to obtain time sequence characteristic data and personnel characteristic data, dividing the time sequence characteristic data by using windows and marking according to falling identifications, dividing the time sequence characteristic data into falling data and non-falling data, screening the non-falling data by adopting a multi-level threshold method based on speed and attitude variation, retaining the data judged to be high-risk behaviors, combining the falling data, the high-risk behavior data and the data generated by data enhancement, and forming the model training data set; The model construction module is configured to construct a layered learning network model, wherein the layered learning network model sequentially comprises a one-dimensional convolution layer, an attention mechanism layer, a bidirectional long-short-time memory network layer and a full-connection layer, train the layered learning network model by using the model training data set, adopt Focal Loss as a Loss function, and identify real-time motion data by using the trained layered learning network model to realize the falling detection of skiers.
9. A computer readable storage medium having stored thereon a program which when executed by a processor performs the steps of a method of detecting a fall of a skier based on hierarchical learning as claimed in any one of claims 1 to 7.
10. An electronic device comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of a method of detecting a fall of a skier based on hierarchical learning as claimed in any one of claims 1 to 7.

Description

Skier falling detection method and system based on hierarchical learning Technical Field The invention belongs to the field of behavior detection, and particularly relates to a skier falling detection method and system based on layered learning. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. As a new fashion for body building and recreation and entertainment for the whole people, ice and snow culture is gradually integrated into the life of the masses. Meanwhile, the skiing safety also faces significant challenges, namely, on one hand, the construction of a skiing site is required to be continuously perfected, the safety consciousness of skiers is enhanced, and on the other hand, the detection of safety accidents is very important, the safety accidents are detected, and the problem of locating the accidents is solved, so that the accurate optimization is realized. And fall events are one of the most common types of ski security incidents. Deep learning methods based on visual information have proven to have significant effects in a number of tasks. However, the method has obvious limitations in practical application, on one hand, because the range of the skiing field is wide, a large number of camera shooting devices are required to be deployed for realizing comprehensive monitoring, and on the other hand, the skiing exercise is often in an outdoor changeable environment, and weather conditions (snowfall, haze, light change and the like) are easy to interfere with visual data quality, so that the recognition performance is further influenced. In contrast, microelectromechanical systems (Micro-Electro-MECHANICAL SYSTEM, MEMS) are receiving considerable attention from researchers due to their small size, light weight, low cost, and ease of integration. Although MEMS have demonstrated potential in motion monitoring, little research is currently done on fall behavior detection in skiing sports. The existing method is mostly based on a threshold method, performs feature selection by experience, designs corresponding thresholds, has poor generalization performance, is difficult to be suitable for wide and complex scenes, does not have personalized characteristics, and has difficult detection precision and robustness to meet actual application requirements. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a skier falling detection method and system based on layered learning, which integrate an inertial sensor, a magnetometer, a GPS and a processing unit on wearable equipment, such as a skiwear, and are responsible for functions of acquisition, data storage and the like, analyze and process acquired data on terminal equipment to realize falling detection and other downstream tasks, and simultaneously provide a layered learning model to reduce performance influence problems caused by category imbalance problems. To achieve the above object, one or more embodiments of the present invention provide the following technical solutions: the first aspect of the invention provides a skier fall detection method based on hierarchical learning, comprising: The method comprises the steps of establishing a model training data set, namely collecting multisource movement data and personnel information of skiers, respectively processing the multisource movement data and the personnel information to obtain time sequence characteristic data and personnel characteristic data, dividing the time sequence characteristic data by using a window with preset length, marking the time sequence characteristic data according to a falling mark, and dividing the time sequence characteristic data into falling data and non-falling data; The method comprises the steps of constructing a layered learning network model, training the layered learning network model by using a model training data set, adopting Focal Loss as a Loss function, and identifying real-time motion data by using the trained layered learning network model to realize the falling detection of skiers. As a further technical feature, the personal information includes skiing experience, selected skiing grade, skiing type and sex, and the personal information is integrated into the feature data by adopting a two-digit coding mode. As a further technical feature, during the skiing season, natural skiing is performed by a skier wearing an integrated sensing device in the inter-scapular region, which integrates an inertial sensor, a magnetometer, a GPS and a processing unit, and a manual triggering device records the current time as a fall identification when a fall occurs during skiing. As a further technical feature, the processing of the multi-source motion data further includes: Carrying out gesture calculation by adopting quaternion error Kalman filtering or complementary filtering to obtain gesture information; converting acce