Method and system for evaluating falling risk of old people based on human body posture estimation

Abstract

The invention discloses an old people falling risk assessment method and system based on human body posture estimation, relates to the technical field of computers, and solves the technical problem that old people falling risk assessment according to gait data is not accurate enough. And calculating and processing the gait features in the video frame, and obtaining an old people falling risk assessment result according to the gait features. No professional equipment or sensor wearing is needed, and the elderly can complete video shooting and risk assessment autonomously to adapt to family scenes. The elderly and family members have an opportunity to take corresponding prevention measures in advance, and serious consequences caused by falling down are avoided to the greatest extent; and the falling risk of the old people can be accurately classified and evaluated.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and system for assessing the risk of falls of the elderly based on human posture estimation. Background Art

[0002] Falls are a common and serious health problem among the elderly. They refer to sudden, involuntary changes in body position that cause individuals to fall to the ground or a lower surface. [1] Wang Liancheng et al. [2] Research has shown that gait parameters differ between elderly individuals with a history of falls within the past year and those without. Therefore, analyzing these gait parameters during daily walking can effectively differentiate between those at high and low risk, providing personalized fall risk estimates for the elderly. This assessment not only aids elderly individuals in self-management and fall prevention, but also contributes to the development of an age-friendly society and improves their quality of life.

[0003] Existing fall risk assessment schemes are mainly divided into two categories: one is manual assessment, and the other is based on machine learning methods. In general, due to the limitation of professional assessors, manual assessment is more widely used in clinical applications, but it is unrealistic to use this method to achieve real-time supervision in the daily lives of the elderly. Liu Xiaoyan et al. [3] Analyze the factors affecting falls in community and hospitalized elderly people. There are many factors affecting community and hospitalized elderly people with a history of falls within 1 year. After research, gait data has been proven to reflect muscle strength, balance ability and other factors that are highly correlated with the risk of falls. [4] , and gait data are easy to obtain.

[0004] Common gait analysis refers to a technology that evaluates human motor function and health status by observing and measuring the posture and movement patterns of the human body when walking. The use of gait analysis can provide a scientific basis for clinical diagnosis, rehabilitation assessment, and auxiliary design. Gait analysis requires the collection of relevant data when elderly subjects walk. The main methods of collection can be roughly divided into wearable sensor-based methods and image-based methods, such as Figure 1 and Figure 2 shown.

[0005] Wearable sensor-based methods mainly include electromyographic sensors and inertial sensors (IMU). The principle of myoelectric sensors to detect gait is to analyze gait characteristics by measuring the electrical activity of muscles during movement. Mehmood et al. [5]The study used data collected by BASN, which is composed of wearable electromyography sensors, to analyze muscle fatigue and proposed an active fall risk assessment mechanism for the elderly. It was verified that fatigue of the tibialis and gastrocnemius muscles of the lower limbs may cause falls in the elderly. Inertial sensors are generally composed of accelerometers and gyroscopes. Some inertial accelerometers also have magnetometers that can measure magnetic fields. Combining these can obtain information such as acceleration, angular velocity, and magnetic field during human movement, thereby estimating gait information. Howcroft et al. [6] A study found that half of the researchers using inertial sensors were able to develop fall risk models with high accuracy (62%-100%), specificity (35%-100%), and sensitivity (55%-99%). In summary, wearable sensor methods are generally unaffected by ambient lighting and limb occlusion, and offer high accuracy. However, these methods require wearing the sensor during measurement, making them unsuitable for daily monitoring.

[0006] Camera-based methods can be divided into three categories based on the type of camera: RGB cameras, RGB depth cameras (RGB-D cameras), and optical gait analysis. RGB cameras are the most commonly used type of camera in life, and most smartphones have cameras of this type. Therefore, using RGB cameras to record gait is the most portable and economical option. RGB-D cameras can simultaneously capture color images and depth information, and can combine the two to extract more accurate gait features. However, this type of camera is expensive and has a short effective depth measurement range. The farther away from the sensor, the worse the depth image quality. [7] Optical gait analysis requires attaching reflective markers to the key points of the subject's body joints, detecting the reflective markers on the subject's body through multiple cameras at different angles, and restoring the subject's movements in a three-dimensional reconstruction method. This method has high accuracy but is expensive and complex to operate, making it unsuitable for daily testing. In summary, using RGB cameras for gait data collection has the advantages of being daily, easy to obtain, and low-cost, and benefits from OpenPose and other [8] The rise of mature 2D human pose estimation models. Through pre-trained 2D human pose estimation models, users can obtain 2D human skeleton joint data using a single RGB camera, such as Figure 3 shown.

[0007] Therefore, experiments based on RGB cameras can support the application of fall risk analysis for elderly people at home. However, using big data to analyze fall risk requires basic clinical data. This data, combined with the patient's condition and medication usage, can refine the sources of fall risk for elderly subjects and improve their quality of life. However, there are the following problems with data sources:

[0008] 1) The data collection method is not suitable for home use. High accuracy and convenience are difficult to achieve simultaneously, and these data collection methods are overly idealistic. It is difficult to extract gait information from random movements in real life to determine whether there is a risk of falling. Collecting a dataset on falls in the elderly used to train the model is difficult, requiring long-term tracking and monitoring of the daily lives of the elderly to obtain sufficient data. Data privacy issues exist, and the collected data needs to be desensitized.

[0009] 2) Currently, there is relatively little research on the use of RGB camera-based human posture estimation models to assess the fall risk of the elderly. Further research is needed on how to assess the fall risk of the elderly based on gait data.

[0010] [1]HAYAKAWAT,HASHIMOTO S,KANDAH,et al.Risk factors offalls ininpatients and their practical use in identifying high-risk persons atadmission:Fukushima Medical UniversityHospital cohort study[J].BMJ Open,2014,4(8):e005385.

[0011] [2] Wang Liancheng, Zhang Liqin, Zhang Yi, et al. Application of balance and gait analysis test in fall risk assessment of the elderly [J]. Chinese Journal of Rehabilitation Medicine, 2012, 27(03): 251-253.

[0012] [3] Liu Xiaoyan, Ding Xia, Dong Chen, et al. Current status and influencing factors of falls in community and hospitalized elderly people [J]. Chinese Journal of Rehabilitation Theory and Practice, 2022, 28(04): 389-398.

[0013] [4]GREENE BR,MCGRATH D,WALSH L,et al.Quantitative falls riskestimation through multi-sensor assessment of standing balance[J].Physiological measurement,2012,33(12):2049.

[0014] [5] MEHMOOD A, NADEEM A, ASHRAF M, et al. A fall risk assessment mechanism for elderly people through muscle fatigue analysis on data from body area sensor network[J]. IEEE Sensors Journal, 2020, 21(5): 6679-90.

[0015] [6] HOWCROFT J, LEMAIRE E D, KOFMAN J. Prospective elderly fall prediction by older-adult fall-risk modeling with feature selection[J]. Biomedical Signal Processing and Control, 2018, 43: 320-8.

[0016] [7] GEERSE D J, COOLEN B H, ROERDINK M. Kinematic validation of a multi-Kinect v2 instrumented 10-meter walkway for quantitative gait assessments[J]. PloS one, 2015, 10(10): e0139913.

[0017] [8] VISWAKUMAR A, RAJAGOPALAN V, RAY T, et al. Development of a robust, simple, and affordable human gait analysis system using bottom-up pose estimation with a smartphone camera[J]. Frontiers in physiology, 2022, 12: 784865. Summary of the Invention

[0018] This application provides a method and system for assessing the fall risk of the elderly based on human posture estimation. Its technical purpose is to use easily available RGB cameras to capture gait videos of the elderly, and through human posture estimation technology and machine learning algorithms, accurately calculate gait parameters and classify and evaluate fall risks, thereby realizing self-assessment of fall risk in a family context, assisting the elderly in preventing falls in advance, reducing the medical burden on families and society, and providing data support and decision-making basis for personalized gait intervention plans.

[0019] The above technical objectives of this application are achieved through the following technical solutions:

[0020] A fall risk assessment method for the elderly based on human posture estimation, including:

[0021] The walking video of the elderly subjects was acquired using a monocular RGB camera;

[0022] Input the walking video into the YOLOv8-Pose model to detect objects in the video frames and obtain key point information of the human body in each video frame; wherein the key point information includes the pixel coordinates and confidence of the left ankle, right ankle, left knee joint, right knee joint, left hip, right hip, left shoulder and right shoulder;

[0023] Calculate the distance between the human body and the camera in each frame based on the key point information of the human body in each video frame;

[0024] The pixel coordinate difference between the left and right ankles in the vertical direction is used as a timing signal. A dynamic curve of step alternation is constructed based on the timing signal. The local extreme points of the dynamic curve are used as the gait event occurrence points, that is, the start and end positions of each step. The local extreme points include local positive peaks and local negative peaks. The local positive peaks represent HS events of the right limb, and the local negative peaks represent HS events of the left limb. HS events represent heel strike events.

[0025] Calculating gait features according to the distance and the gait events to obtain a gait sequence; wherein the gait features include step length, step time and step speed;

[0026] The abnormal steps at the beginning and end of the gait sequence are screened to obtain the final gait sequence;

[0027] Calculate the mean of the gait features in the final gait sequence;

[0028] The mean and cadence of the sequential gait features are input into the machine learning model, and the fall risk classification result is output, i.e., high fall risk or low fall risk.

[0029] Furthermore, the calculation of the distance between the human body and the camera in each frame according to the key point information of the human body in each video frame includes:

[0030]

[0031] Δd i Indicates the distance between the human body and the camera at the i-th frame; S Ratio Represents the ratio of the pixel height of the human body in the i-th frame to the pixel height in the first frame, that is, d Ref Indicates the initial distance between the human body and the camera.

[0032] Furthermore, the step length includes a single step length and a stride length, the step time includes a single step duration and a stride duration, and the step speed includes a single step speed and a stride speed;

[0033] The single-step length represents the displacement of a foot on one side when a HS event occurs relative to the previous HS event of the other side; the stride length represents the displacement of a foot on one side when a HS event occurs relative to the previous HS event of the foot on that side;

[0034] The single step duration represents the time interval between the occurrence of a HS event on one side of the foot and the last HS event on the other side of the foot; the stride duration represents the time interval between the occurrence of a HS event on one side of the foot and the last HS event on that side of the foot;

[0035] The single-step speed represents the average speed between the occurrence of a HS event on one side of the foot and the last HS event on the other side of the foot; the stride speed represents the average speed between the occurrence of a HS event on one side of the foot and the last HS event on that side of the foot.

[0036] Furthermore,

[0037] The stride duration calculation formula is:

[0038] stride_time(i)=HS(i+2)-HS(i);

[0039] The single-step duration calculation formula is:

[0040] step_time(i)=HS(i+1)-HS(i);

[0041] Where HS(i) represents the time when the i-th HS event occurs, in seconds;

[0042] The stride duration calculation formula is:

[0043] stride_length(i)=Δd f(i+2) -Δd f(i) ;

[0044] The single-step duration calculation formula is:

[0045] step_length(i)=Δed f(i+1) -Δd f(i) ;

[0046] Where, f(i) = HS(i) × FPS, f(i) represents the number of frames corresponding to the i-th HS event, and FPS represents the video frame rate per second; Δd f(i) Indicates the distance between the human body and the camera at the f(i)th frame, in meters;

[0047] The stride speed calculation formula is:

[0048]

[0049] The single-step speed calculation formula is:

[0050]

[0051] The calculation formula of the step frequency is:

[0052]

[0053] Where N represents the total number of steps ultimately used to calculate the parameters, and the unit of step frequency is steps per second.

[0054] Furthermore, the abnormal steps at the beginning and end of the gait sequence are screened to obtain the final gait sequence, including:

[0055] Step 200: Delete the last 20% of the data in the gait sequence, and consider the local extreme point with the largest amplitude in the remaining gait sequence as a stop step. Delete the local extreme point with the largest amplitude and all steps after it, and select n local extreme points from the stop step forward, where n≤8;

[0056] Step 201: Evaluate the abnormality of n local extreme points based on the coefficient of variation; when CV < 16%, the gait detection of the n local extreme points is correct, and the final gait sequence is obtained; when CV ≥ 16%, the gait detection of the n local extreme points is abnormal, and the process goes to step 202; where CV represents the abnormality coefficient;

[0057] Step 202: Calculate the Z score of the single step duration of each step in the n local extreme points, and determine whether the maximum or second maximum absolute value of the Z score appears at the beginning and end positions; if so, delete the corresponding single step and go to step 201 to re-determine the abnormal coefficient CV until the abnormal coefficient CV is less than 16% or the maximum or second maximum absolute value of the Z score does not appear at the beginning and end positions, and obtain the final gait sequence; otherwise, select the current gait sequence as the final gait sequence.

[0058] Furthermore, the abnormal coefficient CV is defined as follows:

[0059]

[0060] Among them, σ represents the standard deviation of the feature sequence, and μ is the mean of the feature sequence;

[0061] The Z score is defined as:

[0062]

[0063] Among them, X(i) represents the kth number in the feature sequence.

[0064] Furthermore, the machine learning model performance evaluation indicators include balanced accuracy, sensitivity, specificity and F1 score.

[0065] Furthermore, the machine learning model is a support vector machine (SVM).

[0066] A fall risk assessment system for the elderly based on human posture estimation, which is used to implement a fall risk assessment method for the elderly, including:

[0067] The video acquisition module uses a monocular RGB camera to capture walking videos of elderly subjects;

[0068] An information extraction module inputs the walking video into the YOLOv8-Pose model to detect objects in the video frames and obtain key point information of the human body in each video frame. The key point information includes the pixel coordinates and confidence scores of the left ankle, right ankle, left knee joint, right knee joint, left hip, right hip, left shoulder, and right shoulder.

[0069] The first calculation module calculates the distance between the human body and the camera in each frame based on the key point information of the human body in each video frame;

[0070] A gait event determination module uses the vertical pixel coordinate difference between the left and right ankles as a timing signal, constructs a dynamic curve of step alternation based on the timing signal, and uses the local extreme points of the dynamic curve as the gait event occurrence points, i.e., the start and end positions of each step. The local extreme points include local positive peaks and local negative peaks. The local positive peaks represent HS events of the right limb, and the local negative peaks represent HS events of the left limb. HS events indicate heel strike events.

[0071] A second calculation module calculates gait features according to the distance and the gait event to obtain a gait sequence; wherein the gait features include step length, step time and step speed;

[0072] The screening module filters out abnormal steps at the beginning and end of the gait sequence to obtain the final gait sequence;

[0073] The third calculation module calculates the mean of the gait features in the final gait sequence;

[0074] The evaluation module inputs the mean and cadence of the sequence-based gait features into the machine learning model and outputs the fall risk classification result, i.e., high fall risk or low fall risk.

[0075] The beneficial effects of this application are:

[0076] (1) Simple operation and low cost. Only a monocular RGB camera (such as a smartphone) is required. No professional equipment or wearable sensors are required. The elderly can complete video shooting and risk assessment independently, which is suitable for home scenarios.

[0077] (2) Different from the common post-fall detection, the fall risk assessment scheme proposed in this application estimates the fall risk of the elderly and can predict the fall risk before the injury occurs, so that the elderly and their families have the opportunity to take corresponding preventive measures in advance, such as improving the home environment, adjusting daily exercise plans, etc., to minimize the serious consequences of falls.

[0078] (3) A large number of experiments have shown the universal applicability of this application to chronic disease groups. At the same time, the solution described in this application is not limited to specific scenarios and is applicable to a variety of environments such as homes and communities, and can meet the needs of fall risk assessment for the elderly in different scenarios.

[0079] (4) This application selected seven gait parameters. The independent sample t-test showed that these gait parameters had significant differences between the high-fall risk group and the low-fall risk group (p<0.05), which can effectively reflect the subjects' fall risk. On this basis, the balance accuracy of the support vector machine for fall assessment in the elderly reached 86.35%, the sensitivity reached 87.55%, and the specificity reached 85.16%, which can more accurately classify and assess the fall risk of the elderly. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 Schematic diagram of gait data collection based on wearable sensors;

[0081] Figure 2 Schematic diagram of camera-based gait data collection;

[0082] Figure 3 Schematic diagram of obtaining human skeleton key point data using a single RGB camera;

[0083] Figure 4 This is a flow chart of a method for assessing the risk of falls for elderly people based on human posture estimation in an embodiment of the present application;

[0084] Figure 5 This is a schematic diagram of the depth calculation principle of a monocular RGB camera;

[0085] In the picture: DETAILED DESCRIPTION

[0086] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0087] like Figure 4 As shown, the method for assessing the risk of falls of the elderly based on human posture estimation described in this application includes:

[0088] 100: Acquire walking videos of elderly subjects using a monocular RGB camera.

[0089] 101: Input the walking video into the YOLOv8-Pose model to detect targets in the video frames and obtain key point information of the human body in each video frame; wherein the key point information includes pixel coordinates (unit: px) and confidence levels of the left ankle, right ankle, left knee joint, right knee joint, left hip, right hip, left shoulder, and right shoulder.

[0090] Preferably, in order to eliminate the influence of jitter on key point detection, all video frame data are processed by median filtering, and the window size is set to 8 frames.

[0091] 102: According to the change of the pixel height of the elderly subjects in the video frame, based on the camera imaging principle and the premise of height and walking route facing the camera, the distance between the human body and the camera in each frame is calculated according to the key point information of the human body in each video frame, such as Figure 5 shown.

[0092] Calculate the walking distance Δd of the human body at the i-th frame i It is easy to know that the focal length f can be canceled, so the calculation result is only related to the initial distance between the human body and the camera and the change ratio of the pixel height of the human body in the picture.

[0093] Specifically, the distance is calculated as:

[0094]

[0095] Δd i Indicates the distance between the human body and the camera at the i-th frame; S Ratio Represents the ratio of the pixel height of the human body in the i-th frame to the pixel height in the first frame, that is, d Ref Indicates the initial distance between the human body and the camera.

[0096] 103: The pixel coordinate difference between the left ankle and the right ankle in the vertical direction is used as a timing signal, and a dynamic curve of step alternation is constructed based on the timing signal. The local extreme points of the dynamic curve are used as the gait event occurrence points, that is, the starting and ending positions of each step; the local extreme points include local positive peaks and local negative peaks, the local positive peaks represent HS events of the right limbs, and the local negative peaks represent HS events of the left limbs, and the HS events represent heel-strike events.

[0097] Specifically, as the elderly subjects approached the camera, their left and right feet alternated between supporting and lifting. Therefore, the local maximum and minimum values ​​of the vertical distance between the left and right ankle joints were used to identify left and right gait events. By calculating the timing signal of the difference between the y-coordinates of the left and right ankles, heel strike (HS) events on the right limb were identified at the local positive peak, and heel strike events on the left limb were identified at the local negative peak. Using the vertical pixel coordinate difference between the left and right ankles as the timing signal, a dynamic curve reflecting the alternating steps was constructed. After median filtering and noise reduction, the dynamic curve identified local extreme points as the time points of gait events (i.e., the start and end positions of each step).

[0098] 104: Calculate gait features according to the distance and the gait event to obtain a gait sequence; wherein the gait features include step length, step time, and step speed.

[0099] Preferably, the step length represents the distance the foot moves between two adjacent HS events, the step time represents the time interval between two adjacent HS events, and the step speed represents the average speed of the foot moving between two adjacent HS events.

[0100] Each type of gait feature can be further subdivided into single-step features and stride features based on whether bilateral foot motion is considered when defining adjacent HS events. Therefore, the stride length includes single-step length and stride length, the step duration includes single-step duration and stride duration, and the stride speed includes single-step speed and stride speed.

[0101] The single-step length represents the displacement of a foot on one side when an HS event occurs relative to the last HS event of the other side; the stride length represents the displacement of a foot on one side when an HS event occurs relative to the last HS event of the foot on that side.

[0102] The single step duration represents the time interval between the occurrence of a HS event on one side of the foot and the last HS event on the other side of the foot; the stride duration represents the time interval between the occurrence of a HS event on one side of the foot and the last HS event on that side of the foot.

[0103] The single-step speed represents the average speed between the occurrence of a HS event on one side of the foot and the last HS event on the other side of the foot; the stride speed represents the average speed between the occurrence of a HS event on one side of the foot and the last HS event on that side of the foot.

[0104] The stride duration calculation formula is:

[0105] stride_time(i)=HS(i+2)-HS(i);

[0106] The single-step duration calculation formula is:

[0107] step_time(i)=HS(i+1)-HS(i);

[0108] Where HS(i) represents the time when the i-th HS event occurs, in seconds;

[0109] The stride duration calculation formula is:

[0110] stride_length(i)=Δdd f(i+2) -Δd f(i) ;

[0111] The single-step duration calculation formula is:

[0112] step_length(i)=Δd f(i+1) -Δd f(i) ;

[0113] Where, f(i) = HS(i) × FPS, f(i) represents the number of frames corresponding to the i-th HS event, and FPS represents the video frame rate per second; Δd f(i) Indicates the distance between the human body and the camera at the f(i)th frame, in meters;

[0114] The stride speed calculation formula is:

[0115]

[0116] The single-step speed calculation formula is:

[0117]

[0118] The calculation formula of the step frequency is:

[0119]

[0120] Where N represents the total number of steps ultimately used to calculate the parameters, and the unit of step frequency is steps per second.

[0121] 105: Filter the abnormal steps at the beginning and end of the gait sequence to obtain a final gait sequence, including:

[0122] Step 200: Delete the last 20% of the data in the gait sequence, and consider the local extreme point with the largest amplitude in the remaining gait sequence as a stop step. Delete the local extreme point with the largest amplitude and all subsequent steps, and select n local extreme points from the stop step forward, where n≤8.

[0123] Specifically, the gait of a person at the beginning or end of walking exhibits significant differences compared to their gait during walking. Furthermore, the visual anchoring effect of the finish line may cause elderly people to actively adjust their gait as they approach the finish line. Some elderly people may even unconsciously move after stopping at the finish line. These situations can lead to outliers in the sequence, which in turn affects the calculation of gait parameters. Therefore, outlier detection and processing of gait parameters in the sequence are necessary. First, to eliminate outliers caused by unexpected behavior of elderly people at the finish line, the last 20% of the video frames are excluded. The maximum value is then extracted from the remaining data. The extreme point with the largest amplitude in the detected signal is considered a stop step, and all steps after this point are excluded. A maximum of eight extreme points are selected from the stop step forward, and data abnormalities are then evaluated based on the coefficient of variation (CV).

[0124] Step 201: Evaluate the abnormality of n local extreme points based on the coefficient of variation; when CV < 16%, the gait detection of the n local extreme points is correct, and the final gait sequence is obtained; when CV ≥ 16%, the gait detection of the n local extreme points is abnormal, and go to step 202; where CV represents the abnormality coefficient.

[0125] Preferably, the abnormal coefficient CV is defined as follows:

[0126]

[0127] Among them, σ represents the standard deviation of the feature sequence, and μ is the mean of the feature sequence.

[0128] Step 202: Calculate the Z score of the single step duration of each step in the n local extreme points, and determine whether the maximum or second maximum absolute value of the Z score appears at the beginning and end positions; if so, delete the corresponding single step and go to step 201 to re-determine the abnormal coefficient CV until the abnormal coefficient CV is less than 16% or the maximum or second maximum absolute value of the Z score does not appear at the beginning and end positions, and obtain the final gait sequence; otherwise, select the current gait sequence as the final gait sequence.

[0129] Preferably, the Z score is defined as:

[0130]

[0131] Among them, X(i) represents the kth number in the feature sequence.

[0132] 106: Calculate the mean of the gait features in the final gait sequence.

[0133] 107: Input the mean and cadence of the sequence gait features into the machine learning model and output the fall risk classification result, i.e., high fall risk or low fall risk.

[0134] Preferably, the machine learning model performance evaluation indicators include balanced accuracy, sensitivity, specificity and F1 score, wherein the performance indicator used when performing hyperparameter tuning on the machine learning model is balanced accuracy.

[0135] Preferably, the optimal machine learning model selected according to the indicator of balanced accuracy is a support vector machine (SVM).

[0136] As shown in Table 1, in the comparison of various models, the support vector machine (SVM) has a balanced accuracy of 86.35%, which means that it has good recognition ability for both low-risk and high-risk individuals, and has good classification ability and robustness. Therefore, the support vector machine is preferred as the algorithm model, and its hyperparameter settings are shown in Table 2.

[0137] Table 1 Performance comparison of different models

[0138]

[0139] Table 2 Support vector machine hyperparameters

[0140] Hyperparameter name C Gamma Kernel Function Category weight (low risk: high risk) value 20.0 5.45e-5 Radial Basis Function 1:3.8

[0141] To enhance the model's credibility and interpretability in healthcare and elderly care settings, we further introduced the SHAP (SHapley Additive exPlanations) method to explain the model's classification logic. Analysis of the SHAP values ​​for each feature revealed that stride length and speed had the most significant impact on the model's judgment, while features such as cadence and stride duration provided auxiliary support. This interpretation is consistent with clinical experience, validating the model's validity in professional scenarios.

[0142] Overall, the model system is lightweight and easy to deploy, suitable for mobile devices or edge computing platforms. Its methodology incorporates highly interpretable and scalable feature design and classification mechanisms, significantly outperforming traditional questionnaire assessments or static testing methods. It enables continuous, non-invasive risk monitoring of the elderly population, and has promising prospects for widespread adoption and social value.

[0143] The elderly fall risk assessment system based on human posture estimation described in this application includes a video acquisition module, an information extraction module, a first calculation module, a gait event determination module, a second calculation module, a screening module, a third calculation module and an evaluation module.

[0144] The video acquisition module is used to acquire walking videos of elderly subjects using a monocular RGB camera.

[0145] The information extraction module is used to input walking videos into the YOLOv8-Pose model to detect targets in the video frames and obtain key point information of the human body in each video frame; wherein, the key point information includes the pixel coordinates and confidence levels of the left ankle, right ankle, left knee joint, right knee joint, left hip, right hip, left shoulder, and right shoulder.

[0146] The first calculation module is used to calculate the distance between the human body and the camera in each frame according to the key point information of the human body in each video frame.

[0147] The gait event determination module is used to use the pixel coordinate difference between the left ankle and the right ankle in the vertical direction as a timing signal, construct a dynamic curve of step alternation based on the timing signal, and use the local extreme points of the dynamic curve as the gait event occurrence points, that is, the starting and ending positions of each step; the local extreme points include local positive peaks and local negative peaks, the local positive peaks represent HS events of the right limbs, and the local negative peaks represent HS events of the left limbs, and HS events represent heel-strike events.

[0148] The second calculation module is used to calculate the gait characteristics according to the distance and the gait event to obtain a gait sequence; wherein the gait characteristics include step length, step time and step speed.

[0149] The screening module is used to screen the abnormal steps at the beginning and end of the gait sequence to obtain the final gait sequence.

[0150] The third calculation module is used to calculate the mean of the gait features in the final gait sequence.

[0151] The evaluation module is used to input the mean and cadence of the sequence-form gait features into the machine learning model to obtain the performance classification results of the machine learning model, i.e., high fall risk or low fall risk.

[0152] The above are exemplary embodiments of the present application, and the scope of protection of the present application is determined by the claims and their equivalents.

Claims

1. A method for assessing the risk of falls in the elderly based on human posture estimation, characterized in that: include: The walking video of the elderly subjects was acquired using a monocular RGB camera; Input the walking video into the YOLOv8-Pose model to detect objects in the video frames and obtain key point information of the human body in each video frame; wherein the key point information includes the pixel coordinates and confidence of the left ankle, right ankle, left knee joint, right knee joint, left hip, right hip, left shoulder and right shoulder; Calculate the distance between the human body and the camera in each frame based on the key point information of the human body in each video frame; The pixel coordinate difference between the left and right ankles in the vertical direction is used as a timing signal. A dynamic curve of step alternation is constructed based on the timing signal. The local extreme points of the dynamic curve are used as the gait event occurrence points, that is, the start and end positions of each step. The local extreme points include local positive peaks and local negative peaks. The local positive peaks represent HS events of the right limb, and the local negative peaks represent HS events of the left limb. HS events represent heel strike events. Calculating gait features according to the distance and the gait events to obtain a gait sequence; wherein the gait features include step length, step time and step speed; The abnormal steps at the beginning and end of the gait sequence are screened to obtain the final gait sequence; Calculate the mean of the gait features in the final gait sequence; The mean and cadence of the sequential gait features are input into the machine learning model, and the fall risk classification result is output, i.e., high fall risk or low fall risk.

2. The method for assessing the risk of falls in the elderly according to claim 1, wherein: The calculation of the distance between the human body and the camera in each frame according to the key point information of the human body in each video frame includes: Δd i Indicates the distance between the human body and the camera at the i-th frame; S Ratio Represents the ratio of the pixel height of the human body in the i-th frame to the pixel height in the first frame, that is, d Ref Indicates the initial distance between the human body and the camera.

3. The method for assessing the risk of falls in the elderly according to claim 2, wherein: The step length includes single step length and stride length, the step time includes single step duration and stride duration, and the step speed includes single step speed and stride speed; The single-step length represents the displacement of a foot on one side when a HS event occurs relative to the previous HS event of the other side; the stride length represents the displacement of a foot on one side when a HS event occurs relative to the previous HS event of the foot on that side; The single step duration represents the time interval between the occurrence of a HS event on one side of the foot and the last HS event on the other side of the foot; the stride duration represents the time interval between the occurrence of a HS event on one side of the foot and the last HS event on that side of the foot; The single-step speed represents the average speed between the occurrence of a HS event on one side of the foot and the last HS event on the other side of the foot; the stride speed represents the average speed between the occurrence of a HS event on one side of the foot and the last HS event on that side of the foot.

4. The method for assessing the risk of falls in the elderly according to claim 3, wherein: The stride duration calculation formula is: stride_time(i)=HS(i+2)-HS(i); The single-step duration calculation formula is: step_time(i)=HS(i+1)-HS(i); Where HS(i) represents the time when the i-th HS event occurs, in seconds; The stride duration calculation formula is: stride_length(i)=Δd f(i+2) -Δd f(i) ; The single-step duration calculation formula is: step_length(i)=Δd f(i+1) -Δd f(i) ; Where, f(i) = HS(i) × FPS, f(i) represents the number of frames corresponding to the i-th HS event, and FPS represents the video frame rate per second; Δd f(i) Indicates the distance between the human body and the camera at the f(i)th frame, in meters; The stride speed calculation formula is: The single-step speed calculation formula is: The calculation formula of the step frequency is: Where N represents the total number of steps ultimately used to calculate the parameters, and the unit of step frequency is steps per minute.

5. The method for assessing the risk of falls in the elderly according to claim 4, wherein: The abnormal steps at the beginning and end of the gait sequence are screened to obtain the final gait sequence, including: Step 200: Delete the last 20% of the data in the gait sequence, and consider the local extreme point with the largest amplitude in the remaining gait sequence as a stop step. Delete the local extreme point with the largest amplitude and all steps after it, and select n local extreme points from the stop step forward, where n≤8; Step 201: Evaluate the abnormality of n local extreme points based on the coefficient of variation; when CV < 16%, the gait detection of the n local extreme points is correct, and the final gait sequence is obtained; when CV ≥ 16%, the gait detection of the n local extreme points is abnormal, and the process goes to step 202; where CV represents the abnormality coefficient; Step 202: Calculate the Z score of the single step duration of each step in the n local extreme points, and determine whether the maximum or second maximum absolute value of the Z score appears at the beginning and end positions; if so, delete the corresponding single step and go to step 201 to re-determine the abnormal coefficient CV until the abnormal coefficient CV is less than 16% or the maximum or second maximum absolute value of the Z score does not appear at the beginning and end positions, and obtain the final gait sequence; otherwise, select the current gait sequence as the final gait sequence.

6. The method for assessing the risk of falls in the elderly according to claim 1, wherein: The abnormal coefficient CV is defined as follows: Among them, σ represents the standard deviation of the feature sequence, and μ is the mean of the feature sequence; The Z score is defined as: Among them, X(i) represents the kth number in the feature sequence.

7. The method for assessing the risk of falls in the elderly according to claim 1, wherein: The machine learning model is a support vector machine (SVM).

8. The method for assessing the risk of falls in the elderly according to claim 1, wherein: The machine learning model performance evaluation indicators include balanced accuracy, sensitivity, specificity and F1 score.

9. A system for assessing the risk of falls of the elderly based on human posture estimation, the system being used to implement the method for assessing the risk of falls of the elderly according to any one of claims 1 to 8, characterized in that: include: The video acquisition module uses a monocular RGB camera to capture walking videos of elderly subjects; An information extraction module inputs the walking video into the YOLOv8-Pose model to detect objects in the video frames and obtain key point information of the human body in each video frame. The key point information includes the pixel coordinates and confidence scores of the left ankle, right ankle, left knee joint, right knee joint, left hip, right hip, left shoulder, and right shoulder. The first calculation module calculates the distance between the human body and the camera in each frame based on the key point information of the human body in each video frame; A gait event determination module uses the vertical pixel coordinate difference between the left and right ankles as a timing signal, constructs a dynamic curve of step alternation based on the timing signal, and uses the local extreme points of the dynamic curve as the gait event occurrence points, i.e., the start and end positions of each step. The local extreme points include local positive peaks and local negative peaks. The local positive peaks represent HS events of the right limb, and the local negative peaks represent HS events of the left limb. HS events indicate heel strike events. A second calculation module calculates gait features according to the distance and the gait event to obtain a gait sequence; wherein the gait features include step length, step time and step speed; The screening module filters out abnormal steps at the beginning and end of the gait sequence to obtain the final gait sequence; The third calculation module calculates the mean of the gait features in the final gait sequence; The evaluation module inputs the mean and cadence of the sequence-based gait features into the machine learning model and outputs the fall risk classification result, i.e., high fall risk or low fall risk.

Images