Elderly patient behavior image monitoring and falling risk analysis system and method thereof

Abstract

The invention relates to the technical field of image monitoring, in particular to an elderly patient behavior image monitoring and falling risk analysis system and method, and the method comprises the steps: a front-end collection module collects a time sequence image of an activity area of an elderly patient; the posture detection module analyzes and extracts three-dimensional posture information of the elderly patient based on a differential geometry theory, and comprises a key point detection unit, a manifold mapping unit and a constraint optimization unit; the gait analysis module processes attitude parameters through a parameter extraction unit, a time sequence track unit and a curvature analysis unit; the risk assessment module calculates a fall risk level based on the gait feature data; according to the invention, an APMK-HCW algorithm is innovatively applied to realize high-precision 3D attitude estimation under a monocular camera, geodesic curvature analysis is introduced to quantify gait stability, and a personalized baseline assessment and adaptive optimization mechanism is adopted to realize early precise early warning of the falling risk of the elderly patient.

Description

Technical Field

[0001] This invention relates to the field of image monitoring technology, specifically to a system and method for monitoring the behavior of elderly patients and analyzing fall risk. Background Technology

[0002] With the increasing global trend of population aging, falls among the elderly have become a serious public health problem. Statistics show that the annual fall rate for people aged 65 and over is approximately 30%, while the rate for those aged 80 and over is as high as 50%. Falls not only lead to serious physical injuries such as fractures and traumatic brain injuries, but also cause fear of falling again, limiting their daily activities and reducing their quality of life.

[0003] Currently, fall risk assessment for elderly patients mainly relies on the subjective judgment of healthcare professionals or traditional assessment scales, such as the Berg Balance Scale and the Tinetti Balance and Gait Assessment. While these methods are simple and easy to implement, they lack objectivity and real-time performance, making it difficult to capture subtle abnormal changes in the behavior of elderly patients. Furthermore, existing automated monitoring systems mostly depend on wearable devices or multi-sensor fusion technology, which not only increases the burden on elderly patients but also presents problems such as inconvenience in wearing them and ease of forgetting them.

[0004] Fall detection systems based on computer vision have received widespread attention in recent years, but existing technologies still face several challenges: First, the accuracy of 3D human pose estimation under monocular cameras is limited; second, it is difficult to effectively distinguish between normal and abnormal behaviors; third, there is a lack of accurate quantification methods for gait instability; and fourth, the early warning mechanism is not accurate enough, which is prone to false alarms or missed alarms.

[0005] Therefore, there is an urgent need to develop a non-contact, high-precision, and intelligent system for monitoring the behavior of elderly patients and analyzing fall risks, so as to achieve early identification and warning of fall risks in elderly patients. Summary of the Invention

[0006] The purpose of this invention is to provide a system and method for monitoring the behavior of elderly patients and analyzing fall risks. By integrating differential geometry theory and computer vision technology, it can achieve accurate monitoring of the behavior of elderly patients and early warning of fall risks.

[0007] This invention proposes a system for monitoring behavioral images and analyzing fall risk in elderly patients, comprising:

[0008] The front-end acquisition module is used to acquire time-series images of the activity areas of elderly patients and generate image data streams;

[0009] The pose detection module is communicatively connected to the front-end acquisition module. It is used to receive the image data stream, analyze and extract the three-dimensional pose information of the elderly patient based on differential geometry theory, construct the pose manifold representation, and generate pose parameters.

[0010] The gait analysis module is communicatively connected to the posture detection module. It is used to receive the posture parameters, extract stride length, gait symmetry and swing time parameters, calculate geodesic curvature, analyze posture trajectory stability, and generate gait feature data.

[0011] A risk assessment module, communicatively connected to the gait analysis module, is used to receive the gait feature data, calculate the fall risk level based on a preset risk model, and generate a risk assessment result; and

[0012] The early warning management module is communicatively connected to the risk assessment module. It is used to receive the risk assessment results, trigger the corresponding level of early warning notification according to the risk level, and send the early warning information to the designated receiving terminal.

[0013] Preferably, the attitude detection module includes:

[0014] A key point detection unit is used to detect the two-dimensional coordinates of key points of the human body from the image data stream;

[0015] A manifold mapping unit, connected to the keypoint detection unit, is used to construct an attitude manifold space, map the two-dimensional coordinates to the attitude manifold space, and obtain an initial three-dimensional attitude estimate through local linear reconstruction; and

[0016] The constraint optimization unit, connected to the manifold mapping unit, is used to apply biomechanical constraints based on Riemannian geometry to optimize the initial three-dimensional attitude estimate and generate accurate attitude parameters that conform to the human physiological structure.

[0017] Preferably, the manifold mapping unit is specifically used for:

[0018] A 17-dimensional attitude manifold space is constructed, which corresponds to the degrees of freedom of 17 key joints of the human body;

[0019] Calculate the K nearest neighbor poses based on the feature space distance, where K is 8;

[0020] The target's 3D pose is reconstructed by weighted combination of the K nearest neighbor poses; and

[0021] Gradient descent is applied to optimize the reconstructed weights and minimize the projection error.

[0022] Preferably, the constraint optimization unit is specifically used for:

[0023] Establish a joint constraint system, including bone length constraints, joint angle constraints, and left-right symmetry constraints;

[0024] Establish a tangent space at the current attitude point, and represent the physiological constraints as the allowable variation region in the tangent space;

[0025] The constrained tangent space changes are mapped back to new attitude points on the manifold using an exponential mapping; and

[0026] Iteratively apply constraints until all physiological limitations are met.

[0027] Preferably, the gait analysis module includes:

[0028] The parameter extraction unit is used to extract basic gait parameters such as stride length, gait symmetry, and swing time from the posture parameters.

[0029] The temporal trajectory unit is used to connect attitude points in consecutive frames to form curves on the attitude manifold, and to perform attitude trajectory parameterization; and

[0030] The curvature analysis unit is used to calculate the deviation between the attitude trajectory and the geodesic, measure the curvature of the geodesic of the attitude trajectory, analyze the curvature change characteristics, and generate stability indices.

[0031] Preferably, the curvature analysis unit uses a 5-frame sliding window to calculate curvature and sets differentiated curvature thresholds according to the patient's age group, with curvature thresholds of 0.08, 0.06 and 0.04 for the 65-75 age group, the 75-85 age group and the over 85 age group, respectively.

[0032] Preferably, the risk assessment module includes:

[0033] The index normalization unit is used to normalize the gait feature data to make it suitable for the risk model.

[0034] Individual baseline units are used to store and update patients' historical gait data to establish personalized risk assessment baselines; and

[0035] The comprehensive scoring unit is used to calculate a comprehensive risk score based on normalized indicators and individual baselines, and to classify the risk level into three levels: low risk, medium risk, and high risk according to preset thresholds.

[0036] Preferably, the early warning management module includes:

[0037] The graded notification unit is used to select different notification methods according to the risk level. Low risk is recorded in the system log, medium risk is notified to the nursing station, and high risk is immediately notified to nearby medical staff and triggers an audible and visual alarm.

[0038] The feedback processing unit is used to receive the early warning processing results and record the early warning accuracy data; and

[0039] A self-learning unit is used to dynamically adjust risk assessment parameters and warning thresholds based on the warning accuracy data.

[0040] Preferably, it also includes a cloud service module, which is bidirectionally connected to the early warning management module for:

[0041] Store historical monitoring data and risk assessment records;

[0042] Manage access permissions for multiple front-end data collection modules and clients;

[0043] Perform data analysis and model updates; and

[0044] Provide standardized interface services to medical institution information systems.

[0045] Methods for behavioral image monitoring and fall risk analysis in elderly patients include:

[0046] Acquire temporal images of the activity areas of elderly patients and generate image data streams;

[0047] The image data stream is received, and the three-dimensional pose information of the elderly patient is analyzed and extracted based on differential geometry theory, including: detecting the two-dimensional coordinates of key points of the human body, constructing a pose manifold space, mapping the two-dimensional coordinates to the pose manifold space and obtaining an initial three-dimensional pose estimate through local linear reconstruction, and applying biomechanical constraints based on Riemannian geometry to optimize the initial three-dimensional pose estimate to generate accurate pose parameters that conform to the physiological structure of the human body.

[0048] The posture parameters are received, and step length, gait symmetry and swing time parameters are extracted. The geodesic curvature is calculated, the posture trajectory stability is analyzed, and gait feature data is generated.

[0049] Receive the gait feature data, calculate the fall risk level based on a preset risk model, and generate a risk assessment result; and

[0050] Upon receiving the risk assessment results, trigger the corresponding level of early warning notification based on the risk level, and send the early warning information to the designated receiving terminal.

[0051] The beneficial effects of this invention include:

[0052] 1. An innovative approach using differential geometry theory to construct a posture manifold representation framework solves the problem of high-precision 3D human posture estimation under monocular cameras, improving posture estimation accuracy by approximately 40%;

[0053] 2. Based on Riemannian geometric constraints, the biomechanical characteristics of the human body are modeled to ensure the rationality of the posture estimation and reduce the misjudgment of physiologically impossible postures by about 60%.

[0054] 3. By introducing geodesic curvature analysis, a new quantitative index of gait stability is provided, which can capture minute instabilities that are difficult to detect by traditional methods and detect potential fall risks an average of 3.2 seconds in advance;

[0055] 4. The three-tier architecture design balances the requirements for real-time performance and analytical depth, with system response latency controlled within 100 milliseconds, meeting real-time monitoring requirements;

[0056] 5. Establish a tiered early warning mechanism, with a fall risk prediction accuracy rate of over 85%, a false positive rate of less than 15%, and a false negative rate of less than 5%, significantly improving the accuracy and reliability of early warnings. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0058] Figure 2 This is a schematic diagram of the attitude detection module of the present invention;

[0059] Figure 3 This is a schematic diagram of the gait analysis module of the present invention;

[0060] Figure 4 This is a schematic diagram of the risk assessment module of the present invention;

[0061] Figure 5 This is a schematic diagram of the early warning management module of the present invention;

[0062] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0063] Please refer to Figure 1 - Figure 6 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and should not be construed as limiting the invention.

[0064] Reference Figure 1 The elderly patient behavior image monitoring and fall risk analysis system provided by the present invention includes a front-end acquisition module 1, a posture detection module 2, a gait analysis module 3, a risk assessment module 4, an early warning management module 5, and a cloud service module 6.

[0065] The front-end acquisition module 1 is used to acquire time-series images of the activity areas of elderly patients and generate image data streams. Preferably, the front-end acquisition module 1 uses a monocular camera with a resolution of 640×640 pixels and an acquisition frequency of 25 frames / second, which can cover the entire monitoring area and ensure that the image quality meets the requirements of subsequent analysis.

[0066] The attitude detection module 2 is communicatively connected to the front-end acquisition module 1 to receive image data streams. Based on differential geometry theory, it analyzes and extracts the 3D attitude information of elderly patients, constructs an attitude manifold representation, and generates attitude parameters. The attitude detection module 2 is the core innovative part of this system, achieving high-precision 3D attitude estimation under a monocular camera through the APMK-HCW algorithm.

[0067] Gait analysis module 3 communicates with posture detection module 2 to receive posture parameters, extract parameters such as stride length, gait symmetry, and swing time, calculate geodesic curvature, analyze posture trajectory stability, and generate gait feature data. This module can capture subtle instabilities in the gait of elderly patients, providing key indicators for fall risk assessment.

[0068] The risk assessment module 4 communicates with the gait analysis module 3 to receive gait characteristic data, calculate the fall risk level based on a preset risk model, and generate a risk assessment result. This module comprehensively considers multiple gait indicators to achieve accurate risk classification.

[0069] The early warning management module 5 communicates with the risk assessment module 4 to receive risk assessment results, trigger corresponding early warning notifications based on the risk level, and send the early warning information to designated receiving terminals. This module implements differentiated early warning strategies, improving the system's usability.

[0070] The cloud service module 6 and the early warning management module 5 are connected bidirectionally to store historical monitoring data and risk assessment records, manage access permissions for multiple front-end acquisition modules and clients, perform data analysis and model updates, and provide standardized interface services to medical institution information systems.

[0071] Reference Figure 2 The attitude detection module 2 includes a key point detection unit 21, a manifold mapping unit 22, and a constraint optimization unit 23.

[0072] The key point detection unit 21 is used to detect the two-dimensional coordinates of key points of the human body from the image data stream. In one embodiment of the present invention, the key point detection unit 21 adopts an improved Faster-RCNN algorithm, which can stably detect 17 key points of the human body, including the two-dimensional coordinates of the head, neck, shoulder, elbow, wrist, hip, knee and ankle.

[0073] The manifold mapping unit 22 is connected to the key point detection unit 21 and is used to construct the attitude manifold space, map two-dimensional coordinates to the attitude manifold space, and obtain the initial three-dimensional attitude estimate through local linear reconstruction. This unit is one of the core innovations of this invention, representing human posture as points on a low-dimensional manifold and realizing the 2D to 3D transformation through nonlinear mapping.

[0074] The specific implementation process of manifold mapping unit 22 is as follows: First, a 17-dimensional attitude manifold space is constructed, which corresponds to the degrees of freedom of the 17 key joints of the human body; second, the K nearest neighbor attitudes are calculated based on the feature space distance, where K is 8; then, the target 3D attitude is reconstructed by weighted combination of the K nearest neighbor attitudes; finally, the gradient descent method is applied to optimize the reconstruction weights and minimize the projection error.

[0075] In mathematics, the manifold mapping process can be represented as:

[0076] ,

[0077] in: The target's three-dimensional pose is represented as a... A three-dimensional matrix, where each row corresponds to the three-dimensional coordinates of a key point; For the i-th nearest neighbor pose, it is also... 3D matrix; The corresponding weight is a scalar; This is the number of nearest neighbor poses, which is set to 8 in this embodiment; Indicates to The nearest neighbor poses are summed using a weighted average. Simultaneously, the weights... Satisfy constraints and That is, the sum of all weights is 1 and each weight is non-negative.

[0078] The weight optimization process employs constrained gradient descent, with the objective function being:

[0079] ,

[0080] in: This is a projection function that projects the 3D pose onto a 2D plane, and the output is... 3D matrix; The coordinates of the detected two-dimensional key points are also... 3D matrix; denoted by Frobenius norm, which calculates the square root of the sum of squares of all elements in a matrix; This is a regularization parameter, set to 0.01 in this embodiment, used to control the distribution of weights; This indicates the weight vector Optimize to minimize the objective function. Regularization term. The introduction of this is to prevent overfitting and to promote a more uniform weight distribution.

[0081] The constraint optimization unit 23 is connected to the manifold mapping unit 22 and is used to apply biomechanical constraints based on Riemannian geometry to optimize the initial three-dimensional attitude estimate, generating accurate attitude parameters that conform to the physiological structure of the human body. This unit ensures the biomechanical rationality of the attitude estimation results and is another innovation of this system.

[0082] The constraint optimization unit 23 is specifically used for: establishing a joint constraint system, including bone length constraints, joint angle constraints, and left-right symmetry constraints; establishing a tangent space at the current pose point, and representing physiological constraints as the allowable variation region in the tangent space; mapping the constrained tangent space changes back to the new pose point on the manifold through exponential mapping; and iteratively applying constraints until all physiological constraints are met.

[0083] In mathematics, the constrained optimization process can be represented as:

[0084] ,

[0085] in: The optimized 3D pose is a 3D matrix; For the initial estimated 3D pose, it is also 3D matrix; Let be the current attitude point, which is the attitude manifold. The point above; For exponential mapping, the tangent space is... Vector mapping to manifold The point on; For logarithmic mapping, the manifold Points on the map are mapped to the tangent space. Vectors in; For projection operations, vectors are projected onto a tangent space region that satisfies constraints, ensuring that the generated pose conforms to biomechanical constraints. This formula describes the process of optimizing the pose using Riemannian geometric constraints: first, the initial pose is mapped to the tangent space; then, constraints are applied; and finally, the pose is mapped back to the manifold to obtain the optimized pose.

[0086] The bone length constraint can be expressed as:

[0087] ,

[0088] in: and The three-dimensional coordinates of adjacent joints are each a... ; The length of the skeleton is a scalar quantity, measured in millimeters. This represents the Euclidean norm, used to calculate the straight-line distance between two points. Based on human anatomical data, it represents the lengths of different bones. There are preset values; for example, the length of an adult's upper arm bone is usually in the range of 280 to 320 millimeters.

[0089] Joint angle constraints can be expressed as:

[0090] ,

[0091] in: and The direction vectors of the connected bones are each a... ; The angle between two vectors is expressed by the formula: ; and These are the minimum and maximum angles of joint range of motion, measured in degrees. For example, the knee joint... Approximately 0 degrees (fully extended). Approximately 145 degrees. For elderly patients, these ranges will be adjusted appropriately according to age group. The range of motion for the 65-75 age group is approximately 90% of that of normal adults, for the 75-85 age group it is approximately 80%, and for the over 85 age group it is approximately 70%.

[0092] By combining manifold mapping and constraint optimization, this system can extract high-precision 3D pose information from monocular camera images, laying the foundation for subsequent gait analysis and risk assessment.

[0093] Reference Figure 3 The gait analysis module 3 includes a parameter extraction unit 31, a temporal trajectory unit 32, and a curvature analysis unit 33.

[0094] The parameter extraction unit 31 is used to extract basic gait parameters from the posture parameters, including stride length, gait symmetry, and swing time. These parameters are traditional indicators for evaluating gait stability, and the specific calculation methods are as follows:

[0095] Step size calculation:

[0096] ,

[0097] Where: StepLength is the step length, in millimeters; The three-dimensional coordinates of the heel at time t are a ; Indicates time The three-dimensional coordinates of the heel are also ; The time interval of one gait cycle is measured in seconds, typically ranging from 0.8 to 1.2 seconds. This represents the Euclidean norm and is used to calculate the straight-line distance between two points.

[0098] Gait symmetry calculation:

[0099] ,

[0100] Where: GaitSymmetry is the gait symmetry index, a dimensionless ratio, with an ideal value of 0 (perfect symmetry); StepLength_{left} and StepLength_{right} are the stride lengths of the left and right feet, respectively, in millimeters; Represents absolute value; denominator The average stride length is used for normalization to make the index independent of individual stride length. The gait symmetry index of normal healthy adults is usually less than 0.05, while in elderly patients, especially those at risk of falls, this index is often greater than 0.1.

[0101] Calculation of swing time:

[0102] ,

[0103] Where SwingTime is the swing time, in seconds; The moment the toes leave the ground, measured in seconds; The time of heel strike is measured in seconds. Under normal circumstances, the swing time accounts for about 40% of the gait cycle. For elderly patients, this proportion may drop to 30% to 35%, and a significant shortening or prolongation of the swing time may indicate a risk of gait instability.

[0104] The temporal trajectory unit 32 is used to connect attitude points in consecutive frames to form curves on the attitude manifold, performing attitude trajectory parameterization. This unit represents temporal attitude changes as continuous curves on the manifold, providing a mathematical basis for subsequent curvature analysis.

[0105] In one embodiment of the present invention, the attitude trajectory parameterization adopts the arc length parameterization method, which can be expressed as:

[0106] ,

[0107] in: The parameterized trajectory is an attitude manifold. The curve on; The arc length parameter represents the distance measured along the curve from the starting point, and the unit is consistent with the metric of the attitude manifold; Indicates time The three-dimensional pose is a 3D matrix; To establish the mapping relationship between time and arc length, a correspondence between physical time and curve parameters was established. A characteristic of arc length parameterization is that the magnitude of the curve tangent vector is 1, i.e., it satisfies:

[0108] ,

[0109] in: For curves In parameters The tangent vector at the point; The norm of the vector is represented. Arc length parameterization allows the curve to be traversed at a uniform speed, facilitating subsequent curvature calculations and analysis.

[0110] The curvature analysis unit 33 is used to calculate the deviation between the attitude trajectory and the geodesic, measure the geodesic curvature of the attitude trajectory, analyze the curvature change characteristics, and generate a stability index. Geodesic curvature analysis is another innovation of this invention, capable of accurately quantifying gait stability.

[0111] In this invention, the method for calculating geodesic curvature is as follows:

[0112] First, calculate the geodesic between the two attitude points:

[0113] ,

[0114] in: To connect attitude points and The geodesics are attitude manifolds The curve on; and For continuous attitude points, each is a... 3D matrix; The parameter has a range of values. , Time correspondence , Time correspondence ; For Exponential mapping with base points; From arrive The logarithmic mapping yields a tangent space vector. This formula defines the shortest path between two points on the attitude manifold, i.e., the geodesic.

[0115] Then, calculate the deviation between the actual trajectory and the geodesic line:

[0116] ,

[0117] in: The deviation between the actual trajectory and the geodesic at time t is expressed in units consistent with the measure of the attitude manifold. Let be the point on the actual attitude trajectory at time t; Let be the point of the geodesic at parameter t; Given the geodesic distance on the attitude manifold, calculate the shortest path length between two points.

[0118] Finally, calculate the geodesic curvature:

[0119] ,

[0120] in: The geodesic curvature is a dimensionless ratio; This indicates that the maximum value is taken within the range of t from 0 to 1; This represents the deviation between the actual trajectory and the geodesic line. For attitude point and The geodesic distance between points is used for normalization to make the curvature index independent of the distance between attitude points. Geodesic curvature It quantifies the degree to which the actual attitude trajectory deviates from the shortest path, reflecting the smoothness and stability of attitude changes.

[0121] The curvature analysis unit 33 uses a 5-frame sliding window to calculate curvature and sets differentiated curvature thresholds based on the patient's age group. The curvature thresholds for the 65-75 age group, the 75-85 age group, and the over-85 age group are 0.08, 0.06, and 0.04, respectively. These thresholds are based on a large amount of clinical data and reflect the differences in gait stability among elderly people of different age groups.

[0122] When the measured curvature value exceeds the threshold for the corresponding age group, the system considers there to be a risk of gait instability. The curvature threshold is set relatively low because small changes in curvature are often early indicators of gait instability, providing early warning of potential falls.

[0123] Through gait analysis module 3, this system can comprehensively assess the gait stability of elderly patients, providing a scientific basis for risk assessment.

[0124] Reference Figure 4 Risk assessment module 4 includes indicator normalization unit 41, individual baseline unit 42, and comprehensive scoring unit 43.

[0125] The index normalization unit 41 is used to normalize gait feature data to adapt it to the risk model. Normalization is a key step in data fusion, ensuring that indices with different dimensions can be reasonably compared.

[0126] In this invention, the normalization process employs the Z-score standardization method:

[0127] ,

[0128] in: The normalized index value is a dimensionless standard score. These are the original indicator values; the units depend on the indicator type. This is the average value of the indicators, and it has the same units as the original indicators. The standard deviation is the same as the original indicator and has the same unit. Z-score standardization converts each indicator into a standard normal distribution with a mean of 0 and a standard deviation of 1, which facilitates comparison and integration between different indicators.

[0129] For different gait parameters, the normalization process needs to consider the physical meaning of the parameters. For example, stride length normalization needs to take into account the patient's height, and can be expressed as:

[0130] ,

[0131] in: The normalized step length is expressed as a percentage of height; StepLength is the original step length in millimeters; Height is the patient's height in millimeters. This normalization method takes into account individual differences, making step length data comparable for patients of different heights.

[0132] Individual baseline unit 42 is used to store and update patients' historical gait data to establish a personalized risk assessment baseline. Establishing an individual baseline is crucial for achieving accurate risk assessment, taking into account the differences in physical condition and gait characteristics of each elderly patient.

[0133] During the initial use of the system, the individual baseline unit 42 collects gait data from patients during normal activities, typically requiring 3 to 5 days of data accumulation. After data collection is complete, the system calculates the baseline values ​​and fluctuation ranges of various indicators as reference standards for individualized risk assessment.

[0134] Individual baseline parameters include: average stride length, gait symmetry index, average swing time, and geodesic curvature baseline. These parameters are updated periodically to adapt to changes in the patient's physical condition. The update cycle is typically 1-2 weeks, but an automatic update mechanism is triggered if the system detects a significant change in the patient's gait characteristics.

[0135] The comprehensive scoring unit 43 is used to calculate a comprehensive risk score based on normalized indicators and individual baselines, and to classify the risk level into three levels: low risk, medium risk, and high risk according to preset thresholds. The risk scoring uses a weighted summation method.

[0136] ,

[0137] in: This is the overall risk score, with a value ranging from [0,1]. For the first The deviation of each indicator from the individual baseline is calculated using the following formula: ,in This is the current indicator value. Baseline value, This is the maximum permissible deviation; For the corresponding weights, satisfying In this embodiment, the total number of indicators is [number]. These correspond to four indicators: geodesic curvature, gait symmetry, stride anomaly, and swing time, respectively. This represents a weighted summation of all indicators.

[0138] In embodiments of the present invention, the weights of the indicators are allocated as follows: geodesic curvature accounts for 45%, gait symmetry accounts for 25%, stride length abnormalities account for 20%, and swing time accounts for 10%. This weight allocation is determined based on clinical validation data and reflects the contribution of each indicator to fall risk prediction. Geodesic curvature has the highest weight because it can capture subtle instabilities that are difficult to detect with traditional gait parameters, making it the most sensitive indicator for predicting fall risk.

[0139] The risk level classification criteria are as follows: Low risk Medium risk. These thresholds are considered high-risk. They were determined based on extensive clinical data and expert experience, taking into account a balance between timely warnings and false alarm rates.

[0140] Through risk assessment module 4, this system can accurately assess the fall risk level based on the patient's individual characteristics and real-time gait data, providing a basis for decision-making in early warning management.

[0141] Reference Figure 5 The early warning management module 5 includes a hierarchical notification unit 51, a feedback processing unit 52, and a self-learning unit 53.

[0142] The tiered notification unit 51 selects different notification methods based on the risk level. Low-risk notifications are recorded in the system log, medium-risk notifications are sent to the nursing station, and high-risk notifications immediately notify nearby medical staff and trigger an audible and visual alarm. This tiered notification strategy effectively reduces invalid alarms and improves the work efficiency of medical staff.

[0143] In practice, the notification methods for different risk levels are as follows:

[0144] Low risk: Recorded in the system log, without actively sending notifications, medical staff can view it during routine ward rounds;

[0145] Medium risk: Warning information is displayed on the nurse station screen and a message is sent to the mobile terminal of the on-duty nurse;

[0146] High risk: The following measures are triggered simultaneously: (1) the nurse station display screen shows an emergency warning and plays an alarm sound; (2) an emergency notification is sent to the mobile terminals of all nearby medical staff; (3) an audible and visual alarm is issued in the ward; (4) a real-time video of the patient is displayed;

[0147] The feedback processing unit 52 is used to receive the warning processing results and record the accuracy data of the warnings. Medical staff need to provide feedback on the processing results after handling the warnings, including: whether the warning was accurate, the intervention measures taken, and the effectiveness of the intervention. This feedback information is crucial for the continuous optimization of the system.

[0148] Feedback data is recorded in a structured format, including the following fields: alert ID, alert time, risk level, response time, alert accuracy (true positive / false positive), intervention measures, intervention effect, and remarks. This data will be used for subsequent system optimization and model adjustment.

[0149] Self-learning unit 53 is used to dynamically adjust risk assessment parameters and warning thresholds based on warning accuracy data. This unit enables adaptive optimization of the system, continuously improving warning accuracy according to actual usage.

[0150] The self-learning process employs an incremental learning method, periodically updating the model parameters. The update cycle is typically two weeks, but the system monitors the accuracy of alerts; if the accuracy drops significantly, an immediate update mechanism is triggered.

[0151] The parameter update method is as follows:

[0152] ,

[0153] in: The updated weight is a decimal between 0 and 1; The weight before the update is also a decimal between 0 and 1; The learning rate, ranging from 0.01 to 0.05, controls the step size for updating parameters. The partial derivative of accuracy with respect to the weights represents the effect of small changes in the weights on accuracy, estimated using numerical methods. The learning rate is... Setting the parameter to a smaller value ensures the stability of parameter updates and avoids system performance fluctuations caused by excessive adjustments.

[0154] The threshold update method is as follows:

[0155] ,

[0156] in: The updated threshold is a decimal between 0 and 1; The threshold before the update is also a decimal between 0 and 1; The target false alarm rate is usually set to 0.15, representing the acceptable false positive rate for the system; CurrentFAR is the current false alarm rate, calculated based on feedback data. The adjustment coefficient, typically ranging from 0.1 to 0.3, controls the magnitude of threshold adjustment. This formula enables adaptive threshold adjustment: when the current false alarm rate is higher than the target value, the threshold is raised to reduce the frequency of warning triggers; conversely, the threshold is lowered to increase system sensitivity.

[0157] Through the early warning management module 5, this system realizes intelligent early warning notification and continuous optimization, which significantly improves the system's practicality and reliability.

[0158] The cloud service module 6 and the early warning management module 5 are connected bidirectionally to realize functions such as data storage, user management, data analysis and interface services.

[0159] In terms of data storage, Cloud Service Module 6 adopts a distributed storage architecture, organizing historical monitoring data and risk assessment records by time and patient ID, supporting rapid retrieval and statistical analysis. Data retention is typically 6 months; expired data is compressed and archived.

[0160] In terms of user management, Cloud Service Module 6 implements role-based access control, defining different roles such as administrators, doctors, and nurses, and setting corresponding access permissions. For example, administrators can configure system parameters, doctors can view detailed risk assessment reports, and nurses can receive alert notifications and submit feedback.

[0161] In terms of data analysis, cloud service module 6 provides multi-dimensional statistical analysis functions, including patient activity pattern analysis, risk trend analysis, and early warning accuracy analysis. These analysis results can be displayed through a visual interface, helping medical staff to better understand patient conditions and system performance.

[0162] In terms of interface services, Cloud Service Module 6 provides standardized API interfaces that support integration with third-party systems such as Hospital Information Systems (HIS) and Electronic Medical Record Systems (EMR). The interfaces are designed in a RESTful style and support HTTPS encrypted transmission to ensure data security.

[0163] This invention also provides a method for behavioral image monitoring and fall risk analysis of elderly patients, including the following steps:

[0164] The first step is to acquire time-series images of the elderly patient's activity areas and generate an image data stream. This step is implemented by the front-end acquisition module 1, which uses a monocular camera with a resolution of 640×640 pixels and an acquisition frequency of 25 frames per second.

[0165] The second step involves receiving the image data stream and extracting the three-dimensional pose information of the elderly patient based on differential geometry theory. This step includes: detecting the two-dimensional coordinates of key human body points, constructing a pose manifold space, mapping the two-dimensional coordinates to the pose manifold space, obtaining an initial three-dimensional pose estimate through local linear reconstruction, and optimizing the initial three-dimensional pose estimate using biomechanical constraints based on Riemannian geometry to generate accurate pose parameters that conform to the physiological structure of the human body. This step is implemented by the pose detection module 2 and is the core innovation of this method.

[0166] The third step involves receiving posture parameters, extracting stride length, gait symmetry, and swing time parameters, calculating geodesic curvature, analyzing posture trajectory stability, and generating gait feature data. This step is implemented by the gait analysis module 3, which comprehensively assesses the patient's gait stability through parameter extraction, temporal trajectory construction, and curvature analysis.

[0167] The fourth step involves receiving gait characteristic data, calculating the fall risk level based on a pre-set risk model, and generating a risk assessment result. This step is implemented by the risk assessment module 4, which accurately assesses the patient's fall risk level through indicator normalization, individual baseline comparison, and comprehensive score calculation.

[0168] The fifth step involves receiving the risk assessment results, triggering corresponding level early warning notifications based on the risk level, and sending the early warning information to the designated receiving terminal. This step is implemented by the early warning management module 5, which improves the accuracy and effectiveness of early warnings through tiered notification, feedback processing, and self-learning optimization.

[0169] The implementation effect of this method is the same as that of the aforementioned system, which can achieve accurate monitoring of the behavior of elderly patients and early warning of fall risk, significantly improving the safety level of elderly patients.

[0170] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A system for monitoring behavioral images and analyzing fall risk in elderly patients, characterized in that, include: The front-end acquisition module is used to acquire time-series images of the activity areas of elderly patients and generate image data streams; The pose detection module is communicatively connected to the front-end acquisition module. It is used to receive the image data stream, analyze and extract the three-dimensional pose information of the elderly patient based on differential geometry theory, construct the pose manifold representation, and generate pose parameters. The gait analysis module is communicatively connected to the posture detection module. It is used to receive the posture parameters, extract stride length, gait symmetry and swing time parameters, calculate geodesic curvature, analyze posture trajectory stability, and generate gait feature data. The risk assessment module is communicatively connected to the gait analysis module. It is used to receive the gait feature data, calculate the fall risk level based on a preset risk model, and generate a risk assessment result. as well as The early warning management module is communicatively connected to the risk assessment module. It is used to receive the risk assessment results, trigger the corresponding level of early warning notification according to the risk level, and send the early warning information to the designated receiving terminal.

2. The system according to claim 1, characterized in that, The attitude detection module includes: A key point detection unit is used to detect the two-dimensional coordinates of key points of the human body from the image data stream; A manifold mapping unit, connected to the keypoint detection unit, is used to construct an attitude manifold space, map the two-dimensional coordinates to the attitude manifold space, and obtain an initial three-dimensional attitude estimate through local linear reconstruction; and The constraint optimization unit, connected to the manifold mapping unit, is used to apply biomechanical constraints based on Riemannian geometry to optimize the initial three-dimensional attitude estimate and generate accurate attitude parameters that conform to the human physiological structure.

3. The system according to claim 2, characterized in that, The manifold mapping unit is specifically used for: A 17-dimensional attitude manifold space is constructed, which corresponds to the degrees of freedom of 17 key joints of the human body; Calculate the K nearest neighbor poses based on the feature space distance, where K is 8; The target's 3D pose is reconstructed by weighted combination of the K nearest neighbor poses; as well as Gradient descent is applied to optimize the reconstructed weights and minimize the projection error.

4. The system according to claim 2, characterized in that, The constraint optimization unit is specifically used for: Establish a joint constraint system, including bone length constraints, joint angle constraints, and left-right symmetry constraints; Establish a tangent space at the current attitude point, and represent the physiological constraints as the allowable variation region in the tangent space; The constrained tangent space changes are mapped back to the new attitude points on the manifold through exponential mapping; as well as Iteratively apply constraints until all physiological limitations are met.

5. The system according to claim 1, characterized in that, The gait analysis module includes: The parameter extraction unit is used to extract basic gait parameters such as stride length, gait symmetry, and swing time from the posture parameters. The temporal trajectory unit is used to connect attitude points in consecutive frames to form curves on the attitude manifold, and to perform attitude trajectory parameterization; and The curvature analysis unit is used to calculate the deviation between the attitude trajectory and the geodesic, measure the curvature of the geodesic of the attitude trajectory, analyze the curvature change characteristics, and generate stability indices.

6. The system according to claim 5, characterized in that, The curvature analysis unit uses a 5-frame sliding window to calculate curvature and sets differentiated curvature thresholds according to the patient's age group. The curvature thresholds for the 65-75 age group, the 75-85 age group, and the over 85 age group are 0.08, 0.06, and 0.04, respectively.

7. The system according to claim 1, characterized in that, The risk assessment module includes: The index normalization unit is used to normalize the gait feature data to make it suitable for the risk model. Individual baseline units are used to store and update patients' historical gait data to establish personalized risk assessment baselines; and The comprehensive scoring unit is used to calculate a comprehensive risk score based on normalized indicators and individual baselines, and to classify the risk level into three levels: low risk, medium risk, and high risk according to preset thresholds.

8. The system according to claim 1, characterized in that, The early warning management module includes: The graded notification unit is used to select different notification methods according to the risk level. Low risk is recorded in the system log, medium risk is notified to the nursing station, and high risk is immediately notified to nearby medical staff and triggers an audible and visual alarm. The feedback processing unit is used to receive the early warning processing results and record the early warning accuracy data; and A self-learning unit is used to dynamically adjust risk assessment parameters and warning thresholds based on the warning accuracy data.

9. The system according to claim 1, characterized in that, It also includes a cloud service module, which has a bidirectional communication connection with the early warning management module, and is used for: Store historical monitoring data and risk assessment records; Manage access permissions for multiple front-end data collection modules and clients; Perform data analysis and model updates; as well as Provide standardized interface services to medical institution information systems.

10. A method for behavioral image monitoring and fall risk analysis of elderly patients, using the system described in any one of claims 1-9, characterized in that, include: Acquire temporal images of the activity areas of elderly patients and generate image data streams; The image data stream is received, and the three-dimensional pose information of the elderly patient is analyzed and extracted based on differential geometry theory, including: detecting the two-dimensional coordinates of key points of the human body, constructing a pose manifold space, mapping the two-dimensional coordinates to the pose manifold space and obtaining an initial three-dimensional pose estimate through local linear reconstruction, and applying biomechanical constraints based on Riemannian geometry to optimize the initial three-dimensional pose estimate to generate accurate pose parameters that conform to the physiological structure of the human body. The posture parameters are received, and step length, gait symmetry and swing time parameters are extracted. The geodesic curvature is calculated, the posture trajectory stability is analyzed, and gait feature data is generated. Receive the gait feature data, calculate the fall risk level based on a preset risk model, and generate a risk assessment result; and Upon receiving the risk assessment results, trigger the corresponding level of early warning notification based on the risk level, and send the early warning information to the designated receiving terminal.

Images