Methods and systems for health monitoring of the elderly based on environmental perception and contextual interaction

Abstract

This invention discloses a method and system for health monitoring of the elderly based on environmental perception and contextual interaction, belonging to the field of intelligent health monitoring technology. The method utilizes a mobile terminal application, a wearable terminal, and a cloud-based intelligent AI engine working collaboratively to provide two complementary operating modes: a first monitoring mode, which uses abnormal acoustic event detection combined with forward environmental image analysis to achieve accurate identification and timely warning of safety risks such as falls; and a second interactive mode, which, based on periodically or triggered environmental images, performs contextual understanding in the cloud and dynamically generates personalized cognitive training tasks, improving the cognitive activity of the elderly through voice interaction. The system records and integrates safety event and cognitive interaction data to construct a comprehensive user health profile and regularly generates comprehensive health assessment reports. This invention achieves a deep integration of safety monitoring and cognitive training, providing comprehensive, proactive, and intelligent health protection for the elderly.

Description

Technical Field

[0001] This invention belongs to the field of intelligent health monitoring technology, and in particular relates to a method and system for health monitoring of the elderly based on environmental perception and contextual interaction. Background Technology

[0002] The daily safety and health management of the elderly has become an important social issue. Traditional methods of elderly monitoring mainly include regular in-person visits, emergency call buttons, and monitoring schemes based on single sensors. These methods have significant shortcomings: manual methods are costly and have slow response times; emergency buttons rely on the elderly to actively trigger them and cannot be used in cases of sudden illness or loss of consciousness; single sensors have a high false alarm rate and cannot provide contextualized interaction and cognitive support.

[0003] In recent years, the development of smart home and wearable technology has provided new ideas for elderly monitoring. Some solutions monitor the activities of the elderly through devices such as cameras and infrared sensors, but these pose risks of privacy leaks and are mostly passive monitoring, lacking active interaction and cognitive training functions. Other solutions use sound or motion sensors to detect anomalies, but these are easily interfered with in complex home environments and have difficulty accurately distinguishing between normal activities and risky events.

[0004] Furthermore, existing solutions primarily focus on physiological safety monitoring, neglecting the mental health and cognitive function maintenance of the elderly. There is a lack of an integrated intelligent monitoring solution that can simultaneously address safety alerts, contextual interaction, cognitive training, and health assessment.

[0005] Therefore, there is an urgent need for an intelligent monitoring system that can integrate multimodal environmental perception, artificial intelligence analysis and personalized situational interaction to achieve comprehensive, proactive and personalized protection of the safety and health of the elderly. Summary of the Invention

[0006] Purpose of the Invention: The purpose of this invention is to provide a method and system for health monitoring of the elderly based on environmental perception and contextual interaction, in order to improve the accuracy, initiative, interactivity, and comprehensiveness of elderly monitoring.

[0007] Technical Solution: The present invention provides a method for health monitoring of the elderly based on environmental perception and contextual interaction. The method is executed collaboratively by a mobile terminal application, a wearable terminal, and a cloud-based intelligent AI engine, and includes the following steps:

[0008] S1. The user selects the operating mode through a mobile terminal application. The operating mode includes a first monitoring mode and a second interactive mode. The first monitoring mode judges the risk by combining forward environmental perception and abnormal acoustic events. The second interactive mode performs contextual cognitive interaction based on periodic or event-triggered forward environmental perception.

[0009] S2. Execute the sensor startup and data acquisition process corresponding to the selected mode: If the first monitoring mode is selected, the abnormal acoustic event detection model is started locally on the wearable terminal, and the front camera module is set to enter a low-power standby state; if the second interaction mode is selected, the periodic environmental perception module is started locally on the wearable terminal, and the front camera module is set to a timed wake-up acquisition state or an event-triggered acquisition state based on simple motion detection.

[0010] S3. Execute the cloud analysis and feedback process corresponding to the selected mode: If the current mode is the first monitoring mode, upload the data packet of the risk event to be confirmed to the cloud intelligent AI engine; if the current mode is the second interaction mode, upload the interactive scenario material data to the cloud intelligent AI engine.

[0011] S4. Record and analyze user interaction and event data in the two modes. Based on the risk event records and user confirmation responses in the first monitoring mode and the task completion performance data in the second interaction mode, update the user health profile and regularly generate a comprehensive health assessment report that integrates security situation and cognitive activity.

[0012] Furthermore, step S2 specifically includes the following steps:

[0013] S21. If the first monitoring mode is selected, the abnormal acoustic event detection model is started locally on the wearable terminal, and the front-facing camera module is set to enter a low-power standby state; wherein:

[0014] The abnormal acoustic event detection model analyzes the acoustic signals continuously collected by the microphone of the wearable terminal in real time and detects preset abnormal acoustic patterns, including falling and impact sounds, violent friction sounds, and painful groans.

[0015] When an abnormal acoustic event that meets the preset risk characteristics is detected, the first control command is automatically generated to wake up and control the forward camera module to immediately acquire a frame of forward environment image. The image is used to capture the scene in front of the user, the ground conditions and existing obstacles, and the image is packaged with the acoustic feature fragment at the trigger time and marked as a risk event data packet to be confirmed.

[0016] S22. If the second interaction mode is selected, the periodic environmental perception module is activated locally on the wearable terminal, and the front-facing camera module is set to a timed wake-up acquisition state or an event-triggered acquisition state based on simple motion detection; wherein: according to the preset interaction plan or the user's environment is perceived to change, a second control command is generated periodically or triggered to control the front-facing camera module to acquire an environmental image or a short video clip in front of the current user's field of vision. The material is used to capture the current situation or the situation that the user is concerned about, and is marked as interactive situation material data.

[0017] Furthermore, the abnormal acoustic event detection model is obtained through the following workflow:

[0018] During the training phase, an audio dataset containing various everyday environmental sounds and abnormal acoustic events labeled with falls, collisions, and cries for help was collected on the server side.

[0019] Construct and train a classification model based on Mel spectrogram features and a temporal convolutional neural network, specifically as follows:

[0020] Feature extraction: After the original audio signal is framed, a multi-dimensional feature vector containing static Mel frequency cepstral coefficients and their first and second order differences is extracted to form a time-series feature matrix.

[0021] Core network: It adopts a TCN architecture with causal dilated convolutions; by stacking multiple convolutional layers with exponentially increasing dilation rates, it obtains the temporal receptive field;

[0022] Using model distillation and quantization compression tools, the trained server-side anomalous acoustic event detection model is converted into a lightweight inference model with embedded processor resource constraints suitable for the wearable terminal, ensuring that it can continuously perform low-power real-time monitoring.

[0023] Furthermore, step S3 specifically includes the following steps:

[0024] S31. If the current mode is the first monitoring mode, upload the data packet of the risk event to be confirmed to the cloud intelligent AI engine.

[0025] In the cloud, a multimodal risk joint analysis process is executed: First, scene semantic segmentation and object recognition are performed on the environmental image, including ground type, obstacles in front, furniture positions, and visible body parts; simultaneously, advanced feature extraction and reanalysis are performed on acoustic feature segments; then, the image analysis results and acoustic analysis results are fused, and a comprehensive judgment is made based on a preset risk assessment rule base: if a risk scene that could lead to a fall is identified in the image and the acoustic event is a high-confidence anomaly, or if a human body part is identified in an abnormal position in the image, it is determined to be a high-risk event, and a first alarm message containing environmental image and location information is generated and sent to the remote monitoring terminal; if the risk characteristics are unclear, a first voice confirmation command is generated and sent to the wearable terminal to request the user to respond actively;

[0026] S32. If the current mode is the second interaction mode, the interactive scenario material data is uploaded to the cloud-based intelligent AI engine.

[0027] In the cloud, the contextual understanding and personalized task generation process is executed as follows: First, deep scene understanding is performed on images or videos to identify the main objects, human activities, scene categories, and spatial relationships, and scene description text is generated. Combining the user's historical interaction profile and cognitive ability assessment, contextualized cognitive training tasks are dynamically constructed. These tasks include question-and-answer based on object and scene recognition, describing the current environment, recalling relevant experiences, or completing simple virtual operation guidance. Based on the constructed tasks, structured and conversational second voice interaction content is generated and sent to wearable terminals for playback.

[0028] Furthermore, in step S31, the multimodal risk joint analysis process specifically includes:

[0029] After performing semantic segmentation on the uploaded environmental image in the cloud, the image risk index is calculated:

[0030]

[0031] Ground risk factors Calculated based on the type of ground and its proportion in the image;

[0032] Obstacle risk factors The weighted sum is calculated based on the number and size of the identified tripping obstacles and their distance from the image center.

[0033] Human abnormal factors If semantic segmentation identifies a human body part in an abnormal location, a high constant value of 0.8 is assigned; otherwise, 0 is assigned.

[0034] , , These are the normalized weighting coefficients, and ;

[0035] The cloud-based acoustic analysis submodule extracts and classifies features from the uploaded audio segments, outputting a confidence score between 0 and 1. This indicates the degree of credibility of the audio clip belonging to a real abnormal event;

[0036] Final Risk Score Calculated by the following formula:

[0037]

[0038] in, and To merge weights, and The system presets two levels of decision thresholds:

[0039] High risk threshold Right now : If the event is identified as high-risk, an immediate alert will be triggered;

[0040] Medium risk threshold Right now If the event is determined to require confirmation, a voice confirmation command will be triggered.

[0041] Furthermore, in step S32, the context understanding and personalized task generation process specifically includes:

[0042] The visual analysis submodule first generates a scene graph. Among them, nodes The detected object entities in the image; edge set Semantic relationships between objects;

[0043] Input the scene graph into a pre-trained text generation model to generate scene description text;

[0044] The personalization adapter maintains a dynamic user profile vector. Among them, cognitive ability level Calculated based on historical task performance; average reaction speed Average time for users to answer questions; Interest topic vectors It represents the degree of user preference for different topics;

[0045] After the task template engine generates the initial task, the personalization adapter performs the following operations:

[0046] Difficulty adjustment: based on Choose the matching one from multiple preset difficulty variations of the task template;

[0047] Content bias: computational task topic and Based on similarity, prioritize or enhance topics that users are interested in;

[0048] Interaction rhythm adjustment: based on Insert waiting intervals or speech rate prompts into the generated voice interaction script.

[0049] Furthermore, following step S32, an interactive feedback evaluation process is also included:

[0050] After the second voice interaction content is played, the microphone of the wearable terminal is activated to collect user voice feedback within a preset duration.

[0051] User voice feedback is uploaded to the cloud and converted into text by an automatic speech recognition module;

[0052] The natural language understanding module analyzes the semantic relevance, keyword coverage, and logical coherence between the text and the expected answer to the task.

[0053] The analysis results, including answer accuracy, reaction time, and fluency, will be recorded as performance indicators for this cognitive training and used to dynamically update the user's cognitive ability assessment model.

[0054] Furthermore, the mobile terminal application also provides a mode configuration interface, allowing users or guardians to configure a first monitoring mode and a second interaction mode respectively:

[0055] Scheduled time periods for daily automatic activation;

[0056] In the first monitoring mode, the types of abnormal acoustic events triggered by the forward-facing camera and their sensitivity thresholds are determined.

[0057] In the second interactive mode, the frequency of scene acquisition by the forward-facing camera, and the threshold for triggering the acquisition of ambient light or motion changes.

[0058] The present invention also discloses an elderly health monitoring system based on environmental perception and contextual interaction, including a mobile terminal application, a wearable terminal, a cloud-based intelligent AI engine, and a remote monitoring terminal.

[0059] The wearable terminal includes:

[0060] The sensing unit includes a microphone module and a front-facing camera module;

[0061] A local processing unit, connected to the sensing unit, is used for:

[0062] In the first monitoring mode, a lightweight abnormal acoustic event detection model is loaded and run. When an event is detected, the forward-facing camera module is controlled to wake up from a low-power state and capture an image.

[0063] In the second interactive mode, the periodic environmental perception module is run, and the forward-facing camera module is controlled to collect contextual materials according to the configured period or internal trigger events;

[0064] The first communication unit is used for wireless data interaction with the cloud-based intelligent AI engine;

[0065] The cloud-based intelligent AI engine includes:

[0066] The second communication unit is used to receive data from the wearable terminal;

[0067] The pattern routing and processing module is used to route data to different analysis pipelines based on the pattern labels associated with the data.

[0068] The first analysis pipeline is used to execute the multimodal risk joint analysis process;

[0069] The second analysis pipeline is used to perform contextual understanding and personalized task generation processes;

[0070] The report generation module is used to aggregate analysis results and generate a comprehensive health assessment report.

[0071] The mobile terminal application is used to provide a mode selection, parameter configuration, and report viewing interface;

[0072] The remote monitoring terminal is used to receive and display alarm information and comprehensive health assessment reports.

[0073] Furthermore, the local processing unit of the wearable terminal also includes a power consumption and instruction management module, which manages the switching of the working state of the front-facing camera module to ensure a smooth and rapid transition between the low-power standby and event-triggered capture in the first monitoring mode and the timed wake-up acquisition in the second interactive mode, avoiding instruction conflicts and resource contention.

[0074] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0075] 1. This invention achieves accurate risk identification through multimodal fusion: by combining abnormal acoustic event detection with forward environmental image analysis, the false alarm rate is significantly reduced, and the accuracy and timeliness of risk identification such as falls are improved.

[0076] 2. This invention provides an active and contextualized cognitive interaction: it dynamically generates cognitive training tasks based on the user's environment, transforming everyday scenarios into cognitive training materials, thereby enhancing the fun, relevance, and training effect of the interaction.

[0077] 3. This invention constructs an integrated health assessment system of "safety-cognition": the system not only focuses on physical safety, but also continuously tracks cognitive activity and task performance, and generates a comprehensive health report that integrates safety situation and cognitive ability, providing data support for comprehensive health management.

[0078] 4. This invention achieves a balance between low power consumption and high performance: through the collaboration of a local lightweight anomaly detection model and cloud-based intelligent analysis, it provides powerful AI analysis capabilities while ensuring long battery life for wearable devices.

[0079] 5. This invention improves the system's usability and personalization: it supports multiple operating modes and rich parameter configurations, allowing users or guardians to flexibly customize monitoring strategies according to actual needs, and continuously learns to optimize the user experience. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of the system architecture provided in an embodiment of the present invention;

[0081] Figure 2 This is a general flowchart of the method provided in the embodiments of the present invention;

[0082] Figure 3 This is a flowchart illustrating the execution of the first monitoring mode (risk identification mode) provided in an embodiment of the present invention.

[0083] Figure 4 A flowchart of multimodal risk joint analysis provided in this embodiment of the invention;

[0084] Figure 5 A flowchart of context understanding and personalized task generation provided in this embodiment of the invention;

[0085] Figure 6 This is a schematic diagram of a wearable terminal structure provided in an embodiment of the present invention. Detailed Implementation

[0086] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0087] Example 1: System Architecture and Workflow

[0088] like Figure 1 As shown, the intelligent monitoring system 100 provided by this invention mainly includes: a wearable terminal 110 (such as a smart badge or pendant), a monitoring application (App) running on a mobile terminal 120 (such as a smartphone / tablet), a cloud-based intelligent AI engine 130, and a remote monitoring terminal 140 (such as a mobile phone / computer belonging to a child or medical staff). The wearable terminal 110 is connected to the mobile terminal 120 and the cloud 130 via wireless communication methods (such as 4G / 5G, Wi-Fi, Bluetooth).

[0089] like Figure 2 As shown, the overall process of the method of the present invention includes:

[0090] S200: System initialization and connection establishment. The user wears the wearable terminal, launches the mobile app, and completes device binding and network connection.

[0091] S210: Mode Selection and Configuration. Users or guardians can select the operating mode (monitoring / interaction) through the App and can configure detailed parameters.

[0092] S220: Sensor activation and data acquisition. Depending on the selected mode, the wearable terminal activates the corresponding sensors (microphone, camera) and enters the working state.

[0093] S230: Local Event Detection and Data Preprocessing. In the first monitoring mode, abnormal acoustic events are detected locally; in the second interactive mode, environmental images are collected according to plan or trigger. S240: Data Upload and Cloud Routing. The collected data is packaged and uploaded to the cloud, which routes it to the corresponding analysis pipeline based on data tags. S250: Cloud-based Intelligent Analysis. Multimodal risk analysis or contextual understanding and task generation are performed. S260: Feedback Execution and Interaction. The cloud distributes the analysis results (alarms, voice confirmation commands, or interactive tasks) to the wearable terminal for execution. S270: Data Recording and Profile Update. All interaction and event data are recorded to update the user's health profile. S280: Report Generation and Push. Comprehensive health assessment reports are generated periodically and pushed to the App and remote monitoring terminals.

[0094] Example 2: Detailed Process of First Monitoring Mode (Risk Identification Mode)

[0095] refer to Figure 3 The first guardianship model focuses on safety risks, especially the identification of fall risks.

[0096] S301: The user enables "Monitoring Mode" via the App. The wearable terminal locally loads a lightweight abnormal acoustic event detection model, the microphone continuously monitors, and the front-facing camera enters low-power standby mode.

[0097] S302: The acoustic signals collected by the microphone are analyzed in real time by the model to detect preset abnormal patterns (such as falling and impact sounds, painful groans).

[0098] S303: When a high-confidence abnormal event is detected, the local processing unit immediately generates a control command to wake up the forward-facing camera and capture a frame of the environment in front of the user.

[0099] S304: Package the acoustic feature fragments at the trigger moment with the environmental image, mark it as a "risk event data package to be confirmed", and upload it to the cloud AI engine via a mobile app or directly.

[0100] S305: Launch the "Multimodal Risk Joint Analysis" pipeline in the cloud (e.g., Figure 4 (As shown). First, the image recognition submodule performs semantic segmentation on the image, identifying ground types (floor, carpet), obstacles (wires, toys), furniture layout, and visible human body parts. Simultaneously, the acoustic analysis submodule extracts and classifies deep features from sound segments.

[0101] S306: The fusion decision submodule combines image risk indices (such as slippery ground or blocked passages) with acoustic event confidence to calculate the final risk score.

[0102] S307: Decision based on risk score threshold: If determined to be high risk, immediately generate the first alarm information including environmental snapshot and location, and send it to the remote monitoring terminal; if the risk is unclear, generate the first voice confirmation command (such as "Are you okay? Please speak if you need help"), send it to the wearable terminal for playback, and listen for the user's response.

[0103] S308: Record this event (including the judgment result, user response, and environment snapshot) in the user event log.

[0104] Example 3: Detailed Process of the Second Interaction Mode (Contextual Interaction Mode)

[0105] refer to Figure 4 The second interaction mode aims to maintain cognitive activity through environmental interaction.

[0106] S401: The user enables "Interactive Mode" via the App. The wearable terminal activates the periodic environmental awareness module, and the front-facing camera is triggered to wake up according to the settings (such as every 30 minutes) or based on simple motion detection.

[0107] S402: The camera captures images or videos of the main environment in front of the user's current field of vision, marks them as "interactive contextual material data", and uploads them to the cloud.

[0108] S403: Cloud-based launch of the "Contextual Understanding and Personalized Task Generation" pipeline (e.g.) Figure 5 (As shown). The visual analysis submodule performs in-depth understanding of the materials, identifies the scene (living room, kitchen), main objects (sofa, dining table, TV), human activities and spatial relationships, and generates scene description text.

[0109] S404: The task template engine matches the initial task (such as "Describe what furniture you see?") from the preset cognitive task template library based on the scene description.

[0110] S405: The personalized adapter retrieves the user's historical interaction data (cognitive difficulty preference, reaction speed, and interest topics) and adjusts the sentence complexity, prompting method, and content focus of the initial task to form the final personalized voice interaction script.

[0111] S406: Convert the script into structured, conversational second voice interaction content (such as "Good morning! I see you are in the kitchen. The flower on the windowsill is blooming beautifully today. Do you remember its name?") via text-to-speech (TTS) and send it to the wearable terminal for playback.

[0112] S407: After playback is complete, the wearable device turns on its microphone and collects user voice feedback within a preset duration.

[0113] S408: The feedback speech is uploaded, converted into text by automatic speech recognition (ASR), and then analyzed by natural language understanding (NLU) to determine its semantic relevance and logical coherence with the task expectation.

[0114] S409: Record the analysis results (accuracy, reaction time, and fluency) as performance indicators for this cognitive training and use them to dynamically update the user's cognitive ability assessment model.

[0115] Example 4: Pattern Collaboration and Health Profile Construction

[0116] The two modes of this invention do not operate in isolation, but rather work together to build a comprehensive user health profile.

[0117] Data aggregation: The frequency of risk events, user confirmation response rate, and false alarms in the first monitoring mode; and the task completion rate, response time, answer accuracy, and participation enthusiasm in the second interactive mode are all continuously recorded.

[0118] User profile updates: Based on this multi-dimensional data, the system dynamically updates the user's health profile. This profile not only includes traditional physiological safety indicators (such as the fall risk index), but also innovatively incorporates dimensions such as cognitive activity index, contextual interaction participation, and daily activity patterns.

[0119] Report Generation: Based on health profiles, the system automatically generates comprehensive health assessment reports with illustrations and text on a regular (e.g., weekly / monthly) basis. These reports not only highlight potential safety risks but also showcase cognitive training results and trends, providing remote caregivers with a comprehensive and intuitive view of the elderly person's health status.

[0120] Intelligent Recommendations: Based on health profiles, the system can intelligently recommend switching modes or adjusting parameters. For example, if the completion rate of cognitive tasks has declined recently, it can suggest increasing the frequency of interaction or adjusting the task difficulty.

[0121] Regarding wearable terminals and system modules

[0122] Figure 6 A specific implementation of a wearable terminal 110 is shown. It includes a housing 701 and a back clip 702. Internally, it integrates a main control chip (local processing unit) 703, a microphone array (voice acquisition unit) 704, a speaker (voice playback unit) 705, a front-facing camera 706, and a communication module 707. The local processing unit 703 is responsible for running a lightweight anomaly detection model, managing sensor states, and executing low-power strategies. A power consumption and instruction management submodule ensures that the camera switches efficiently and without conflict between several states: "low-power standby," "event-triggered capture," and "timed wake-up acquisition."

[0123] This invention creates a novel paradigm for intelligent health monitoring of the elderly through deep collaboration between wearable terminals, mobile applications, and cloud-based AI. It is not merely a "fall alarm," but a smart assistant that functions as both a "safety guardian" and a "cognitive partner." By seamlessly integrating accurate risk perception and positive situational interaction into real-life scenarios, this invention provides a practical technical solution for improving the quality of life for the elderly and reducing the caregiving burden on families and society.

[0124] It should be noted that, in the embodiments of the present invention, functions such as abnormal acoustic event detection, environmental image analysis, context understanding, dialogue generation, speech synthesis, and natural language understanding can be flexibly deployed on the wearable terminal, mobile terminal, or cloud according to computing power requirements and real-time requirements, which does not deviate from the core idea of ​​the system architecture of the present invention.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for health monitoring of the elderly based on environmental perception and contextual interaction, characterized in that, The method is executed collaboratively by a mobile terminal application, a wearable terminal, and a cloud-based intelligent AI engine, and includes the following steps: S1. The user selects the operating mode through a mobile terminal application. The operating mode includes a first monitoring mode and a second interactive mode. The first monitoring mode judges the risk by combining forward environmental perception and abnormal acoustic events. The second interactive mode performs contextual cognitive interaction based on periodic or event-triggered forward environmental perception. S2. Execute the sensor startup and data acquisition process corresponding to the selected mode: If the first monitoring mode is selected, the abnormal acoustic event detection model is started locally on the wearable terminal, and the front camera module is set to enter a low-power standby state; if the second interaction mode is selected, the periodic environmental perception module is started locally on the wearable terminal, and the front camera module is set to a timed wake-up acquisition state or an event-triggered acquisition state based on simple motion detection. Step S2 specifically includes the following steps: S21. If the first monitoring mode is selected, the abnormal acoustic event detection model is started locally on the wearable terminal, and the front-facing camera module is set to enter a low-power standby state; wherein: The abnormal acoustic event detection model analyzes the acoustic signals continuously collected by the microphone of the wearable terminal in real time and detects preset abnormal acoustic patterns, including falling and impact sounds, violent friction sounds, and painful groans. When an abnormal acoustic event that meets the preset risk characteristics is detected, the first control command is automatically generated to wake up and control the forward camera module to immediately acquire a frame of forward environment image. The image is used to capture the scene in front of the user, the ground conditions and existing obstacles, and the image is packaged with the acoustic feature fragment at the trigger time and marked as a risk event data packet to be confirmed. S22. If the second interaction mode is selected, the periodic environmental perception module is activated locally on the wearable terminal, and the front-facing camera module is set to a timed wake-up acquisition state or an event-triggered acquisition state based on simple motion detection; wherein: according to the preset interaction plan or the user's environment is perceived to change, a second control command is generated periodically or triggered to control the front-facing camera module to acquire an environmental image or a short video clip in front of the current user's field of vision. The material is used to capture the current situation or the situation that the user is concerned about, and is marked as interactive situation material data; The abnormal acoustic event detection model is obtained through the following workflow: During the training phase, an audio dataset containing various everyday environmental sounds and abnormal acoustic events labeled with falls, collisions, and cries for help was collected on the server side. Construct and train a classification model based on Mel spectrogram features and a temporal convolutional neural network, specifically as follows: Feature extraction: After the original audio signal is framed, a multi-dimensional feature vector containing static Mel frequency cepstral coefficients and their first and second order differences is extracted to form a time-series feature matrix. Core network: It adopts a TCN architecture with causal dilated convolutions; by stacking multiple convolutional layers with exponentially increasing dilation rates, it obtains the temporal receptive field; Using model distillation and quantization compression tools, the trained server-side abnormal acoustic event detection model is converted into a lightweight inference model with embedded processor resource constraints suitable for the wearable terminal, ensuring that it can continuously perform low-power real-time monitoring. S3. Execute the cloud analysis and feedback process corresponding to the selected mode: If the current mode is the first monitoring mode, upload the data packet of the risk event to be confirmed to the cloud intelligent AI engine; if the current mode is the second interaction mode, upload the interactive scenario material data to the cloud intelligent AI engine. S4. Record and analyze user interaction and event data in the two modes. Based on the risk event records and user confirmation responses in the first monitoring mode and the task completion performance data in the second interaction mode, update the user health profile and regularly generate a comprehensive health assessment report that integrates security situation and cognitive activity.

2. The method for health monitoring of the elderly based on environmental perception and contextual interaction according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. If the current mode is the first monitoring mode, upload the data packet of the risk event to be confirmed to the cloud intelligent AI engine. In the cloud, a multimodal risk joint analysis process is executed: First, scene semantic segmentation and object recognition are performed on the environmental images, including ground type, obstacles in front, furniture positions, and visible body parts; at the same time, advanced feature extraction and reanalysis are performed on acoustic feature segments. Subsequently, by integrating the image analysis results and acoustic analysis results, a comprehensive judgment is made based on a preset risk assessment rule base: if a risk scene that could lead to a fall is identified in the image and the acoustic event is a high-confidence anomaly, or if a human body part is identified in an abnormal position in the image, it is determined to be a high-risk event, and a first alarm message containing environmental image and location information is generated and sent to the remote monitoring terminal; if the risk characteristics are unclear, a first voice confirmation command is generated and sent to the wearable terminal to request the user to respond actively. S32. If the current mode is the second interaction mode, the interactive scenario material data is uploaded to the cloud-based intelligent AI engine. In the cloud, the process of contextual understanding and personalized task generation is executed: First, deep scene understanding is performed on images or videos to identify the main objects, human activities, scene categories and spatial relationships, and scene description text is generated; combined with the user's historical interaction profile and cognitive ability assessment, contextualized cognitive training tasks are dynamically constructed, including question and answer based on object and scene recognition, describing the current environment, recalling relevant experiences or completing simple virtual operation guidance. Based on the constructed task, structured, conversational second voice interaction content is generated and sent to wearable terminals for playback.

3. The method for health monitoring of the elderly based on environmental perception and contextual interaction according to claim 2, characterized in that, In step S31, the multimodal risk joint analysis process is specifically as follows: After performing semantic segmentation on the uploaded environmental image in the cloud, the image risk index is calculated: ； Ground risk factors Calculated based on the type of ground and its proportion in the image; Obstacle risk factors The weighted sum is calculated based on the number and size of the identified tripping obstacles and their distance from the image center. Human abnormal factors If semantic segmentation identifies a human body part in an abnormal location, a high constant value of 0.8 is assigned; otherwise, 0 is assigned. , , These are the normalized weighting coefficients, and ; The cloud-based acoustic analysis submodule extracts and classifies features from the uploaded audio segments, outputting a confidence score between 0 and 1. This indicates the degree of credibility of the audio clip belonging to a real abnormal event; Final Risk Score Calculated by the following formula: ； in, and To merge weights, and The system presets two levels of decision thresholds: High risk threshold Right now : If the event is identified as high-risk, an immediate alert will be triggered; Medium risk threshold Right now If the event is determined to require confirmation, a voice confirmation command will be triggered.

4. The method for health monitoring of the elderly based on environmental perception and contextual interaction according to claim 2, characterized in that, In step S32, the context understanding and personalized task generation process is specifically as follows: The visual analysis submodule first generates a scene graph. Among them, nodes The detected object entities in the image; edge set Semantic relationships between objects; Input the scene graph into a pre-trained text generation model to generate scene description text; The personalization adapter maintains a dynamic user profile vector. Among them, cognitive ability level Calculated based on historical task performance; average reaction speed Average time for users to answer questions; Interest topic vectors It represents the degree of user preference for different topics; After the task template engine generates the initial task, the personalization adapter performs the following operations: Difficulty adjustment: based on Choose the matching one from multiple preset difficulty variations of the task template; Content bias: computational task topic and Based on similarity, prioritize or enhance topics that users are interested in; Interaction rhythm adjustment: based on Insert waiting intervals or speech rate prompts into the generated voice interaction script.

5. A method for health monitoring of the elderly based on environmental perception and contextual interaction according to claim 2, characterized in that, Following step S32, an interactive feedback evaluation process is also included: After the second voice interaction content is played, the microphone of the wearable terminal is activated to collect user voice feedback within a preset duration. User voice feedback is uploaded to the cloud and converted into text by an automatic speech recognition module; The natural language understanding module analyzes the semantic relevance, keyword coverage, and logical coherence between the text and the expected answer to the task. The analysis results, including answer accuracy, reaction time, and fluency, will be recorded as performance indicators for this cognitive training and used to dynamically update the user's cognitive ability assessment model.

6. The method for health monitoring of the elderly based on environmental perception and contextual interaction according to claim 1, characterized in that, The mobile terminal application also provides a mode configuration interface, allowing users or guardians to configure a first monitoring mode and a second interactive mode respectively: Scheduled time periods for daily automatic activation; In the first monitoring mode, the types of abnormal acoustic events triggered by the forward-facing camera and their sensitivity thresholds are determined. In the second interactive mode, the frequency of scene acquisition by the forward-facing camera, and the threshold for triggering the acquisition of ambient light or motion changes.

7. A health monitoring system for the elderly based on environmental perception and contextual interaction, used to implement the method as described in any one of claims 1-6, characterized in that, This includes mobile terminal applications, wearable terminals, cloud-based intelligent AI engines, and remote monitoring terminals; The wearable terminal includes: The sensing unit includes a microphone module and a front-facing camera module; A local processing unit, connected to the sensing unit, is used for: In the first monitoring mode, a lightweight abnormal acoustic event detection model is loaded and run. When an event is detected, the forward-facing camera module is controlled to wake up from a low-power state and capture an image. In the second interactive mode, the periodic environmental perception module is run, and the forward-facing camera module is controlled to collect contextual materials according to the configured period or internal trigger events; The first communication unit is used for wireless data interaction with the cloud-based intelligent AI engine; The cloud-based intelligent AI engine includes: The second communication unit is used to receive data from the wearable terminal; The pattern routing and processing module is used to route data to different analysis pipelines based on the pattern labels associated with the data. The first analysis pipeline is used to execute the multimodal risk joint analysis process; The second analysis pipeline is used to perform contextual understanding and personalized task generation processes; The report generation module is used to aggregate analysis results and generate a comprehensive health assessment report. The mobile terminal application is used to provide a mode selection, parameter configuration, and report viewing interface; The remote monitoring terminal is used to receive and display alarm information and comprehensive health assessment reports.

8. The elderly health monitoring system based on environmental perception and contextual interaction according to claim 7, characterized in that, The local processing unit of the wearable terminal also includes a power consumption and instruction management module, which manages the switching of the working state of the front-facing camera module to ensure a smooth and fast transition between the low-power standby and event-triggered capture in the first monitoring mode and the timed wake-up acquisition in the second interactive mode, avoiding instruction conflicts and resource contention.

Images