Old people outdoor falling risk early warning processing method and system based on multi-source data

Abstract

The invention, which relates to the technical field of the computer, discloses a multi-source-data-based outdoor falling risk early warning processing method and system for the elderly, comprising the following steps: collecting and fusing multi-source data; performing multi-dimensional feature extraction and correlation analysis; carrying out risk identification based on business rules; small sample machine learning identification; performing relation network auxiliary analysis; according to the method, Internet of Things devices such as video monitoring and smart bracelets are integrated, three types of recognition algorithms of a rule model, a classification model and a relation network model are creatively fused, model cooperation is realized through a weighted voting mechanism, and high recognition accuracy is still kept under the condition of small samples. A differential response process is automatically triggered based on a risk level, flexible expansion of system function modules is supported by adopting a micro-service architecture design, a whole-process closed loop of old people outdoor fall risk early warning disposal is finally realized, and the intelligent level of grassroots social governance is remarkably improved.

Description

A method and system for early warning and handling of outdoor fall risks among the elderly based on multi-source data Technical Field

[0001] This invention relates to the field of computer technology, specifically to a method and system for early warning and handling of outdoor fall risks among the elderly based on multi-source data. Background Technology

[0002] With the accelerating aging process, outdoor safety protection for the elderly in communities has become a crucial issue in grassroots governance. Falls among the elderly are characterized by their sudden occurrence, serious consequences, and high demands for timely rescue. Establishing an effective risk warning and response mechanism is imperative. Currently, there are two major technical bottlenecks in the field of fall detection technology for the elderly: First, the scarcity of sample data. In real-world scenarios, the frequency of falls among the elderly is low, with an average of fewer than 30 cases per year in a single community. Existing supervised learning methods heavily rely on large-scale labeled data, and the model's generalization ability significantly decreases when samples are insufficient. Second, there is a lack of a closed-loop business process. Existing technologies are mostly limited to algorithmic optimization and fail to effectively integrate with community grid management systems and emergency response mechanisms in medical institutions. After warning information is issued, there is a lack of standardized on-site verification, multi-party collaboration, and resource allocation procedures, making it difficult to translate technological achievements into practical protective effectiveness. Currently, a fall monitoring and multi-terminal linkage emergency response system for the elderly (application number CN202511164582.1) is disclosed. This method uses an electromagnetic wave radar sensor array to detect blind spots in low-light conditions. The non-intrusive collection of human dynamic information in the environment, the confirmation of the fall command output by the behavior perception through the voice interaction verification judgment module, and the setting of priority queue management mechanism to realize the multi-protocol linkage scheduling of heterogeneous terminals, effectively protect user privacy and reduce false alarm rate; however, the existing technology has the following three shortcomings: (1) The problem of single data collection source is prominent. Most systems rely on a single data source such as video surveillance or wearable devices, resulting in the lack of dimensions in the constructed user profile, which cannot fully reflect the correlation between the physiological state, behavior pattern and environmental factors of the caring population, especially when identifying complex risk scenarios; (2) The risk identification model architecture is single. Traditional systems either only use rule engines or simply rely on machine learning algorithms. In small sample scenarios, there is a common overfitting phenomenon. The sensitivity to sparse abnormal events is insufficient. In the complex environment of the actual community, the generalization performance is significantly reduced, and the false alarm rate remains high; (3) The existing technology generally lacks a complete business closed-loop design. Risk warning and response are disconnected. There is no effective linkage mechanism between warning events and community resource scheduling, resulting in a long response delay from risk discovery to on-site handling. Summary of the Invention

[0003] This invention provides a method and system for early warning and handling of outdoor falls among the elderly based on multi-source data. It can effectively solve the problems mentioned in the background technology, such as the sparse occurrence of falls among the elderly, the reliance of existing models on a large amount of labeled data, the difficulty in maintaining robustness when the sample is insufficient, and the lack of a business closed loop that links with community management, grid personnel, and medical institutions, resulting in low response efficiency after early warning and the inability to form a complete chain.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for early warning and handling of outdoor falls among the elderly based on multi-source data, realizing closed-loop management of the entire process from risk perception to response, including the following steps: Step 1, multi-source data collection and fusion; Step 2, multi-dimensional feature extraction and correlation analysis; Step 3, risk identification based on business rules; Step 4, small-sample machine learning identification; Step 5, relationship network-assisted analysis; Step 6, multi-model fusion and coordinated handling.

[0005] According to the above technical solution, in step one, multi-source data related to the elderly is first collected, including community public area surveillance video data, smart bracelet data worn by the elderly, and basic information data of the elderly. The surveillance video data is obtained through high-definition cameras deployed in the community public area. The smart bracelet data includes physiological indicators such as blood pressure, blood oxygen, and heart rate. The basic information data is obtained through community registration or questionnaires. The data fusion process uses multi-dimensional identifiers with device number, location code, and timestamp for spatiotemporal alignment. In specific implementation, the time of various data sources is first synchronized, then the video data and bracelet data are geographically associated through spatial location coding, and finally the basic information is associated with the dynamic monitoring data through a unique personnel identifier.

[0006] According to the above technical solution, step two involves multi-dimensional feature extraction of the fused data, including behavioral features, posture features, physiological features, and spatiotemporal features. Regarding behavioral features, parameters such as walking speed, gait stability, and activity frequency are extracted in real-time using computer vision technologies based on YOLO and DeepSORT. Regarding posture features, a deep learning posture estimation algorithm based on OpenPose combined with an LSTM network is used to extract temporal indicators of body tilt angle, joint flexion, and center of gravity changes. Regarding physiological features, wavelet transform is used to extract the coefficient of variation of heart rate from wristband sensor data, peak acceleration is obtained through a peak detection algorithm, and blood oxygen saturation is analyzed using a sliding window statistical method. In terms of spatiotemporal characteristics, parameters for degree change trends are extracted using a DBSCAN-based trajectory mining algorithm to identify features such as activity trajectory density, dwell time distribution, and regional preferences. During feature correlation analysis, a Transformer-based spatiotemporal correlation algorithm is employed to establish a mapping relationship between individual behavioral patterns and risk levels. By analyzing normal behavioral patterns in historical data, an exponentially weighted moving average method is used to construct an individual dynamic behavioral baseline. Real-time monitoring of the deviation between the data and the baseline serves as a crucial basis for risk assessment. Simultaneously, K-means clustering analysis is used to discover group behavioral patterns, and combined with the isolated forest algorithm, a reference benchmark is provided for anomaly detection, forming a dual-perspective anomaly assessment system encompassing both individual and group aspects.

[0007] According to the above technical solution, in step three, based on medical knowledge and real-world scenario data analysis, a multi-level business rule base for elderly fall recognition is constructed. The rule design needs to fully consider the physical characteristics and physiological response characteristics of the fall event, and adopts sliding window real-time analysis technology, specifically including body motion feature rules, posture persistence feature rules, and physiological response rules. The body motion feature rules are based on OpenPose posture estimation combined with Kalman filtering, the posture persistence feature rules are based on computer vision target detection technology, and the physiological response rules are based on dynamic baseline adaptive algorithm.

[0008] According to the above technical solution, step four, which employs few-shot learning technology to construct a risk identification model, specifically includes a prototype network construction stage, a query set processing stage, a model optimization stage, and a model evaluation stage. The prototype network construction stage uses a metric learning method, extracting deep features of samples through a convolutional neural network, clustering similar samples and separating dissimilar samples in the embedding space. The support set construction stage selects representative samples from a limited set of labeled samples, covering typical cases of different body types and different fall postures. The query set processing stage calculates the distance between new samples and prototypes of each category using distance metrics, employing Euclidean distance to measure the similarity between samples. For each query sample, the distance between it and prototypes of each category in the support set is calculated, and its category is determined based on the nearest neighbor principle. The model optimization stage employs a meta-learning strategy, simulating the few-shot learning process in real-world scenarios through multiple training tasks. The model evaluation stage uses five-fold cross-validation.

[0009] According to the above technical solution, step five involves constructing individual behavioral profiles and group relationship networks by analyzing the daily activity patterns and social relationship networks of the elderly. Specifically, this includes activity pattern network analysis, social relationship network analysis, spatiotemporal co-occurrence network analysis, and risk propagation network analysis. Activity pattern network analysis uses activity areas and time periods as nodes, with transfer frequency and dwell time as edge weights to construct an individual activity pattern network. By analyzing network characteristics, it identifies the regularity and stability of behavioral patterns and detects abnormal behavioral patterns. Social relationship network analysis uses elderly individuals as nodes, with interaction frequency and intimacy as edge weights to construct a social relationship network. By analyzing topological characteristics, it assesses the individual's social activity and social support strength. Spatiotemporal co-occurrence network analysis combines three-dimensional information to construct a spatiotemporal co-occurrence network. By analyzing the co-occurrence patterns of individuals in the spatiotemporal dimension, it discovers potential social interaction groups and behavioral imitation phenomena. Risk propagation network analysis uses risk factors and individuals as nodes, with influence intensity and propagation path as edges to construct a risk propagation network.

[0010] According to the above technical solution, in step six, the final risk level is determined by a weighted fusion strategy based on the identification results of integrated business rules, machine learning, and relationship networks. The fusion weights are dynamically adjusted according to the historical performance of each model. Specifically, the weight of the business rule model is 30%, the weight of the few-shot learning model is 40%, and the weight of the relationship network model is 30%. Based on the fused risk score, the risk level is divided into four levels: high risk, medium risk, low risk, and normal state. A three-level response system is established in conjunction with a linkage and handling mechanism to respond to high risk, medium risk, and low risk respectively, and all handling results are fed back to the system in real time.

[0011] A risk warning and response system for outdoor falls among the elderly based on multi-source data is proposed. This system integrates multi-source data fusion, intelligent identification, and collaborative response to achieve risk warning and response. Specifically, it includes a multi-source data acquisition and fusion processing module, a multi-dimensional feature extraction and correlation analysis module, a risk identification module based on business rules, a small-sample machine learning identification module, a relationship network-assisted analysis module, and a multi-model fusion and linkage response module.

[0012] According to the above technical solution, the multi-source data acquisition and fusion processing module collects surveillance video data from public areas of the community and data from wristbands worn by the elderly. It uses multidimensional identifiers to perform spatiotemporal alignment and fusion processing on the original data, constructing a complete intelligent device collaborative data system. Basic information about the elderly is collected through community surveys or questionnaires. The multidimensional feature extraction and correlation analysis module uses computer vision and deep learning technologies to extract multidimensional features of the elderly from the fused video data. Through spatiotemporal correlation analysis, it establishes a correspondence between individual behavioral patterns and risk levels. The risk identification module based on business rules analyzes the behavioral characteristics of elderly people falling, combines medical knowledge and real-world scenario data to formulate a business rule base, and achieves preliminary risk identification by setting thresholds for key indicators.

[0013] According to the above technical solution, the few-sample machine learning recognition module addresses the sparsity problem of fall events by employing few-sample learning technology. It constructs a risk recognition model based on a prototype network and achieves rapid model adaptation and optimization under limited labeled sample conditions through support sets and query sets. The relationship network auxiliary analysis module analyzes the daily activity patterns and social relationship networks of the elderly to construct individual behavioral profiles. It then uses graph neural network technology to identify abnormal behavior patterns, providing auxiliary decision-making basis for risk warning. The multi-model fusion and coordinated response module integrates the recognition results of business rules, machine learning, and relationship networks, determines the final risk level through a weighted fusion strategy, and automatically triggers a three-level coordinated response mechanism involving community grid workers, guardians, and medical institutions based on the warning level.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: 1. It integrates multi-source data fusion, intelligent identification, and collaborative handling. By integrating IoT devices such as video surveillance and smart bracelets, it establishes a multi-dimensional data collection network covering physical space, physiological indicators, and environmental parameters, forming a complete user profile system. It innovatively integrates three types of identification algorithms: rule model, classification model, and relationship network model. Through a weighted voting mechanism, it achieves model collaboration and maintains a high identification accuracy rate even under small sample conditions. It establishes a tiered emergency response plan library and automatically triggers differentiated response processes based on risk levels, enabling rapid linkage between guardians, community grid workers, medical institutions, and emergency departments. It adopts a microservice architecture design to support the flexible expansion of system functional modules, providing a technical foundation for integration into the smart city governance system. Ultimately, it achieves a closed-loop process for early warning and handling of outdoor falls among the elderly, significantly improving the level of intelligence in grassroots social governance.

[0015] 2. By fusing heterogeneous data from video surveillance and smart bracelets, and utilizing device IDs, location codes, and timestamps, spatiotemporal alignment and iterative correlation of video, sensor data, and basic information are achieved. This constructs a comprehensive risk perception capability that goes beyond simple fall detection. A hierarchical rule base is established by combining posture, physiological, and spatiotemporal features, and key threshold feature sets such as body tilt and acceleration abrupt changes are defined. By integrating rule models, few-shot learning, and relational network models, weighted voting ensures high accuracy while solving the overfitting problem in complex scenarios. Multi-source data fusion provides comprehensive monitoring data support, multi-dimensional feature extraction enables accurate behavioral analysis, and multi-model fusion ensures the reliability of risk assessment. Finally, a coordinated response mechanism forms a business closed loop, effectively improving the safety level of elderly people's outdoor activities.

[0016] 3. By integrating data from IoT devices such as video surveillance and smart bracelets, a perception system covering physical and physiological dimensions has been established. The data collection dimensions have expanded from single fall detection to multiple scenarios such as health monitoring and behavior analysis. It integrates three types of algorithms: rule models, classification models, and relationship network models. Through a weighted voting mechanism, the model collaboration is achieved, maintaining an accuracy rate of over 85% even under small sample conditions. This effectively solves the overfitting problem of single models in complex community scenarios. A complete hierarchical emergency response plan library and cross-departmental linkage mechanism, as well as a closed-loop mechanism, have been established to achieve a closed loop from human risk perception to response feedback. The microservice architecture design supports modular expansion of sensing devices, analysis models, and response resources. Standardized interfaces with government platforms and medical systems are reserved, providing a better technical foundation for integration into the smart city governance system. In contrast, existing patents are mostly limited to specific fall detection scenarios. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] In the accompanying drawings: Figure 1 is a flowchart of the risk warning and handling method of the present invention; Figure 2 is a flowchart of the multi-source data acquisition and fusion of the present invention; Figure 3 is a schematic diagram of the multi-dimensional feature extraction and correlation analysis of the present invention; Figure 4 is a flowchart of the small sample machine learning recognition of the present invention; Figure 5 is a flowchart of the multi-model fusion and linkage handling of the present invention; Figure 6 is a logical block diagram of the risk warning and handling system of the present invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] Example: As shown in Figures 1-5, this invention provides a technical solution, a method for early warning and handling of outdoor falls among the elderly based on multi-source data. This method achieves closed-loop management of the entire process from risk perception to response, providing reliable technical support for early warning and handling of falls among the elderly and improving the safety level of outdoor activities for the elderly. The method includes the following steps: Step 1, multi-source data collection and fusion; Step 2, multi-dimensional feature extraction and correlation analysis; Step 3, risk identification based on business rules; Step 4, small-sample machine learning identification; Step 5, relationship network-assisted analysis; Step 6, multi-model fusion and coordinated response.

[0021] Based on the above technical solution, the first step involves collecting multi-source data related to the elderly, including surveillance video data from community public areas, data from smart bracelets worn by the elderly, and basic information data about the elderly. Surveillance video data is acquired through high-definition cameras deployed in community public areas to capture the elderly's behavior and movement trajectories. Smart bracelet data includes physiological indicators such as blood pressure, blood oxygen, and heart rate. Basic information data is obtained through community registration, covering static attribute information such as age, history of underlying diseases, emergency contacts, and permanent address. The data fusion process uses multi-dimensional identifiers, including device numbers, location codes, and timestamps, for spatiotemporal alignment. In practice, the time of various data sources is first synchronized to ensure data consistency. Secondly, spatial location codes are used to geographically associate video data with bracelet data. Finally, unique personal identifiers are used to associate basic information with dynamic monitoring data. Through multiple iterative associations, a complete data system for monitoring the outdoor activities of the elderly is constructed.

[0022] Based on the above technical solution, step two involves multi-dimensional feature extraction of the fused data, including behavioral features, posture features, physiological features, and spatiotemporal features. Regarding behavioral features, walking speed, gait stability, and activity frequency parameters are extracted in real-time using computer vision technologies based on YOLO and DeepSORT. For posture features, a deep learning posture estimation algorithm based on OpenPose combined with an LSTM network is used to extract temporal indicators of body tilt angle, joint flexion, and center of gravity changes. Regarding physiological features, wavelet transform is used to extract the coefficient of variation of heart rate from wristband sensor data, peak acceleration is obtained through a peak detection algorithm, and blood oxygen saturation is analyzed using a sliding window statistical method. In terms of spatiotemporal characteristics, the parameters for trend changes are extracted using a DBSCAN-based trajectory mining algorithm to identify features such as activity trajectory density, dwell time distribution, and regional preferences. During feature correlation analysis, a Transformer-based spatiotemporal correlation algorithm is employed to establish a mapping relationship between individual behavioral patterns and risk levels. By analyzing normal behavioral patterns in historical data, an exponentially weighted moving average method is used to construct an individual dynamic behavioral baseline. Real-time monitoring of the deviation between the data and the baseline serves as a crucial basis for risk assessment. Simultaneously, K-means clustering analysis is used to discover group behavioral patterns, and the isolated forest algorithm is combined to provide a reference benchmark for anomaly detection, forming a dual-perspective anomaly assessment system encompassing both individual and group aspects.

[0023] Based on the above technical solution, step three involves constructing a multi-layered business rule base for fall recognition in the elderly, based on medical knowledge and real-world scenario data analysis. Rule design must fully consider the physical and physiological characteristics of fall events, employing a sliding window real-time analysis technique. Specifically, this includes body motion characteristic rules, posture continuity characteristic rules, and physiological response rules. The body motion characteristic rules, based on OpenPose posture estimation combined with Kalman filtering, detect whether the body tilt angle exceeds a 60-degree threshold. This threshold is determined based on ergonomic research and clinical data analysis. Using triaxial accelerometer data, it monitors whether there is a sudden change in vertical acceleration exceeding three times the gravitational acceleration; this indicator reflects the dynamic characteristics of a sudden fall. The posture continuity characteristic rules are based on… The computer vision-based target detection technology detects whether the contact state with the ground lasts for more than 2 seconds. Combined with the geographic location information system, it determines whether the person is in a normal resting area. Time series analysis is used to monitor whether the duration of stillness exceeds 5 minutes. At the same time, the activity area features are correlated and a logistic regression model is used to exclude normal resting situations. The physiological response rules are based on a dynamic baseline adaptive algorithm. It monitors whether the heart rate changes by more than 30% of the baseline value in a short period of time. Combined with the movement status analysis, it analyzes whether there is a stress response. Change point detection technology is used to analyze sudden changes in blood oxygen saturation. A support vector machine classifier is used to assess whether there is physiological dysfunction caused by the fall. All rule outputs are comprehensively judged by a confidence-weighted fusion model to improve the accuracy of risk assessment.

[0024] Based on the above technical solution, step four addresses the problem of sparse fall event samples and limited labeled data by employing few-shot learning techniques to construct a risk identification model. This process includes prototype network construction, query set processing, model optimization, and model evaluation. The prototype network construction stage uses metric learning, extracting deep features from samples via convolutional neural networks. Similar samples are clustered in the embedding space, while dissimilar samples are separated. The support set construction stage selects representative samples from the limited labeled samples, covering typical cases of different body types and fall postures. The query set processing stage calculates the distance between new samples and prototypes of each category using distance metrics. Euclidean distance is used to measure the similarity between samples. For each query sample, the distance between it and prototypes of each category in the support set is calculated, and its category is determined based on the nearest neighbor principle. The model optimization stage employs a meta-learning strategy, simulating the few-shot learning process in real-world scenarios through multiple training tasks. During model training, cross-validation is used to evaluate model performance, and model parameters are adjusted based on the validation results. The model evaluation stage uses five-fold cross-validation to ensure the reliability of the model evaluation results.

[0025] Based on the above technical solution, step five involves constructing individual behavioral profiles and group relationship networks by analyzing the daily activity patterns and social relationship networks of the elderly. Specifically, this includes activity pattern network analysis, social relationship network analysis, spatiotemporal co-occurrence network analysis, and risk transmission network analysis. Activity pattern network analysis uses activity areas and time periods as nodes, with transfer frequency and dwell time as edge weights to construct an individual activity pattern network. By analyzing network characteristics, the regularity and stability of behavioral patterns are identified, and abnormal behavioral patterns are detected. Social relationship network analysis uses individual elderly individuals as nodes, with interaction frequency and intimacy as edge weights. The analysis methods include: constructing social relationship networks to assess an individual's social activity and social support strength by analyzing topological characteristics, such as network centrality and clustering coefficients; constructing spatiotemporal co-occurrence network analysis by combining three-dimensional information, including people, location, and time, and identifying potential social interaction groups and behavioral imitation phenomena by analyzing the co-occurrence patterns of individuals in the spatiotemporal dimensions; and constructing risk propagation network analysis by using risk factors and individuals as nodes and influence intensity and propagation path as edges, and assessing systemic risk and individual vulnerability by simulating the propagation process of risk in the network.

[0026] Based on the above technical solution, in step six, the final risk level is determined by a weighted fusion strategy using the identification results from integrated business rules, machine learning, and relationship networks. The fusion weights are dynamically adjusted based on the historical performance of each model. Specifically, the business rule model has a weight of 30%, focusing on the identification of high-risk emergencies; the few-shot learning model has a weight of 40%, mainly responsible for routine risk identification; and the relationship network model has a weight of 30%, focusing on abnormal behavior detection. Based on the fused risk score, the risk level is divided into four levels: high risk, medium risk, low risk, and normal state. High risk (80-100 points) triggers the emergency response mechanism of medical institutions; medium risk... Risk: 60-79 points, initiate guardian notification and community attention procedures; Low risk: 40-59 points, community grid workers handle on-site; Normal status: 0-39 points, maintain continuous monitoring; A three-tiered response system is established in conjunction with a joint response mechanism, responding to high, medium, and low risks respectively. Specifically, in high-risk situations, the green channel for medical institutions is activated to ensure a response within 5 minutes; in medium-risk situations, guardians and community service centers are notified, and initial contact is completed within 15 minutes; in low-risk situations, community grid workers are dispatched for on-site handling, arriving at the scene within 30 minutes. All handling results are fed back to the system in real time for model optimization and rule updates.

[0027] A complete early warning and response system for outdoor falls among the elderly has been established. It provides comprehensive monitoring data support through multi-source data fusion, achieves accurate behavioral analysis through multi-dimensional feature extraction, ensures the reliability of risk assessment through multi-model fusion identification, and finally forms a business closed loop through a joint response mechanism, effectively improving the safety level of outdoor activities for the elderly.

[0028] As shown in Figure 6, a multi-source data-based outdoor fall risk warning and response system for the elderly is presented. This system integrates multi-source data fusion, intelligent identification, and collaborative response to achieve risk warning and response. Specifically, it includes a multi-source data acquisition and fusion processing module, a multi-dimensional feature extraction and correlation analysis module, a business rule-based risk identification module, a small-sample machine learning identification module, a relationship network-assisted analysis module, and a multi-model fusion and linkage response module.

[0029] Based on the above technical solution, the multi-source data acquisition and fusion processing module collects surveillance video data from public areas of the community and data from wristbands worn by the elderly. It uses multidimensional identifiers (DIDs) to perform spatiotemporal alignment and fusion processing on the raw data. These DIDs include device number, location code, and timestamp, constructing a complete intelligent device collaborative data system. Basic information about the elderly is collected from the community, including age, underlying diseases, family members, and address. The multidimensional feature extraction and correlation analysis module uses computer vision and deep learning technologies to extract multidimensional features of the elderly from the fused video data. These features include behavioral characteristics, movement trajectories, and posture changes. Through spatiotemporal correlation analysis, a correspondence between individual behavioral patterns and risk levels is established. The risk identification module based on business rules analyzes the behavioral characteristics of elderly falls. Combining medical knowledge and real-world scenario data, it develops a business rule base. By setting thresholds for key indicators, it achieves preliminary risk identification. Key indicators include body tilt angle, acceleration, and ground contact state.

[0030] Based on the above technical solutions, the few-sample machine learning recognition module addresses the sparsity problem of fall events by employing few-sample learning techniques. It constructs a risk recognition model based on a prototype network and achieves rapid model adaptation and optimization under limited labeled sample conditions through support sets and query sets. The relationship network auxiliary analysis module analyzes the daily activity patterns and social relationship networks of the elderly to construct individual behavioral profiles. Using graph neural network technology, it identifies abnormal behavior patterns, providing auxiliary decision-making basis for risk warning. The multi-model fusion and coordinated response module integrates the recognition results from business rules, machine learning, and relationship networks. Through a weighted fusion strategy, it determines the final risk level and automatically triggers a three-level coordinated response mechanism involving community grid workers, guardians, and medical institutions based on the warning level, forming a complete business closed loop.

[0031] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning and handling of outdoor fall risks among the elderly based on multi-source data, characterized in that: To achieve closed-loop management of the entire process from risk perception to response, the following steps are included: Step 1, multi-source data collection and fusion; Step 2, multi-dimensional feature extraction and correlation analysis; Step 3, risk identification based on business rules; Step 4, small-sample machine learning identification; Step 5, relationship network-assisted analysis; Step 6, multi-model fusion and coordinated response.

2. The method for early warning and handling of outdoor fall risks among the elderly based on multi-source data according to claim 1, characterized in that: Step one involves collecting multi-source data related to the elderly, including surveillance video data from public areas in the community, data from smart bracelets worn by the elderly, and basic information data of the elderly. The data fusion process uses multidimensional identifiers, including device number, location code, and timestamp, for spatiotemporal alignment. In practice, the time of various data sources is first synchronized, then the video data and wristband data are geographically associated through spatial location coding, and finally, basic information is associated with dynamic monitoring data through unique personnel identifiers.

3. The method for early warning and handling of outdoor fall risks among the elderly based on multi-source data according to claim 2, characterized in that: Step two involves extracting multi-dimensional features from the fused data, including behavioral features, posture features, physiological features, and spatiotemporal features. Simultaneously, K-means clustering analysis is used to discover group behavior patterns, and the isolated forest algorithm is combined to provide a reference benchmark for anomaly detection, forming an anomaly assessment system from both individual and group perspectives.

4. The method for early warning and handling of outdoor fall risks among the elderly based on multi-source data according to claim 1, characterized in that: In step three, based on medical knowledge and real-world scenario data analysis, a multi-level business rule base for elderly fall recognition is constructed. The rule design needs to fully consider the physical and physiological characteristics of the fall event and adopts sliding window real-time analysis technology, specifically including body movement feature rules, posture continuity feature rules, and physiological response rules.

5. A method for early warning and handling of outdoor fall risks among the elderly based on multi-source data, as described in claim 1, characterized in that: Step four involves constructing a risk identification model using few-shot learning techniques. This process includes prototype network construction, query set processing, model optimization, and model evaluation.

6. The method for early warning and handling of outdoor fall risks among the elderly based on multi-source data according to claim 1, characterized in that: Step five involves constructing individual behavioral profiles and group relationship networks by analyzing the daily activity patterns and social relationship networks of the elderly. Specifically, this includes activity pattern network analysis, social relationship network analysis, spatiotemporal co-occurrence network analysis, and risk transmission network analysis.

7. A method for early warning and handling of outdoor fall risks among the elderly based on multi-source data, as described in claim 1, is characterized in that: In step six, the final risk level is determined by a weighted fusion strategy based on the identification results of integrated business rules, machine learning, and relationship networks. The fusion weights are dynamically adjusted according to the historical performance of each model. Specifically, the weight of the business rule model is 30%, the weight of the few-shot learning model is 40%, and the weight of the relationship network model is 30%. Based on the fused risk score, the risk level is divided into four levels: high risk, medium risk, low risk, and normal state. A three-level response system is established in conjunction with a coordinated handling mechanism to respond to high risk, medium risk, and low risk respectively, and all handling results are fed back to the system in real time.

8. A multi-source data-based outdoor fall risk warning and response system for the elderly, used in the response system according to any one of claims 1-7, characterized in that: This system integrates multi-source data fusion, intelligent identification, and collaborative handling to achieve risk early warning and response. Specifically, it includes a multi-source data acquisition and fusion processing module, a multi-dimensional feature extraction and correlation analysis module, a risk identification module based on business rules, a small-sample machine learning identification module, a relationship network-assisted analysis module, and a multi-model fusion and linkage handling module.

9. A multi-source data-based early warning and response system for outdoor fall risks among the elderly, as described in claim 8, is characterized in that: The multi-source data acquisition and fusion processing module collects surveillance video data from public areas of the community and data from wristbands worn by the elderly. It uses multidimensional identifiers to perform spatiotemporal alignment and fusion processing on the raw data to build a complete intelligent device collaborative data system. It collects basic information about the elderly through the community or questionnaires. The multidimensional feature extraction and correlation analysis module uses computer vision and deep learning technologies to extract multidimensional features of the elderly from the fused video data. Through spatiotemporal correlation analysis, it establishes a correspondence between individual behavioral patterns and risk levels. The risk identification module based on business rules analyzes the behavioral characteristics of elderly people falling, combines medical knowledge and real-world scenario data to formulate a business rule base, and achieves preliminary risk identification by setting thresholds for key indicators.

10. A multi-source data-based early warning and response system for outdoor fall risks among the elderly, as described in claim 8, is characterized in that: The few-shot machine learning recognition module addresses the sparsity problem of fall events by employing few-shot learning techniques. It constructs a risk recognition model based on a prototype network and utilizes support sets and query sets to achieve rapid model adaptation and optimization under limited labeled samples. The relationship network auxiliary analysis module analyzes the daily activity patterns and social relationship networks of the elderly to construct individual behavioral profiles. Using graph neural network technology, it identifies abnormal behavior patterns, providing auxiliary decision-making basis for risk warning. The multi-model fusion and coordinated response module integrates business rules, machine learning, and relationship network recognition results, determines the final risk level through a weighted fusion strategy, and automatically triggers a three-tiered coordinated response mechanism involving community grid workers, guardians, and medical institutions based on the warning level.

Images