A fall prevention early warning system based on intrinsic capacity assessment of the elderly

Abstract

The application discloses a kind of based on the anti-falling early warning system of old person internal capacity evaluation, it is related to intelligent health monitoring technical field, including data acquisition module, acceleration data and angular velocity data and camera image data in the walking process of old person near the edge of narrow corridor carpet are collected by multimodal sensor, initial multi-source signal sequence is formed by fusion data to obtain preliminary risk feature set, feature extraction module, according to preliminary risk feature set using convolutional neural network extracts spatial features, separates sequence from acceleration data and angular velocity data Dynamic stage change mode is determined phase separation sequence;The anti-falling early warning system based on the old person internal capacity evaluation, realizes the continuous balance monitoring and individualized intervention under dynamic environment, effectively reduces the risk of falling of old people.

Description

Technical Field

[0001] This invention relates to the field of intelligent health monitoring technology, specifically to a fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly. Background Technology

[0002] Fall detection and intervention are crucial for ensuring the safety of the elderly. Falls in the elderly can lead to serious consequences such as fractures and head injuries, threatening not only their lives and health but also significantly increasing the medical burden and social costs. Developing intelligent systems capable of identifying fall risks in real time and intervening promptly is a key direction for improving the quality of life for the elderly and reducing the risk of accidental injuries.

[0003] Current research largely focuses on capturing fall movements using sensors or cameras. However, existing methods are limited by their inability to accurately distinguish the dynamic stages of a fall and their insufficient ability to capture subtle risk signals in complex environments. These shortcomings can cause systems to miss critical intervention opportunities in practical applications, reducing their protective effectiveness. Existing solutions lack adaptability to dynamic scenarios, especially in the daily lives of the elderly, where environments are complex and changeable, such as indoor furniture obstructing the view or uneven outdoor terrain, leading to significant interference with sensor data. Furthermore, the movement patterns of the elderly vary considerably from person to person, making it difficult for the system to accurately determine which movement signals indicate an impending fall. This interaction between complex environments and individual differences constitutes the primary technical challenge in identifying critical interruption points in the fall process. Moreover, since the fall process is often rapid, taking only a few seconds from loss of balance to the completion of the fall, the system needs to complete signal analysis, risk assessment, and intervention triggering within an extremely short time window. Accurately capturing and processing multi-source heterogeneous data within such a short time window becomes another core technical challenge. Therefore, in real-world business scenarios, such as when an elderly person loses balance due to carpet folds while walking at home, the system needs to instantly distinguish between a normal gait and a pre-fall warning from acceleration, angular velocity, and other data received from sensors, and quickly decide whether to issue a warning or activate protective devices. Existing technologies lack the ability to quickly analyze weak signals in complex environments and are poorly adaptable to individual differences in movement, leading to delayed warnings or frequent false alarms. Summary of the Invention

[0004] The purpose of this invention is to provide a fall prevention early warning system based on the assessment of the elderly’s intrinsic abilities, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a fall prevention early warning system based on the assessment of the elderly's intrinsic abilities, comprising a data acquisition module that collects acceleration and angular velocity data of elderly people walking near the edge of a carpet in a narrow corridor, as well as camera image data, through multimodal sensors, and fuses the data to form an initial multi-source signal sequence to obtain a preliminary risk feature set; a feature extraction module that extracts spatial features based on the preliminary risk feature set using a convolutional neural network, and separates dynamic stage change patterns from acceleration and angular velocity data to determine a stage separation sequence; an imbalance judgment module that processes the time series dependency of the stage separation sequence using a recurrent neural network, captures the fluctuation pattern of the signal in the environment, and judges potential imbalance indicators; and a personalized assessment module that, if the potential imbalance indicator exceeds a preset threshold, retrieves historical movement patterns from an individual difference database and matches and adjusts the current stage separation sequence to obtain a more accurate assessment result. The system comprises a personalized risk scoring module; an environmental analysis module that combines personalized risk scores with real-time image data captured by cameras to analyze environmental interference factors such as furniture obstruction or uneven terrain to determine the environmental adaptation coefficient; an intervention judgment module that corrects the personalized risk score based on the environmental adaptation coefficient and integrates angular velocity data from multi-source signal sequences to determine immediate intervention needs; an alarm and response module that triggers the system alarm device if immediate intervention needs are confirmed and generates a protective instruction sequence from the judged risk score to obtain the final intervention response; a database update module that extracts signal fluctuation patterns and obstruction impact assessments from the final intervention response and updates the individual difference database to optimize the historical movement pattern matching and adjustment process; and a continuous monitoring module that processes subsequent multi-source signal sequences for complex environmental dynamic changes through the updated individual difference database, integrating potential imbalance index judgments and terrain unevenness adaptation coefficients to continuously monitor the elderly's walking balance.

[0006] Preferably, the data acquisition module collects acceleration and angular velocity data of the elderly person walking near the edge of the carpet in a narrow corridor, as well as camera image data, using a multimodal sensor. The data is then fused to form an initial multi-source signal sequence to obtain a preliminary risk feature set. This includes collecting acceleration, angular velocity, and camera image data of the elderly person walking using an inertial measurement unit and a camera device to obtain an original multimodal dataset; synchronizing the acceleration, angular velocity, and camera image data using a timestamp alignment method, and fusing the multimodal data using a Kalman filter algorithm to obtain an initial signal sequence; extracting peak features from the acceleration data, rate of change features from the angular velocity data, and edge contour features from the camera image based on the initial signal sequence; identifying potential risk features if the feature values ​​exceed a preset threshold, thus obtaining a preliminary risk feature set; and clustering the preliminary risk feature set using a K-means algorithm to determine the risk characteristics of the elderly person walking, resulting in a classified risk feature set.

[0007] Preferably, the feature extraction module extracts spatial features using a convolutional neural network based on a preliminary risk feature set, and separates dynamic stage change patterns from acceleration and angular velocity data to determine the stage separation sequence. This includes obtaining acceleration and angular velocity data from multimodal data, performing convolution operations on the acceleration and angular velocity data using a convolutional neural network to obtain a spatial feature set; using principal component analysis to reduce the dimensionality of the spatial features based on the spatial feature set to obtain a dimensionality-reduced feature vector set; using a dynamic time warping algorithm to compare the temporal differences between the feature vectors to determine the dynamic stage change pattern; if the temporal difference of the dynamic stage change pattern exceeds a preset threshold, then a clustering algorithm is used to group the change patterns to obtain the stage separation sequence.

[0008] Preferably, the imbalance judgment module, for the stage-separated sequence, processes the time series dependency using a recurrent neural network to capture the signal fluctuation patterns in the environment and judge potential imbalance indicators. This includes obtaining time series data from a multimodal data stream, using a recurrent neural network to perform time-series modeling on the time series data to obtain a dynamic dependency feature set; for the dynamic dependency feature set, segmenting the time series data using a sliding window method to obtain the local fluctuation patterns of the signal; if the amplitude of the local fluctuation pattern exceeds a preset threshold, performing a frequency domain transformation on the local fluctuation pattern using a fast Fourier transform to determine the signal frequency distribution in the environment; and based on the signal frequency distribution, using a clustering algorithm to group the fluctuation patterns and determine potential imbalance indicators.

[0009] Preferably, the personalized assessment module, if the potential imbalance index exceeds a preset threshold, retrieves historical action patterns from the individual difference database, matches and adjusts the current stage separation sequence to obtain a personalized risk score. This includes retrieving historical action patterns from the individual difference database, using a data stream processing method to segment and extract behavioral pattern features from the time series data; if the deviation between the behavioral pattern features and the stage separation sequence exceeds a preset threshold, a dynamic matching algorithm is used to compare the stage separation sequence with the historical action patterns to obtain pattern adjustment parameters; based on the pattern adjustment parameters, the stage separation sequence is corrected, and a weighted average method is used to fuse the behavioral pattern features to generate an adjusted sequence feature set; the adjusted sequence feature set is scored using a risk assessment model, and the k-nearest neighbor algorithm is used to quantify the potential imbalance index to determine the personalized risk score.

[0010] Preferably, the environmental analysis module combines personalized risk scores with real-time image data captured by a camera to analyze environmental interference factors such as furniture obstruction or uneven terrain to determine the environmental adaptability coefficient. This includes obtaining an original image from the real-time image data captured by the camera, removing noise and blur through image preprocessing to obtain a first image; for the first image, using feature extraction methods to identify image features of obstruction and uneven terrain to obtain an environmental interference feature set; if the deviation between the environmental interference feature set and a preset threshold exceeds a standard, then using a dynamic matching algorithm to compare the environmental interference feature set with behavioral pattern features to obtain environmental adjustment parameters; based on the environmental adjustment parameters, using a weighted average method to fuse the personalized risk score and the environmental interference feature set to obtain the environmental adaptability coefficient.

[0011] Preferably, the intervention judgment module, in accordance with the environmental adaptation coefficient, corrects the personalized risk score and determines the immediate intervention requirement by fusing angular velocity data from multi-source signal sequences. This includes obtaining angular velocity data from the multi-source signal sequences, denoising and standardizing the angular velocity data through real-time data processing to obtain a first angular velocity dataset; extracting motion features from the first angular velocity dataset using a behavior pattern recognition method, dynamically matching them with a preset behavior pattern library to obtain behavior pattern parameters; fusing the personalized risk score using a weighted average method based on the behavior pattern parameters and the environmental adaptation coefficient to obtain a corrected risk score value; and determining the immediate intervention requirement by combining the angular velocity data and the environmental adaptation coefficient through signal fusion processing.

[0012] Preferably, the alarm and response module, if an immediate intervention requirement is confirmed, triggers the system alarm device and generates a protection instruction sequence from the assessed risk score to obtain the final intervention response. This includes acquiring real-time image data and angular velocity data from the monitoring equipment, denoising and standardizing the real-time image data through image preprocessing to obtain a first image dataset and a first angular velocity dataset; for the first image dataset, using an edge detection algorithm to identify edge features, and combining the behavioral pattern parameters of the first angular velocity dataset to obtain a tripping risk assessment value; based on the tripping risk assessment value and the environmental adaptability coefficient, using a weighted average method to fuse the risk score value; if the risk score value exceeds a preset threshold, an immediate intervention requirement is determined; when an immediate intervention requirement is determined, signal fusion processing is used to integrate the time series dependency and the first image dataset to generate a protection instruction sequence, triggering the system alarm device to obtain the final intervention response.

[0013] Preferably, the database update module extracts signal fluctuation patterns and occlusion impact assessments from the final intervention response, and updates the individual difference database to optimize the historical action pattern matching adjustment process. This process includes acquiring real-time environmental data and individual behavior data from monitoring equipment, denoising and standardizing the real-time environmental data and individual behavior data through data preprocessing to obtain a first environmental dataset and a first behavior dataset; for the first environmental dataset, an edge detection algorithm is used to identify occlusion features, and combined with the time series dependence of the first behavior dataset, a signal fluctuation pattern is obtained; if the signal fluctuation pattern does not match the preset behavior anomaly threshold, a clustering algorithm is used to extract features from the first behavior dataset and update the individual difference database; based on the updated individual difference database, a weighted average method is used to fuse environmental adaptation parameters and historical action patterns to generate a dynamic matching adjustment instruction, resulting in the final intervention response.

[0014] Preferably, the continuous monitoring module performs complex environmental dynamic change processing on subsequent multi-source signal sequences through an updated individual difference database, and integrates potential imbalance index judgment and terrain unevenness adaptation coefficient to continuously monitor the walking balance of the elderly. This includes acquiring multi-source signal sequences from sensor data acquisition devices, denoising and standardizing the multi-source signal sequences through real-time data processing to obtain a first signal dataset; for the first signal dataset, using time series analysis to extract environmental dynamic features, and combining it with the terrain unevenness adaptation coefficient to obtain an environmental dynamic feature set; if the imbalance index in the environmental dynamic feature set exceeds a preset threshold, then using a feature extraction algorithm to perform in-depth analysis on the first signal dataset to obtain an updated individual difference database; based on the updated individual difference database, using a weighted average method to integrate environmental adaptation parameters and time series analysis results to generate dynamic matching adjustment instructions, thereby obtaining walking balance monitoring results.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This fall prevention and early warning system, based on the assessment of the elderly's intrinsic abilities, solves the problem of how to achieve personalized, real-time fall risk assessment and immediate intervention in business scenarios through multimodal data fusion. This invention collects acceleration, angular velocity, and camera image data, fuses them to form an initial multi-source signal sequence, extracts a preliminary risk feature set, uses convolutional neural networks to mine spatial features, and combines recurrent neural networks to capture subtle fluctuation patterns in the time series to determine potential imbalance indicators. When the imbalance indicator exceeds a threshold, this invention retrieves historical movement patterns from an individual difference database, matches and adjusts them to generate a personalized risk score, and combines real-time image analysis of environmental interference factors such as furniture obstruction or uneven terrain to correct the score and determine the need for immediate intervention. Once intervention is triggered, this invention generates a protective instruction sequence, integrates time series dependencies and image analysis, accurately addresses tripping risks, and updates the database to optimize subsequent monitoring. This invention achieves continuous balance monitoring and personalized intervention in dynamic environments, effectively reducing the risk of falls among the elderly. Attached Figure Description

[0016] Figure 1 This is a connection diagram of the fall prevention early warning system based on the assessment of the elderly's intrinsic abilities, as described in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1As shown, this invention provides a technical solution: a fall prevention and early warning system based on the assessment of the elderly's intrinsic abilities, comprising a data acquisition module that collects acceleration and angular velocity data of elderly people walking near the edge of a carpet in a narrow corridor, as well as camera image data, through multimodal sensors, and fuses the data to form an initial multi-source signal sequence to obtain a preliminary risk feature set; a feature extraction module that uses a convolutional neural network to extract spatial features based on the preliminary risk feature set, and separates dynamic stage change patterns from acceleration and angular velocity data to determine a stage separation sequence; an imbalance judgment module that uses a recurrent neural network to process the time series dependency of the stage separation sequence, captures the fluctuation pattern of the signal in the environment, and judges potential imbalance indicators; and a personalized assessment module that, if the potential imbalance indicator exceeds a preset threshold, retrieves historical movement patterns from an individual difference database, matches and adjusts the current stage separation sequence to obtain a personalized assessment. The system comprises the following modules: Personalized Risk Scoring; Environmental Analysis Module, which combines personalized risk scores with real-time image data captured by cameras to analyze environmental interference factors such as furniture obstruction or uneven terrain to determine the environmental adaptation coefficient; Intervention Judgment Module, which corrects the personalized risk score based on the environmental adaptation coefficient and integrates angular velocity data from multi-source signal sequences to determine immediate intervention needs; Alarm and Response Module, which triggers the system alarm device if immediate intervention needs are confirmed and generates a protective instruction sequence from the judged risk score to obtain the final intervention response; Database Update Module, which extracts signal fluctuation patterns and obstruction impact assessments from the final intervention response and updates the individual difference database to optimize the historical movement pattern matching and adjustment process; and Continuous Monitoring Module, which processes subsequent multi-source signal sequences for complex environmental dynamic changes through the updated individual difference database and integrates potential imbalance index judgments and terrain unevenness adaptation coefficients to continuously monitor the elderly's walking balance.

[0019] This system, based on a deep learning network architecture, collects dynamic behavioral data of elderly individuals in confined spaces such as carpet edges using multimodal sensors (e.g., triaxial accelerometers, gyroscopes, and high-definition cameras), forming an initial multi-source signal sequence that integrates spatial and temporal features. A convolutional neural network (CNN) is used to automatically extract stability-related spatial features from this multi-source data and, combined with acceleration and angular velocity signals, constructs a phase separation sequence reflecting the dynamic phase divisions of an individual's walking process. A recurrent neural network (RNN) further models the phase separation sequence over time, identifying subtle signs of imbalance and forming potential imbalance indicators. The system's built-in individual difference database retrieves historical behavior templates to perform high-dimensional matching and adjustment between the current stage's actions and the individual's history, improving the accuracy and personalization of risk scoring. Subsequently, the system combines images captured by the camera to perform structured analysis of the real-time environment, identifying potential interference factors (e.g., furniture position, carpet edge curvature, terrain elevation changes, etc.) and quantifying them into an environmental adaptation coefficient. This coefficient is used to correct the personalized score, enhancing the predictive ability for immediate fall risk. If the risk exceeds the intervention threshold, the system activates an alarm mechanism and generates a sequence of protective instructions matching the elderly person's behavioral characteristics. It can also coordinate with relevant devices to implement intervention measures (such as issuing audible alerts, waking caregivers, and activating fall prevention devices). Finally, the system inputs the signal fluctuation characteristics and occlusion assessment information involved in this intervention response into the database to update the individual difference model. Through this closed-loop structure, the system can achieve more sensitive dynamic change recognition and adaptive adjustment in complex environments during subsequent monitoring.

[0020] This implementation significantly improves the accuracy of fall risk identification during elderly individuals' walking. It not only captures individual imbalance trends through dynamic multi-source signal analysis but also integrates visual information to enhance the understanding of environmental complexity, exhibiting extremely high real-time performance and personalized adaptability. The combination of convolutional neural networks and recurrent neural networks makes spatial-temporal feature modeling more accurate, especially demonstrating excellent performance in dynamic stage identification and potential imbalance indicator extraction. The personalized risk scoring system effectively solves the problem of decreased prediction accuracy in traditional models when faced with individual differences. The introduction of environmental adaptation factors makes the scoring results more closely match actual usage scenarios, enhancing the overall practical value of the early warning system. Furthermore, through the synergistic effect of the alarm response and database update modules, the system can achieve self-evolution and continuous optimization, possessing advantages such as strong long-term operational stability and a low false alarm rate.

[0021] The data acquisition module collects acceleration and angular velocity data of elderly people walking near the edge of carpet in narrow corridors, as well as camera image data, using multimodal sensors. The data is fused to form an initial multi-source signal sequence, resulting in a preliminary risk feature set. This includes collecting acceleration, angular velocity, and camera image data during the elderly person's walking process using an inertial measurement unit and camera equipment to obtain the original multimodal dataset; synchronizing the acceleration, angular velocity, and camera image data using a timestamp alignment method, and fusing the multimodal data using a Kalman filter algorithm to obtain the initial signal sequence; extracting peak features from the acceleration data, rate of change features from the angular velocity data, and edge contour features from the camera images based on the initial signal sequence; identifying potential risk features if the feature values ​​exceed a preset threshold, thus obtaining the preliminary risk feature set; and clustering the preliminary risk feature set using a K-means algorithm to determine the risk characteristics during the elderly person's walking process, resulting in a classified risk feature set.

[0022] In one possible implementation, the data acquisition module specifically includes an inertial measurement unit (IMU) and a camera device. The IMU is installed at the elderly person's waist level, and the camera is fixed to the ceiling or corner of the elderly person's activity area to capture the complete walking path of the elderly person as they pass the edge of the carpet. After the system starts, the IMU synchronously acquires three-axis acceleration and three-axis angular velocity data at a frequency of 100 samples per second, and the camera device captures images of the elderly person at a rate of 30 frames per second. All acquired sensor data and image data are accompanied by millisecond-level timestamps. This timestamp uses the system's internal unified clock, controlled by the main control processing unit, to ensure that all types of data are accurately aligned in the time dimension.

[0023] For the acquired raw multimodal dataset, time synchronization processing is first performed. Specifically, using the timestamp of the image frame as the main axis, the acceleration and angular velocity data closest to the time of each frame are selected. Linear interpolation is then used to calculate the interpolated data based on the ratio of the image timestamp to the timestamps of two consecutive acceleration or angular velocity frames, thus ensuring a one-to-one correspondence between acceleration, angular velocity, and image data. After alignment, multimodal data fusion is performed using the Kalman filter algorithm. In the Kalman filter algorithm, the initial state is set to the first set of measured values ​​for acceleration and angular velocity. The state transition equation is the weighted predicted value from the previous moment. The measurement noise covariance is determined based on the error statistics obtained during the sampling stabilization period, with the acceleration noise variance set to 0.1 and the angular velocity noise variance set to 0.2. During the filtering process, at each moment, the difference between the predicted value from the previous step and the current measured value, along with the Kalman gain, is used for weighted updates, outputting the estimated fused value, thus obtaining a continuous fused initial signal sequence containing acceleration, angular velocity, and image index.

[0024] Next, feature extraction is performed. First, for acceleration data, a sliding window length of 1 second is set, meaning each window consists of 100 data points with a step size of 50 data points. The absolute values ​​of the triaxial accelerations are calculated for each window, and the maximum value is taken as the peak acceleration feature for that window. Then, the same sliding window processing is performed on the angular velocity data, calculating the maximum value of the triaxial angular velocity change within each time interval as the angular velocity rate of change feature. For image data, the Canny edge detection algorithm with a fixed threshold of 50 is used to extract image edge contours, and the total length of the edge lines is counted as the image contour feature. The system sets the threshold for judging the peak acceleration feature to 2.0 meters per square second, meaning that if the peak acceleration in a certain window exceeds this threshold, it indicates an abnormal change in the elderly person's gait. The threshold for the angular velocity rate of change feature is set to 1.5 radians per second to identify cases of violent twisting or imbalance. The image edge feature is set to have a contour length that increases by more than 20% compared to the baseline edge length of the image under unobstructed conditions, used to determine whether new obstructions or terrain changes have occurred in the environment. When any of the three types of features exceeds its corresponding threshold, that time period is marked as a potential risk segment, and the extracted feature values ​​form a preliminary risk feature set.

[0025] After obtaining the initial risk feature set, the system performs risk classification. First, the peak acceleration, rate of change of angular velocity, and image contour length features are normalized to between 0 and 1 using a maximum-minimum standardization method. The maximum and minimum values ​​are derived from statistical results of data collected over a period of continuous monitoring during the initial acquisition phase. After normalization, K-means clustering is used for clustering. Initially, the number of cluster centers is set to 3, representing low, medium, and high risk levels. At the start of the algorithm, three initial feature sets are randomly selected as initial cluster centers. Then, for each feature data set, based on its Euclidean distance to the current cluster center, the feature is assigned to the category corresponding to the nearest cluster center. The mean of all data points in that category is then recalculated as the new cluster center. This process is repeated until the cluster centers no longer change, indicating algorithm convergence. Finally, all initial risk features are assigned a risk level label, forming the classified risk feature set.

[0026] The feature extraction module uses a convolutional neural network to extract spatial features based on a preliminary risk feature set. It separates dynamic stage change patterns from acceleration and angular velocity data to determine the stage separation sequence. This includes obtaining acceleration and angular velocity data from multimodal data, performing convolution operations on the acceleration and angular velocity data using a convolutional neural network to obtain a spatial feature set; using principal component analysis to reduce the dimensionality of the spatial features based on the spatial feature set to obtain a dimensionality-reduced feature vector set; and using a dynamic time warping algorithm to compare the temporal differences between the feature vectors to determine the dynamic stage change patterns. If the temporal differences of the dynamic stage change patterns exceed a preset threshold, a clustering algorithm is used to group the change patterns to obtain the stage separation sequence.

[0027] In one possible implementation, the feature extraction module extracts acceleration and angular velocity data from the multimodal data that has undergone preliminary risk feature labeling. The data consists of triaxial data obtained by an inertial measurement unit at a sampling frequency of 100 times per second, meaning each 1-second time interval contains 600 data points across three axes of acceleration and angular velocity, arranged in chronological order to form a data matrix, with each row corresponding to one time interval. First, this data matrix is ​​input into a convolutional neural network for spatial feature extraction. The convolutional neural network model consists of two one-dimensional convolutional layers and one max-pooling layer. The first layer has 32 kernels, a kernel length of 3, and a stride of 1; the second layer has 64 kernels, a kernel length of 3, and a stride of 1. The convolutional layers have no edge padding. The first convolutional layer operates on the acceleration data, and the second convolutional layer operates on the angular velocity data, extracting local temporal variation patterns respectively. Then, max-pooling is used to compress the feature map to half its original length to enhance feature stability and anti-interference capabilities. Finally, the outputs of the two pooling methods are concatenated along the time dimension to form a spatial feature set. Each spatial feature set is a fixed-length vector with a length of 128, corresponding to the dynamic feature description fused within 1 second.

[0028] Subsequently, principal component analysis (PCA) is performed on the aforementioned spatial feature set to reduce feature dimensionality and minimize redundant information. The PCA process includes the following steps: first, the covariance matrix of the spatial feature set is calculated, representing the correlation between features of each dimension; then, eigenvalue decomposition is performed on the covariance matrix to obtain all eigenvalues ​​and their corresponding eigenvectors; the top principal components are selected based on the eigenvalue size, with the cumulative explained variance required to exceed 90%. In this implementation, based on extensive sample validation, retaining the first 10 principal components satisfies an information retention rate of over 90%. Therefore, the system projects each set of original 128-dimensional spatial feature vectors onto the directions of the first 10 principal components, resulting in a dimensionality-reduced feature vector set of length 10, ensuring the information representation capability after dimensionality reduction while significantly reducing the computational burden.

[0029] The system performs a dynamic time warping algorithm on the dimensionality-reduced feature vector set to compare the temporal structure differences between adjacent time periods. Specifically, it selects feature vectors from two consecutive time periods and constructs a distance matrix, where each element represents the absolute difference between the principal component values ​​in the two time periods. Then, it searches this matrix for the path with the minimum cumulative cost, i.e., the optimal alignment path. The cumulative cost of this path is the dynamic time warping distance for that pair of features. The system presets a temporal difference threshold of 15, which is derived from pre-collected data samples of normal walking among elderly people and represents the maximum acceptable range of change in consecutive time periods under normal gait conditions. If the dynamic time warping distance between a pair of features exceeds 15, it is determined that there is a dynamic structural abrupt change between the current time period and the previous one, and this is marked as a stage transition point.

[0030] Finally, clustering is performed on all time-period features identified as changing, using the K-means clustering algorithm. The initial number of cluster centers is set to 3, corresponding to the three typical dynamic stages: the initial stage, the stable walking stage, and the final stage. The clustering algorithm executes as follows: first, three feature vectors are randomly selected as initial cluster centers; then, all feature vectors are assigned to the nearest centers based on their Euclidean distance from the three centers; next, the center vector for each cluster is recalculated as the new center; this process is repeated until the centers remain stable, completing cluster convergence. Ultimately, the system divides the entire data into several segments based on the clustering results, each segment corresponding to a dynamic stage, and arranges them in chronological order to form a stage separation sequence.

[0031] The imbalance judgment module uses a recurrent neural network to process time series dependencies in the stage-separated sequences, capturing signal fluctuation patterns in the environment to determine potential imbalance indicators. This includes acquiring time series data from multimodal data streams, using a recurrent neural network to perform time-series modeling on the time series data to obtain a dynamic dependency feature set; segmenting the time series data using a sliding window method to obtain local fluctuation patterns of the signal; if the amplitude of a local fluctuation pattern exceeds a preset threshold, performing a frequency domain transformation on the local fluctuation pattern using a fast Fourier transform to determine the signal frequency distribution in the environment; and grouping the fluctuation patterns according to the signal frequency distribution using a clustering algorithm to determine potential imbalance indicators.

[0032] In one possible implementation, the imbalance judgment module reads the time period index from the stage separation sequence and extracts the acceleration and angular velocity data corresponding to each stage time period as the input time series. The acquisition frequency is 100 times per second, and the data includes three-axis acceleration and three-axis angular velocity, with each data point accompanied by a precise timestamp. First, the system organizes the time series data into a multi-dimensional vector sequence based on time and inputs it into a recurrent neural network. This recurrent neural network adopts a long short-term memory network structure, containing one hidden layer with 64 hidden units. Each input is one second of data, i.e., 100 vectors, each with a dimension of 6, containing three-axis acceleration and three-axis angular velocity. The network captures the temporal dependencies between data through a gating mechanism and outputs the dynamic dependency features corresponding to each second. The output features are 64-dimensional vectors representing the comprehensive dynamic state of the signal within that time period. Multiple time period output vectors form a dynamic dependency feature set.

[0033] Next, the system uses a sliding window method to segment and analyze the dynamic dependency feature set. The sliding window length is 2 seconds (200 data points), and the step size is 1 second (100 data points are slid across the window each time), performing local analysis on the signal over a continuous time period. For the 64-dimensional feature vector set within each window, the system first calculates the difference between the maximum and minimum values ​​for each dimension within that segment, and then averages the differences across all 64 dimensions to obtain the local fluctuation amplitude for that window. The system presets a local fluctuation amplitude threshold of 1.5, which is obtained statistically from normal gait data extracted from a sample of 200 elderly individuals aged 60 and above. This means that the average local amplitude under stable walking conditions does not exceed 1.5. If the detected value exceeds this value, it indicates significant abnormal fluctuation in the current time period, and the system marks it as a high-fluctuation time period.

[0034] For all high-fluctuation time periods, the system continues to perform frequency domain analysis to identify abnormal frequency structures in the signal. The dynamic dependency features within each marked high-fluctuation window are arranged in chronological order to form a one-dimensional signal sequence. Fast Fourier Transform (FFT) processing is then performed, and the system selects the frequency and energy corresponding to the dominant frequency in the signal as frequency domain features. The dominant frequency is the frequency component with the highest energy. The system determines whether the dominant frequency exceeds 3 Hz. If the dominant frequency exceeds 3 Hz and the spectral energy is concentrated in the high-frequency region, it is considered an abnormal frequency pattern. The dominant frequency threshold of 3 Hz is set based on the upper limit statistical value of gait frequency extracted from the aforementioned elderly samples. Exceeding this threshold indicates excessively rapid gait changes or body swaying. If the spectrum meets the above conditions, the time position of that window is recorded, and this frequency domain feature is used as input to enter the clustering judgment stage.

[0035] In the clustering judgment phase, the system inputs all frequency domain features that meet the conditions into the K-means clustering model for classification. The initial number of cluster centers is 2, representing normal fluctuation patterns and imbalanced fluctuation patterns, respectively. The input features used for clustering include the dominant frequency value and the proportion of high-frequency energy. The system first performs standardization processing, scaling all data to between 0 and 1. Then, using Euclidean distance as the distance metric, each data point is assigned to the nearest center, and the center value is updated. This process is repeated iteratively until the clustering results stabilize and no longer change. The time periods corresponding to the data points ultimately classified as imbalanced fluctuations are identified as potential imbalance indicators, which the system uses as the input for the next module's evaluation.

[0036] The personalized assessment module, if the potential imbalance index exceeds a preset threshold, retrieves historical action patterns from the individual difference database and matches and adjusts the current stage separation sequence to obtain a personalized risk score. This includes retrieving historical action patterns from the individual difference database, using data stream processing methods to segment and extract behavioral pattern features from the time series data; if the deviation between the behavioral pattern features and the stage separation sequence exceeds a preset threshold, a dynamic matching algorithm is used to compare the stage separation sequence with the historical action patterns to obtain pattern adjustment parameters; based on the pattern adjustment parameters, the stage separation sequence is corrected, and a weighted average method is used to fuse the behavioral pattern features to generate an adjusted sequence feature set; the adjusted sequence feature set is scored using a risk assessment model, and the k-nearest neighbor algorithm is used to quantify the potential imbalance index to determine the personalized risk score.

[0037] In one possible implementation, upon receiving a potential imbalance indicator and detecting that the indicator value exceeds the system-defined risk trigger threshold of 50, the personalized assessment module immediately retrieves the elderly user's historical movement pattern data from the individual difference database. Each piece of historical movement pattern data in the database includes a complete phase separation sequence, and each phase contains specific time periods, cadence, stride length, angular velocity changes, body tilt angles, and other parameter features. The system first uses a data stream processing method to segment the current phase separation sequence into 2-second time slices with a step size of 1 second, ensuring overlap between continuous time segments and improving the ability to capture behavioral changes. Similarly, the historical movement pattern data is segmented in the same way, and corresponding behavioral pattern features are extracted for each time slice, including: average cadence (steps per minute); average stride length (meters); average angular velocity (radians per second); average body tilt angle (degrees); and transition time between adjacent phases (seconds). These data form a five-dimensional behavioral feature vector, representing the characteristics of each movement segment.

[0038] Subsequently, the system compares the differences between each segment of the current stage's separated sequence and each segment of the historical action pattern, using Euclidean distance to calculate the vector difference. If the average of all differences exceeds a set deviation threshold of 3.0, it is considered that there is a significant difference between the current behavioral state and the historical pattern. This threshold of 3.0 is the maximum reasonable average deviation derived from comparing normal and abnormal states based on 2000 sets of historical behavioral data of elderly people, effectively distinguishing between stable states and signs of imbalance. After determining that a deviation exists, the system executes a dynamic matching process, calling a dynamic time warping algorithm to perform time alignment calculations between the current stage's separated sequence and the historical action pattern sequence. During the matching process, each segment of the current sequence searches for the best matching segment in the historical data. The algorithm outputs the matching path and the time scaling factor and feature adjustment coefficient for each segment, called the pattern adjustment parameters, for subsequent correction.

[0039] Next, the system performs correction operations on the stage separation sequence based on the obtained pattern adjustment parameters. For stages with a time scaling factor greater than 1, the system proportionally extends the stage's time length, i.e., interpolates and fills in the intermediate features; for stages with a time scaling factor less than 1, the system proportionally compresses the stage's time length, i.e., deletes redundant segments to align with historical references. Simultaneously, a weighted average fusion process is performed on each segment's five-dimensional feature vector, with a fusion ratio of 0.7 and 0.3, meaning the current observation accounts for 70% and historical pattern features account for 30%, primarily preserving the individual's current state while incorporating stable individual references. The resulting adjusted sequence feature set will be used as the evaluation object and input into the risk scoring model.

[0040] The risk scoring model employs the k-nearest neighbor algorithm, where k is set to 5. This means the system will find the five samples from the historical sample set that are most similar to the current feature. Similarity calculation still uses Euclidean distance, sorting the feature vectors by distance and selecting the five samples with the smallest distances. For the historical imbalance scores of these five samples, the system performs a weighted average calculation, where the smaller the distance between the sample and the current data, the higher its score weight. The weight allocation ratio is generated based on the normalized inverse proportion method to ensure a total weight of 1. The final output score is a real number between 0 and 100, representing the personalized imbalance risk level of the elderly person in their current behavioral state; a higher score indicates a greater imbalance risk.

[0041] The environmental analysis module combines personalized risk scores with real-time image data captured by cameras to analyze environmental interference factors such as furniture obstruction or uneven terrain to determine the environmental adaptability coefficient. This includes acquiring the original image from the real-time image data captured by the camera, removing noise and blur through image preprocessing to obtain a first image; for the first image, using feature extraction methods to identify image features of obstruction and uneven terrain to obtain an environmental interference feature set; if the deviation between the environmental interference feature set and a preset threshold exceeds a standard, a dynamic matching algorithm is used to compare the environmental interference feature set with behavioral pattern features to obtain environmental adjustment parameters; based on the environmental adjustment parameters, a weighted average method is used to fuse the personalized risk score and the environmental interference feature set to obtain the environmental adaptability coefficient.

[0042] In one possible implementation, the environmental analysis module acquires image data from the camera in real time, obtaining 30 frames per second, each with a resolution of 1280 x 720 pixels. The system first performs image preprocessing on each frame, including noise removal and image sharpening. Noise removal uses a 3x3 median filter, selecting the median value within each pixel region to replace the center pixel value, thereby removing high-frequency interference points. Image sharpening uses the Laplacian operator to perform second-order derivative edge enhancement, and after performing a full-image convolution operation, it improves contour sharpness, removes image blur, and finally outputs the first image.

[0043] Next, the system performs image feature extraction on the first image to identify occlusion and uneven terrain. First, edge detection is performed using the Canny algorithm, with a low threshold of 50 and a high threshold of 150 to determine edge responses and extract strong contours from the image. Then, contour closure analysis is performed on the edge contours, measuring the boundary closure and area of ​​each contour. If a completely closed boundary is detected with a vertical edge length exceeding 30% of the image height (240 pixels) and a top height exceeding 140 pixels from the ground, it is identified as furniture occlusion. The system accumulates the number of such occluded contours and calculates the pixel length of each contour boundary; the total length is used as a reference for the occlusion level.

[0044] To determine uneven terrain, the system extracts pixel blocks located in the lower quarter of the image as the ground region. A grayscale image is generated using a grayscale conversion method. Then, the average grayscale difference between adjacent blocks is calculated, using 10x10 pixel units as the unit. If the grayscale difference in a continuous region exceeds 30 and the corresponding boundary drop is greater than 5 cm (calculated based on the pixel-to-actual distance mapping calibration value in the ground region, where 100 pixels equals 20 cm), it is identified as a terrain abrupt change region. The number of abrupt change blocks and their total area are then counted. Finally, an environmental interference feature set is formed, including four indicators: the number of occlusions, the total length of the occlusion outline, the number of terrain abrupt changes, and the percentage of abrupt change area.

[0045] The system calculates the difference between the current image environment interference feature set and the environmental baseline model. The environmental baseline model is derived from the image analysis results collected in a typical safe residential environment. The following thresholds are set as judgment criteria: the number of obstructions exceeds 3, the total length of the obstruction outline exceeds 300 pixels, the number of terrain abrupt changes exceeds 2, and the total area of ​​the abrupt changes exceeds 15% of the ground area. If three or more of the above four items exceed the threshold, the system determines that the current environment has an interference risk.

[0046] Once an environment is identified as disruptive, the system uses a dynamic matching algorithm to compare the current environment feature set with environmental samples in the behavioral pattern feature library. It performs a minimum Euclidean distance search to find the most similar environment template and extracts its corresponding environment adjustment parameters. These parameters include the disruption level (1 to 5) and a risk factor. The disruption level is determined by the historical intervention success rate in the matched template; for example, level 1 represents mild disruption, and level 5 represents high-risk disruption. The risk factor is a decimal between 0 and 1, representing the percentage of the current disruptive environment's impact on the elderly person's behavioral stability. The system sets the risk factor based on training samples: 0.1 for mild disruption, 0.3 for moderate disruption, 0.6 for severe disruption, and 0.8 or higher for extremely severe disruption.

[0047] The system integrates the current personalized risk score with environmental disturbance risk factors, using a weighted average method. The personalized score has a weight of 0.6, and the environmental risk factor has a weight of 0.4. For example, if the personalized score is 70 and the environmental factor is 0.6 multiplied by 100, which equals 60, then the environmental adaptation coefficient is calculated as 70 multiplied by 0.6 plus 60 multiplied by 0.4, resulting in a final value of 66. The environmental adaptation coefficient ranges from 0 to 100 and represents the elderly person's overall ability to adapt to the environment in the current state. A lower value indicates a more dangerous environment and the need for immediate intervention.

[0048] The intervention judgment module corrects personalized risk scores based on environmental adaptability coefficients and determines immediate intervention needs by fusing angular velocity data from multi-source signal sequences. This includes acquiring angular velocity data from the multi-source signal sequences, denoising and standardizing the angular velocity data through real-time data processing to obtain a first angular velocity dataset; extracting motion features from the first angular velocity dataset using a behavior pattern recognition method and dynamically matching them with a preset behavior pattern library to obtain behavior pattern parameters; fusing personalized risk scores using a weighted average method based on the behavior pattern parameters and the environmental adaptability coefficients to obtain a corrected risk score value; and determining immediate intervention needs by combining signal fusion processing with the angular velocity data and the environmental adaptability coefficients if the corrected risk score value exceeds a preset intervention threshold.

[0049] In one possible implementation, the intervention judgment module first extracts angular velocity data from a multi-source signal sequence. This data consists of three-axis angular velocity signals acquired by an inertial measurement unit, corresponding to left-right rotation, forward-backward tilt, and up-down twisting of the body, with a sampling frequency of 100 times per second, generating a total of 300 raw data points per second across the three axes. The system first performs noise reduction processing on this angular velocity data using a moving average filtering algorithm. The sliding window size is set to 5 data points; that is, for each data point, the average of the values ​​of the two points before and after it, plus the current value, is taken as the current smoothed value, resulting in the noise-reduced angular velocity data. Then, the processed angular velocity data is standardized. The standardization method is maximum-minimum normalization, with the maximum value set to 4 radians per second and the minimum value set to -4 radians per second. These values ​​are derived from the extreme value range of statistical analysis of the movement states of normal elderly individuals during the system training phase. Through standardization, all angular velocity values ​​are mapped to the range of 0 to 1, forming a unified first angular velocity dataset, which facilitates subsequent model processing.

[0050] Subsequently, the system divides the first angular velocity dataset into segments every 2 seconds, with each segment containing 200 data points, and extracts motion features for each segment. The extracted features include: maximum angular velocity (the largest absolute value among the three axial angular velocities in the current segment); minimum angular velocity (the smallest absolute value); average angular velocity (the arithmetic mean of the angular velocity values ​​at all points in the segment); rate of change of angular velocity (the maximum amplitude of the first-order difference of the angular velocity in the current segment); and fluctuation frequency (obtained by multiplying the number of times the angular velocity crosses zero in the current segment by the unit time). Each segment corresponds to a 5-dimensional feature vector, which is input into the behavior pattern recognition module. The system matches this feature vector with standard feature templates stored in a behavior pattern library containing 6 categories of standard behavioral features: standing, normal walking, fast walking, turning, lateral movement, and swaying. The matching process uses a dynamic time warping algorithm, determining the similarity by calculating the total cost of the shortest alignment path between the current feature and the template. The system selects the template corresponding to the path with the minimum cost as the recognition result of the current segment, and outputs the behavior category and its corresponding behavior sensitivity level. The level range is from 1 to 5, where level 1 represents stable behavior, such as standing, and level 5 represents actions that are very easy to induce imbalance, such as swaying or turning around suddenly.

[0051] Next, the system integrates and corrects the personalized risk score with behavioral pattern parameters and environmental adaptation coefficients using a weighted average method. In the correction process, the system first multiplies the behavioral sensitivity level by the environmental adaptation coefficient and then normalizes the result. This normalized product is used to calculate the behavioral and environmental weights of the risk score. For example, if the current behavioral sensitivity level is 4 and the environmental adaptation coefficient is 60, the product is 240. The system compares this value with the historical maximum product value of 300 and normalizes it to 0.8, setting the behavioral weight to 0.6 and the environmental weight to 0.4. If the current personalized risk score is 70 and the environmental adaptation coefficient is 60, the corrected risk score is calculated as 70 multiplied by 0.6 plus 60 multiplied by 0.4, which equals 66. The corrected risk score represents the level of imbalance risk at the current moment after comprehensively considering individual status and environmental factors.

[0052] The system then determines whether the revised risk score exceeds the intervention threshold, which is set at 65. This threshold is derived from analyzing the risk score distribution of 1000 elderly individuals before their actual falls and serves as the critical value for a high-risk alarm. If the revised score is greater than or equal to 65, the system enters the immediate intervention determination phase, further integrating the current angular velocity change rate with the environmental adaptation coefficient. The angular velocity change rate is the maximum change in angular velocity per second; if this value exceeds 2 radians per second, it indicates a violent movement. The system standardizes the angular velocity change rate and multiplies it by the normalized value of the environmental adaptation coefficient to obtain the fused signal strength. If this strength value exceeds 0.8, it indicates a violent movement and low environmental adaptability. The system will immediately generate an intervention request signal, triggering an alarm or executing physical intervention mechanisms such as voice prompts or vibration feedback to ensure timely risk control before a fall occurs.

[0053] The alarm and response module, if an immediate intervention requirement is confirmed, triggers the system alarm device. It generates a protective instruction sequence from the assessed risk score to obtain the final intervention response. This includes acquiring real-time image data and angular velocity data from monitoring equipment; performing image preprocessing to denoise and standardize the real-time image data to obtain a first image dataset and a first angular velocity dataset; using an edge detection algorithm to identify edge features in the first image dataset, and combining this with behavioral pattern parameters from the first angular velocity dataset to obtain a tripping risk assessment value; fusing the risk score value and environmental adaptability coefficient using a weighted average method; if the risk score value exceeds a preset threshold, an immediate intervention requirement is determined; when an immediate intervention requirement is determined, signal fusion processing is used to integrate the time series dependency and the first image dataset to generate a protective instruction sequence, triggering the system alarm device to obtain the final intervention response.

[0054] In one possible implementation, upon receiving an immediate intervention request from the intervention judgment module, the alarm and response module immediately retrieves the current image data and angular velocity data from the monitoring system. Image data is acquired via a camera device at a sampling frequency of 30 frames per second, with each frame having a resolution of 1280 x 720 pixels. Angular velocity data is acquired via an inertial measurement unit at 100 samples per second, covering three axes: left-right rotation, forward-backward tilt, and up-down twisting. The system first performs image preprocessing on the image data, including two steps: first, median filtering for noise reduction using a 3 x 3 pixel sliding window to remove isolated bright spots and interfering pixels; second, grayscale normalization, converting the image to a single-channel grayscale image and linearly mapping all pixel values ​​to the 0-1 range. During normalization, the maximum and minimum pixel values ​​in the current image frame are used as normalization boundaries. The processed image data constitutes the first image dataset. The angular velocity data is processed by a moving average filtering algorithm with a sliding window size of 5 data points to remove high-frequency jitter noise from the signal. Then, it is subjected to maximum and minimum normalization processing, with the maximum angular velocity set to 4 radians per second and the minimum to -4 radians per second. The normalization result generates the first angular velocity dataset.

[0055] Next, the system performs edge detection on the first image dataset using the Canny algorithm, setting a low threshold of 50 and a high threshold of 150. The algorithm performs edge enhancement processing between these two thresholds, extracting structural feature lines representing object boundaries from the image. The system performs statistical analysis on the extracted edge features, including the number of edge segments, total length, average density, and spatial distribution of edge regions. If continuous closed edge contours appear in the ground area of ​​the image, the system determines that there is furniture occlusion or abrupt terrain change. The system combines these image features with behavioral pattern parameters from the first angular velocity dataset for the corresponding time period. These behavioral pattern parameters include maximum angular velocity, rate of change of angular velocity, standard deviation of angular velocity, and the currently identified action category (e.g., standing, walking, turning, swaying). By combining the edge density in the image features with the behavioral fluctuation index in the angular velocity, the system assigns contribution values ​​to the image and angular velocity factors with weight ratios of 0.6 and 0.4 respectively, calculating the tripping risk assessment value for the current time period. The assessment value ranges from 0 to 100, with higher values ​​indicating a greater probability of tripping. The image edge density threshold is set to the presence of closed edge regions with a length exceeding 300 pixels in the lower third of the image, and the angular velocity change rate threshold is set to a change of more than 2 radians per second. Both of these thresholds are derived from statistical results of real-world data recorded before elderly people fell.

[0056] The system integrates the tripping risk assessment value with the calculated environmental adaptation coefficient using a weighted average method. The tripping risk value accounts for 0.7 of the weight, and the environmental adaptation coefficient accounts for 0.3. The weighting ratio is set based on the optimal combination of false positive and false negative rates determined during the system validation phase. If the final risk score obtained after integration exceeds the intervention trigger threshold of 65, it is considered that an immediate intervention requirement has been confirmed. This intervention threshold of 65 was determined by statistical analysis of images and angular velocity scores within 10 seconds prior to actual falls in 500 elderly individuals, effectively identifying over 90% of genuine high-risk events.

[0057] Upon confirming the need for immediate intervention, the system initiates a signal fusion processing flow, extracting the angular velocity time series and image frame sequence within 3 seconds before and after the current time point, forming a time series dependency matrix and an image spatial feature matrix, respectively. The time series dependency matrix contains angular velocity data and behavioral pattern label sequences over a continuous 6 seconds, while the image matrix contains edge distribution, occlusion area, and obstacle location markers. The system aligns the two types of data using time labels, analyzes the spatial overlap between the current action trajectory and risk areas in the image, and generates a comprehensive protective instruction sequence based on the location of the overlapping area, the direction of angular velocity change, and the action type. The instruction sequence includes the current action state, risk direction, priority intervention method, and target, such as "The current action is a rapid turn; there is an obstacle to the right front. It is recommended to issue a voice prompt and vibration alert." This instruction is encoded in a structured format to ensure that it can be directly parsed by the alarm execution module.

[0058] Finally, the alarm and response module automatically selects the appropriate alarm device and executes the response based on the type of protection instruction. If the instruction type is a voice prompt, the system calls the speaker device to play a fixed instruction audio, such as "Watch your step, there is an obstacle ahead"; if the instruction type is a physical intervention, the vibration function of wearable devices such as smart bracelets or canes is activated; if the instruction level is emergency, a high-decibel alarm is activated simultaneously, and an alarm message is sent to the caregiver's mobile phone. A structured alarm message is sent through a dedicated communication interface, including the user ID, timestamp, risk level, action type, and intervention suggestions.

[0059] The database update module extracts signal fluctuation patterns and occlusion impact assessments from the final intervention response, and updates the individual difference database to optimize the historical action pattern matching adjustment process. This includes acquiring real-time environmental data and individual behavior data from monitoring equipment; preprocessing the real-time environmental data and individual behavior data to remove noise and standardize them, resulting in a first environmental dataset and a first behavior dataset; for the first environmental dataset, an edge detection algorithm is used to identify occlusion features, and combined with the time-series dependence of the first behavior dataset, signal fluctuation patterns are obtained; if the signal fluctuation pattern does not match a preset behavior anomaly threshold, a clustering algorithm is used to extract features from the first behavior dataset and update the individual difference database; based on the updated individual difference database, a weighted average method is used to fuse environmental adaptation parameters and historical action patterns to generate dynamic matching adjustment instructions, resulting in the final intervention response.

[0060] In one possible implementation, the database update module is activated immediately after each system intervention response. It extracts real-time environmental and individual behavior data from the monitoring equipment by calling a 6-second continuous data segment stored before and after the current intervention response. The environmental data source is image frame data at a frame rate of 30 frames per second and an image resolution of 1280 x 720 pixels; the behavior data source is a triaxial angular velocity sequence with a sampling frequency of 100 times per second. The system first preprocesses both types of data. For the environmental image data, median filtering is used for noise reduction, with a window size of 3 x 3 pixels. Each pixel is replaced by its neighborhood median to remove high-frequency noise points. Then, grayscale processing is performed, converting the RGB color image to a single-channel image and performing linear normalization to map all pixel values ​​to between 0 and 1, generating the first environmental dataset. The angular velocity data were denoised using a moving average algorithm with a window size of 5 data points. Each point's value was replaced by the average of itself and the two points before and after it. Then, a maximum-minimum normalization process was performed, with the maximum value set to 4 radians per second and the minimum value set to -4 radians per second. These two values ​​were set based on the extreme values ​​of 1,000 elderly people in normal and abnormal states collected by the system. The normalization results formed the first row of the dataset.

[0061] Subsequently, the system performs edge detection on the first environmental dataset, using the Canny algorithm to process the images. Edge extraction for each frame is set with a low threshold of 50 and a high threshold of 150. The detected edges are assessed for closure and spatial location. If a continuous closed edge with a length greater than 300 pixels appears in the lower third of the image, it is considered an effective occlusion area. Time-series dependency analysis is performed on the first behavioral dataset within the same time period. The system extracts five features from the angular velocity data: maximum angular velocity, average angular velocity, rate of change of angular velocity, standard deviation, and fluctuation period. The 6-second time period is divided into 1-second sub-windows, and the above five features are calculated for each window to form a time-series feature vector. The system compares this feature sequence with preset behavioral anomaly criteria. These criteria include: a maximum rate of change of angular velocity exceeding 2 radians per second within any time window, or a fluctuation period less than 0.5 seconds. This threshold is based on real fall sample data and represents significant imbalance behavior characteristics. If the current behavioral data does not meet the above abnormality criteria, but the system triggers an intervention response, it is considered that this response is not driven by the established abnormality criteria, and there is a new behavioral pattern, so the database needs to be updated.

[0062] After confirming that the data deviated from the preset standard, the system used the K-means clustering algorithm to perform feature clustering analysis on the first behavior dataset. The system input the aforementioned 5-dimensional feature vectors per second into the clustering model, with the initial number of cluster centers set to 5, representing five basic action types: normal standing, slow walking, fast walking, turning, and swaying. Before input, each feature dimension was normalized to 0 to 1, with the normalization range determined by the maximum and minimum values ​​statistically analyzed from the training set. During the clustering process, Euclidean distance was used as a metric to assign each feature vector to the nearest center, and the center values ​​were iteratively updated until all cluster centers stabilized. After clustering, the system determined whether the data segment formed a new category of behavioral features different from those in the existing database. If it was a new category, the cluster center was used as a new behavior template and added to the individual difference database along with the corresponding occlusion environment image label. The recorded content included the new action name, feature mean vector, occlusion image edge parameters, time period, and system intervention result.

[0063] Subsequently, the system optimizes the historical action pattern matching mechanism based on the newly added template. The newly identified template is used as the primary reference and fused with historical action patterns using a weighted average. The weights are set to 0.7 for the current template and 0.3 for the historical template. This weight ratio is derived from the training results that minimize matching error during model optimization, prioritizing the retention of current high-confidence data while avoiding the complete discarding of old templates. The fused template is then input into the matching module as the new version, used in subsequent stages of sequence separation to determine individual behavioral states and risk scores, thereby improving the system's sensitivity to abnormal behavior and its ability to personalize adjustments.

[0064] Finally, the system generates a dynamic matching adjustment instruction based on the updated template. The instruction includes the current user ID, update timestamp, new template feature vector, matching priority label, and matching correction coefficient. These are stored uniformly and used as the object called by the behavior matching module in the next round.

[0065] The continuous monitoring module processes subsequent multi-source signal sequences for complex environmental dynamic changes using an updated individual difference database. It integrates potential imbalance indicators and terrain unevenness adaptation coefficients to continuously monitor the walking balance of the elderly. This includes acquiring multi-source signal sequences from sensor data acquisition devices, denoising and standardizing these sequences through real-time data processing to obtain a first signal dataset; extracting environmental dynamic features from this first signal dataset using time series analysis, and combining this with the terrain unevenness adaptation coefficient to obtain an environmental dynamic feature set; if the imbalance indicators in the environmental dynamic feature set exceed a preset threshold, a feature extraction algorithm is used to perform in-depth analysis on the first signal dataset to obtain an updated individual difference database; based on this updated database, a weighted average method is used to integrate environmental adaptation parameters and time series analysis results to generate dynamic matching adjustment instructions, resulting in walking balance monitoring results.

[0066] In one possible implementation, the continuous monitoring module acquires multi-source signal sequences in real time from wearable sensor devices for the elderly and a fixed camera via continuous acquisition. Acceleration and angular velocity signals are provided by a waist-mounted inertial measurement unit, providing triaxial acceleration and triaxial angular velocity signals respectively, with a sampling frequency of 100 times per second. Image data is acquired by a camera installed above the corridor, at a frame rate of 30 frames per second and an image resolution of 1280 x 720 pixels. The system first performs data cleaning and preprocessing operations on the acquired multi-source signal sequences. Acceleration and angular velocity data were denoised using a moving average filtering algorithm with a filtering window length of 5 data points, meaning the value at each moment was the average of the two points before and after it, plus the current value. Image data underwent a 3x3 median filter to remove isolated high and low grayscale pixel noise and normalize the image grayscale values ​​to between 0 and 1. The normalization formula used the minimum and maximum pixel values ​​in the current frame as upper and lower bounds for linear transformation. The normalization interval for acceleration data was set to -20 to +20 meters per square second, and the normalization interval for angular velocity data was set to -4 to +4 radians per second. These values ​​were based on the maximum boundary values ​​collected by the system during the walking process of elderly people. All processed data were integrated into a unified format first signal dataset.

[0067] After acquiring the first signal dataset, the system performs time series analysis to extract dynamic change features in the current environment. The system divides the signal sequence into windows with a 3-second sliding window and a 1-second step size, and extracts the following five environmental dynamic features from the signal data within each window: 1) Acceleration rate of change, calculated as the difference between the maximum and minimum acceleration values ​​in the current window divided by the time length; 2) Angular velocity rate of change, calculated as the difference between the maximum and minimum angular velocity values ​​divided by the time length; 3) Acceleration standard deviation, representing the signal fluctuation amplitude within the window; 4) Dominant frequency, extracted by performing a Fast Fourier Transform on the acceleration signal within the window to extract the most concentrated energy frequency in the spectrum; and 5) Acceleration direction change frequency, counting the number of times the acceleration vector direction deflects by more than 15 degrees per unit time. Subsequently, the system uses the terrain unevenness adaptation coefficient extracted from the image as an auxiliary reference. The adaptation coefficient is calculated by analyzing the gray-level gradient changes in the ground area of ​​the image and comparing it structurally with a standard terrain template. If the gray-level change area of ​​the ground in the image is greater than 200 pixels and exhibits an irregular continuous slope, it is considered uneven, and the adaptation coefficient is set to a range of 0 to 1. A standard level ground is set to 1, and severely uneven ground is set to 0. The above five dynamic features and the terrain unevenness adaptation coefficient constitute an environmental dynamic feature set, which the system uses to determine whether there are any abnormal trends.

[0068] The system performs scoring on potential imbalance indicators within the dynamic environmental feature set. The system sets a comprehensive imbalance judgment threshold of 0.6. The scoring method is as follows: an acceleration change rate exceeding 2 meters per square second is scored as 0.2; an angular velocity change rate exceeding 1.5 radians per second is scored as 0.2; a standard deviation greater than 1.0 is scored as 0.1; a dominant frequency exceeding 3 Hz is scored as 0.1; a direction change frequency greater than 3 times per second is scored as 0.2; and a terrain adaptability coefficient less than 0.4 is scored as 0.2. If the cumulative score exceeds 0.6, it is judged as a potential imbalance risk state. This threshold was derived from analyzing walking data from over 2000 elderly individuals during the system training phase, ensuring that the system maintains a recognition accuracy rate of over 95% while reducing the false alarm rate.

[0069] Upon identifying an imbalance trend, the system immediately performs deep feature extraction on the first signal dataset within the current time window. It uses continuous wavelet transform to extract high-frequency and low-frequency energy distribution features from the acceleration signal, and principal component analysis to extract the principal direction changes in the time series features to form principal component feature vectors. Simultaneously, it performs edge distribution continuity statistics on the image frame sequence to identify the trend of occlusion edge changes, and comprehensively generates a set of multimodal deep feature vectors as a joint expression of the current behavioral state and environmental state of the elderly.

[0070] The newly generated deep feature vectors are compared with historical feature templates already stored in the individual difference database. If the Euclidean distance between the current feature and the closest historical template is greater than 0.3, the system identifies it as a new behavioral pattern and writes the feature vector as a new sample into the individual difference database, along with labels such as the current occurrence time, environmental adaptation coefficient, and trigger intervention result. To improve the system's predictive consistency, an update mechanism is used to perform a weighted fusion of the current template and the closest historical template in the database. The fusion weight is 0.7 for the current template and 0.3 for the historical template. The weight ratio is derived from the actual prediction accuracy evaluation in the validation set, ensuring that the new template becomes the primary reference in a short period of time.

[0071] The system uses the merged template to construct new dynamic matching and adjustment instructions. These instructions include a timestamp, user ID, current feature vector, risk level, terrain coefficient, behavioral state classification label, and suggested intervention strategy. This instruction, as output, is sent in real-time by the monitoring module to the evaluation and response module for the next round of judgment. Based on the instruction result, the system outputs a walking balance monitoring status value, ranging from 0 to 100. A value below 40 indicates a warning state, between 40 and 70 indicates a moderate fluctuation state, and above 70 indicates a stable state.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly, characterized in that, include: The data acquisition module uses multimodal sensors to collect acceleration and angular velocity data of elderly people walking near the edge of carpet in narrow corridors, as well as camera image data. The data is fused to form an initial multi-source signal sequence to obtain a preliminary risk feature set. The feature extraction module uses a convolutional neural network to extract spatial features based on the preliminary risk feature set, and separates the dynamic stage change patterns from acceleration data and angular velocity data to determine the stage separation sequence; The imbalance judgment module uses a recurrent neural network to process the time series dependencies of the stage-separated sequences and capture the fluctuation patterns of the signals in the environment to judge potential imbalance indicators. The personalized assessment module retrieves historical action patterns from the individual difference database if potential imbalance indicators exceed preset thresholds, and then matches and adjusts the current stage separation sequence to obtain a personalized risk score. The environmental analysis module combines personalized risk scores with real-time image data captured by cameras to analyze environmental interference factors such as furniture obstruction or uneven terrain to determine the environmental adaptability coefficient. The intervention assessment module corrects personalized risk scores based on environmental adaptability coefficients and integrates angular velocity data from multi-source signal sequences to determine immediate intervention needs. The alarm and response module, if the need for immediate intervention is confirmed, triggers the system alarm device and generates a sequence of protection instructions from the risk score after assessment to obtain the final intervention response; The database update module extracts signal fluctuation patterns and occlusion impact assessments from the final intervention response and updates the individual difference database to optimize the historical action pattern matching and adjustment process. The continuous monitoring module processes subsequent multi-source signal sequences for complex environmental dynamic changes using an updated individual difference database, and integrates potential imbalance indicators and terrain unevenness adaptation coefficients to continuously monitor the walking balance of the elderly.

2. The fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The data acquisition module collects acceleration and angular velocity data of elderly people walking near the edge of carpet in narrow corridors, as well as camera image data, through multimodal sensors. The data is then fused to form an initial multi-source signal sequence, resulting in a preliminary risk feature set, including: By using an inertial measurement unit and camera equipment, acceleration data, angular velocity data, and camera image data of elderly people during their walking process are collected to obtain the original multimodal dataset; For the original multimodal dataset, a timestamp alignment method is used to synchronize acceleration data, angular velocity data, and camera image data. The multimodal data is then fused using a Kalman filter algorithm to obtain an initial signal sequence. Based on the initial signal sequence, peak features of acceleration data, rate of change features of angular velocity data, and edge contour features of camera images are extracted. If the feature values ​​exceed the preset threshold, they are judged as potential risk features, and a preliminary risk feature set is obtained. The initial risk feature set is clustered using the K-means algorithm to determine the risk characteristics of elderly people during their walking process, thus obtaining a classified risk feature set.

3. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The feature extraction module uses a convolutional neural network to extract spatial features based on the preliminary risk feature set, and separates the dynamic stage change patterns from acceleration and angular velocity data to determine the stage separation sequence, including: Acceleration and angular velocity data are obtained from multimodal data, and convolutional neural networks are used to perform convolution operations on the acceleration and angular velocity data to obtain a spatial feature set. Based on the spatial feature set, principal component analysis is used to reduce the dimensionality of the spatial features, resulting in a dimensionality-reduced feature vector set. For the dimensionality-reduced feature vector set, a dynamic time warping algorithm is used to compare the temporal differences between the feature vectors and determine the dynamic stage change pattern. If the temporal differences of the dynamic stage change patterns exceed a preset threshold, the change patterns are grouped using a clustering algorithm to obtain stage separation sequences.

4. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The imbalance judgment module uses a recurrent neural network to process the time series dependencies of the stage-separated sequences and captures the fluctuation patterns of signals in the environment to judge potential imbalance indicators, including: Time series data is obtained from multimodal data streams, and a recurrent neural network is used to perform time series modeling on the time series data to obtain a dynamic dependency feature set; For dynamically dependent feature sets, the time series data is segmented using a sliding window method to obtain the local fluctuation patterns of the signal; If the amplitude of a local fluctuation pattern exceeds a preset threshold, the local fluctuation pattern is converted to the frequency domain by a fast Fourier transform to determine the signal frequency distribution in the environment. Based on the signal frequency distribution, a clustering algorithm is used to group the fluctuation patterns and determine potential imbalance indicators.

5. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The personalized assessment module, if the potential imbalance indicator exceeds a preset threshold, retrieves historical action patterns from the individual difference database and matches and adjusts the current stage separation sequence to obtain a personalized risk score, including: Historical action patterns are obtained from an individual difference database, and the time series data is segmented and extracted using a data stream processing method to obtain behavioral pattern features. If the deviation between the behavioral pattern features and the stage separation sequence exceeds a preset threshold, the stage separation sequence is compared with the historical action pattern through a dynamic matching algorithm to obtain the pattern adjustment parameters. Based on the pattern adjustment parameters, the stage separation sequence is corrected, and the behavioral pattern features are fused using a weighted average method to generate an adjusted sequence feature set. The adjusted sequence feature set is scored using a risk assessment model, and the k-nearest neighbor algorithm is used to quantify potential imbalance indicators to determine personalized risk scores.

6. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The environmental analysis module combines personalized risk scoring with real-time image data captured by cameras to analyze environmental interference factors such as furniture obstruction or uneven terrain to determine the environmental adaptability coefficient, including: The original image is obtained from the real-time image data captured by the camera, and noise and blur are removed through image preprocessing to obtain the first image; For the first image, a feature extraction method is used to identify image features of occlusion and uneven terrain to obtain an environmental interference feature set; If the deviation between the environmental interference feature set and the preset threshold exceeds the standard, the environmental adjustment parameters are obtained by comparing the environmental interference feature set with the behavioral pattern features through a dynamic matching algorithm. Based on the environmental adjustment parameters, the personalized risk score and the environmental disturbance feature set are fused using a weighted average method to obtain the environmental adaptability coefficient.

7. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The intervention judgment module corrects the personalized risk score based on the environmental adaptability coefficient and integrates angular velocity data from multi-source signal sequences to determine immediate intervention needs, including: Angular velocity data is obtained from a multi-source signal sequence, and the angular velocity data is denoised and standardized through real-time data processing to obtain the first angular velocity dataset; For the first angular velocity dataset, motion features are extracted using a behavior pattern recognition method and dynamically matched with a pre-set behavior pattern library to obtain behavior pattern parameters; Based on the behavioral pattern parameters and the environmental adaptation coefficient, a weighted average method is used to fuse the personalized risk scores to obtain the corrected risk score value. If the corrected risk score exceeds the preset intervention threshold, the need for immediate intervention is determined by combining the angular velocity data and the environmental adaptation coefficient through signal fusion processing.

8. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The alarm and response module, if an immediate intervention requirement is confirmed, triggers the system alarm device and generates a sequence of protective instructions from the assessed risk score to obtain the final intervention response, including: Real-time image data and angular velocity data are acquired from monitoring equipment. The real-time image data is denoised and standardized through image preprocessing to obtain the first image dataset and the first angular velocity dataset. For the first image dataset, an edge detection algorithm is used to identify edge features, and combined with the behavior pattern parameters of the first angular velocity dataset, a tripping risk assessment value is obtained; Based on the tripping risk assessment value and the environmental adaptability coefficient, a weighted average method is used to integrate the risk score value. If the risk score value exceeds the preset threshold, an immediate intervention is required. When the need for immediate intervention is determined, signal fusion processing is used to integrate time-series dependencies and the first image dataset to generate a sequence of protective instructions, trigger the system alarm device, and obtain the final intervention response.

9. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The database update module extracts signal fluctuation patterns and occlusion impact assessments from the final intervention response, and updates the individual difference database to optimize the historical action pattern matching adjustment process, including: Real-time environmental data and individual behavior data are acquired from monitoring equipment. Data preprocessing is used to denoise and standardize the real-time environmental data and individual behavior data to obtain the first environmental dataset and the first behavior dataset. For the first environmental dataset, an edge detection algorithm is used to identify occlusion features, and combined with the time series dependency of the first row dataset, the signal fluctuation pattern is obtained; If the signal fluctuation pattern does not match the preset abnormal behavior threshold, then the first behavior dataset is subjected to feature extraction by clustering algorithm, and the individual difference database is updated. Based on the updated individual difference database, a weighted average method is used to integrate environmental adaptation parameters and historical action patterns to generate dynamic matching adjustment instructions, resulting in the final intervention response.

10. A fall prevention early warning system based on the assessment of the intrinsic abilities of the elderly according to claim 1, characterized in that: The continuous monitoring module performs complex environmental dynamic changes processing on subsequent multi-source signal sequences through an updated individual difference database, and integrates potential imbalance index judgment and terrain unevenness adaptation coefficient to continuously monitor the walking balance of the elderly, including: The first signal dataset is obtained by acquiring multi-source signal sequences from sensor data acquisition devices and denoising and standardizing the multi-source signal sequences through real-time data processing. For the first signal dataset, time series analysis is used to extract environmental dynamic features, and combined with the terrain unevenness adaptation coefficient, an environmental dynamic feature set is obtained; If the imbalance index in the environmental dynamic feature set exceeds the preset threshold, the first signal dataset is analyzed in depth using a feature extraction algorithm to obtain an updated individual difference database. Based on the updated individual difference database, a weighted average method is used to integrate environmental adaptation parameters and time series analysis results to generate dynamic matching adjustment instructions and obtain walking balance monitoring results.

Images