An elderly fall risk assessment and dynamic protection intervention system and method
By combining multimodal sensors and lightweight algorithms, real-time and accurate assessment and graded protection of fall risk for the elderly are achieved, solving the problems of false alarms and high resource consumption in existing technologies, and providing efficient and reliable protective measures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEFEI CAREER TECHNICAL COLLEGE
- Filing Date
- 2025-07-23
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies for fall detection and protection for the elderly suffer from problems such as environmental interference leading to false alarms, high cloud processing latency, lack of proactive protection capabilities, and high consumption of computing resources.
It employs multimodal sensors to collect data in real time, combines lightweight intelligent algorithms for attitude analysis and fall risk prediction, predicts imbalance trends through biomechanical simulation, utilizes MEMS micro-jet airbags for graded and precise protection, and achieves efficient deployment through embedded edge computing and privacy-preserving collaborative learning.
It improves the accuracy and real-time performance of fall risk assessment, reduces false alarm rate, provides a reliable basis for tiered protection strategies, and enhances the efficiency of system computing resource utilization.
Smart Images
Figure CN120892990B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, and in particular to a fall risk assessment and dynamic protective intervention system and method for the elderly. Background Technology
[0002] Many older adults experience falls, with approximately 20% of these falls resulting in serious injuries such as hip fractures and traumatic brain injury. These falls not only significantly reduce their quality of life but also impose a heavy medical burden and socioeconomic pressure. Medical research indicates that falls in the elderly are primarily attributed to declining physical function (such as weakened muscle strength and decreased balance), chronic diseases (such as Parkinson's disease and cardiovascular diseases), medication side effects (such as dizziness caused by antihypertensive drugs), and environmental factors (such as slippery surfaces and dim lighting).
[0003] While current technologies for fall detection and prevention among the elderly have made some progress, significant drawbacks remain: traditional fall detection technologies suffer from problems such as false alarms due to environmental interference, high cloud processing latency, lack of proactive protection capabilities, and high computational resource consumption. Therefore, this invention provides a fall risk assessment and dynamic protective intervention system and method for the elderly. Summary of the Invention
[0004] The purpose of this invention is to solve the problems in the prior art by proposing a fall risk assessment and dynamic protection intervention system and method for the elderly.
[0005] A fall risk assessment and dynamic prevention intervention system for the elderly includes a hardware module, a software and algorithm module, an intervention mechanism module, a system integration and optimization module, and an application scenario module, wherein:
[0006] The hardware module is used to collect real-time data on the elderly’s movement posture, gait and environment through multimodal sensors, providing hardware support for risk assessment and protection.
[0007] The software and algorithm module is used to perform real-time posture analysis of the elderly, fall risk prediction and generation of graded protection strategies by integrating multimodal data and lightweight intelligent algorithms, and integrates privacy-preserving encrypted computing.
[0008] The intervention mechanism module is used to predict imbalance trends through biomechanical simulation, and combine MEMS micro-jet airbags with reinforcement learning dynamic thresholds to carry out graded and precise protection and fall emergency linkage.
[0009] The system integration and optimization module is used for efficient deployment and continuous optimization through embedded edge computing, privacy-preserving collaborative learning, and dynamic resource management;
[0010] The application scenario module is designed to adapt to different environments such as homes, nursing homes, and public places, providing customized fall protection solutions.
[0011] Preferably, the hardware module includes a visual perception unit, an inertial sensing unit, an environmental perception unit, a wearable monitoring unit, and a data preprocessing unit, wherein:
[0012] The visual perception unit is used to acquire human three-dimensional posture coordinates and environmental semantic information through a depth camera and an RGB camera;
[0013] The inertial sensing unit is used to collect motion parameters such as acceleration and angular velocity through the IMU array to compensate for visual occlusion errors;
[0014] The environmental sensing unit is used to detect environmental risk factors such as ground slippage and obstacle distribution using millimeter-wave radar and infrared sensors.
[0015] The wearable monitoring unit is used to integrate a flexible pressure sensor to monitor the distribution of pressure on the sole of the foot and changes in the center of gravity.
[0016] The data preprocessing unit is used to perform noise reduction, format conversion, and synchronization processing on the raw sensor data, providing standardized data for subsequent analysis.
[0017] Preferably, the software and algorithm module includes a multimodal fusion unit, a lightweight analysis unit, a risk prediction unit, a strategy generation unit, and a privacy computing unit, wherein:
[0018] The multimodal fusion unit is used to weightedly fuse visual, IMU, and environmental data using an attention mechanism to generate a spatiotemporally consistent motion representation.
[0019] The lightweight analysis unit performs real-time pose estimation and gait cycle analysis based on a distilled and pruned CNN-LSTM network.
[0020] The risk prediction unit is used to calculate the relationship between the center of mass projection and the support surface through a biomechanical model, and output the probability of falling.
[0021] The strategy generation unit is used to generate corresponding graded protection strategies based on the risk assessment results, and guide the execution of the intervention mechanism;
[0022] The privacy computing unit is used to perform data desensitization and encrypted aggregation at the edge using a federated learning framework.
[0023] Preferably, the lightweight analysis unit performs real-time pose estimation and gait period analysis based on a distilled and pruned CNN-LSTM network, specifically as follows:
[0024] (1) Spatial feature extraction: Depthwise separable convolution is used to replace standard convolution. The calculation formula is as follows:
[0025]
[0026] in, The decomposed depthwise convolution kernel has its parameter count reduced to 1 / 8 of that of the standard convolution.
[0027] (2) Temporal modeling: Channel pruning is performed on the LSTM layer, and the number of hidden layer units is reduced from 256 to 128;
[0028] (3) Knowledge distillation: The teacher model outputs a pose angle heatmap as a soft label to guide student model training. The loss function is:
[0029]
[0030] Where α = 0.7 is the task weight. This represents the KL divergence distillation loss.
[0031] Preferably, the privacy computing unit employs a federated learning framework to perform data desensitization and encrypted aggregation at the edge, as detailed below:
[0032] (1) Data anonymization: k-anonymization is used on the local training data to ensure that at least 5 records are indistinguishable;
[0033] (2) Encryption Aggregation: The Paillier homomorphic encryption algorithm is used to aggregate model parameters. The aggregation process satisfies the following:
[0034]
[0035] in, Here, n is the encryption parameter, and n is the modulus of the public key.
[0036] (3) Differential privacy: Injecting Laplacian noise during global model updates Noise scale Privacy budget ∈ ≤ 0.5.
[0037] Preferably, the intervention mechanism module includes a risk prediction unit, a protection execution unit, a dynamic optimization unit, and an emergency response unit, wherein:
[0038] The risk prediction unit is based on a biomechanical simulation model and analyzes the movement trends of the elderly in real time to predict the risk of imbalance.
[0039] The protection execution unit is used to trigger the MEMS micro-jet airbag for physical buffer protection according to the graded protection strategy.
[0040] The dynamic optimization unit is used to dynamically adjust the risk threshold based on the user's historical data using the PPO reinforcement learning algorithm.
[0041] The emergency response unit is used to automatically trigger a GPS location alarm and push medical records to the emergency center after a fall is confirmed.
[0042] Preferably, the protection execution unit triggers the MEMS micro-jet airbag for physical buffering protection according to the graded protection strategy, and the specific operation is as follows:
[0043] (1) Gas generation mechanism: A solid sodium azide micro combustion and explosion device is used to generate ≥40L of nitrogen gas within 20ms after triggering;
[0044] (2) Multi-zone control: Based on the fall direction prediction results, the inflation pressure of the 6 airbag zones in the waist and hip is independently controlled to 20-30 kPa;
[0045] (3) Linked braking: The airbag triggers synchronously to send a PWM braking signal to the walker, and outputs a reverse torque ≥5N·m.
[0046] Preferably, the system integration and optimization module includes an edge computing unit, a collaborative learning unit, a resource scheduling unit, and a human-computer interaction unit, wherein:
[0047] The edge computing unit deploys the model via Jetson Xavier NX and supports INT8 quantized inference;
[0048] The collaborative learning unit is used to perform cross-institutional model updates using differential privacy-preserving federated learning;
[0049] The resource scheduling unit is used to dynamically allocate computing power and storage resources according to task priority;
[0050] The human-computer interaction unit is used to provide voice prompts, vibration feedback, and AR interface guidance.
[0051] Preferably, the application scenario module includes an environment adaptation unit, an organization management unit, a public deployment unit, and a personalized configuration unit, wherein:
[0052] The environment adaptation unit is used to adjust the deployment scheme of the device according to different layouts and needs of homes, nursing homes, and public places;
[0053] The institutional management unit is used to deploy a multi-node monitoring network, and the central monitoring platform displays the group risk heat map and early warning statistics in real time;
[0054] The public deployment unit is used to achieve low-cost renovation by utilizing existing surveillance cameras and edge AI boxes to cover key areas such as restrooms and staircases;
[0055] The personalized configuration unit is used to customize the protection level via a mobile app.
[0056] A method for assessing fall risk and implementing dynamic preventive interventions in the elderly is also proposed, including the following steps:
[0057] S1. Real-time synchronous acquisition of motion posture, gait and environmental data through depth camera, IMU array and millimeter wave radar, and spatiotemporal alignment and noise reduction processing.
[0058] S2. Multimodal data is weighted and fused using an attention mechanism. Real-time pose estimation is performed based on a lightweight CNN-LSTM network. The relationship between the center of mass and the support surface is calculated using a biomechanical model, and the graded fall probability is output.
[0059] S3. Dynamically generate protection strategies based on the probability of falling. When the risk level exceeds the threshold, trigger the multi-zone inflation of the MEMS micro-jet airbag and the linkage braking of the walking aid.
[0060] S4. Aggregate encrypted model parameters through a federated learning framework, inject ε≤0.5 differential privacy noise, and dynamically adjust the risk judgment threshold using PPO reinforcement learning.
[0061] S5. Based on the environmental characteristics of homes, nursing homes, or public places, adaptively adjust the sensor deployment scheme and protection response strategy, and realize personalized configuration through a mobile APP.
[0062] Compared with existing technologies, the advantages of this invention are:
[0063] This invention achieves deep fusion of visual, inertial, and environmental data through an attention mechanism, combines a lightweight CNN-LSTM network to quickly extract spatiotemporal features, and then accurately calculates the dynamic relationship between the center of mass and the support surface based on a biomechanical model. This effectively improves the real-time performance of posture analysis and the accuracy of fall risk assessment. While reducing computational resource consumption, it greatly improves the accuracy of the system's prediction of fall risk in the elderly and reduces the false alarm rate, providing a reliable basis for graded protection strategies. Attached Figure Description
[0064] Figure 1 This is a schematic diagram of the system framework of the present invention.
[0065] Figure 2 This is a flowchart of the method of the present invention.
[0066] Figure 3 This is a flowchart of the protection decision execution process in this invention. Detailed Implementation
[0067] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0068] Reference Figure 1 As shown, a fall risk assessment and dynamic prevention intervention system for the elderly includes a hardware module, a software and algorithm module, an intervention mechanism module, a system integration and optimization module, and an application scenario module, wherein:
[0069] The hardware modules are used to collect real-time data on the elderly’s movement posture, gait and environment through multimodal sensors, providing hardware support for risk assessment and protection.
[0070] The software and algorithm modules integrate multimodal data with lightweight intelligent algorithms to perform real-time posture analysis of the elderly, predict fall risk, and generate graded protection strategies, while also integrating privacy-preserving encrypted computing.
[0071] The intervention mechanism module is used to predict imbalance trends through biomechanical simulation, and combined with MEMS micro-jet airbags and reinforcement learning dynamic thresholds, it can carry out graded and precise protection and fall emergency linkage.
[0072] The system integration and optimization module is used for efficient deployment and continuous optimization through embedded edge computing, privacy-preserving collaborative learning, and dynamic resource management;
[0073] The application scenario module is designed to adapt to different environments such as homes, nursing homes, and public places, providing customized fall protection solutions.
[0074] It should be noted that the hardware modules enable multimodal data acquisition, the software and algorithm modules perform data fusion and intelligent analysis to generate protection strategies, the intervention mechanism module executes hierarchical protection and emergency response, the system integration and optimization module achieves efficient deployment and continuous evolution, and the application scenario module completes multi-environment adaptation.
[0075] In this embodiment, the hardware components include a visual perception unit, an inertial sensing unit, an environmental perception unit, a wearable monitoring unit, and a data preprocessing unit, wherein:
[0076] The visual perception unit is used to acquire the three-dimensional pose coordinates of the human body and the semantic information of the environment through a depth camera and an RGB camera;
[0077] The inertial sensing unit is used to collect motion parameters such as acceleration and angular velocity through the IMU array to compensate for visual occlusion errors;
[0078] The environmental sensing unit is used to detect environmental risk factors such as ground slippage and obstacle distribution using millimeter-wave radar and infrared sensors.
[0079] Wearable monitoring units are used to integrate flexible pressure sensors to monitor plantar pressure distribution and changes in the center of gravity.
[0080] The data preprocessing unit is used to perform noise reduction, format conversion, and synchronization processing on the raw sensor data, providing standardized data for subsequent analysis.
[0081] It should be noted that the visual perception unit acquires three-dimensional posture information, combines the motion parameters of the inertial sensing unit to compensate for visual blind spots, coordinates with the environmental perception unit to detect external risk factors, and is supplemented by plantar pressure data from the wearable monitoring unit. Finally, the data preprocessing unit performs multi-source data fusion and standardization processing.
[0082] Furthermore, it should be noted that the depth camera and RGB camera of the visual perception unit acquire the three-dimensional posture coordinates of the human body (accuracy up to ±5mm) and environmental semantic information (such as furniture layout and ground texture) through stereo vision technology, providing visual data for posture analysis. The IMU array (including a three-axis accelerometer and gyroscope) of the inertial sensing unit captures motion parameters in real time at a sampling frequency of 100Hz to compensate for visual blind spots. The environmental perception unit uses millimeter-wave radar and infrared sensors to accurately identify the slipperiness of the ground (judged by the attenuation rate of reflected signals) and the distribution of obstacles. Combined with infrared sensors to monitor environmental risk factors such as temperature and light, an environmental risk database is constructed. The wearable monitoring unit integrates flexible pressure sensors into insoles or belts to monitor the distribution of foot pressure and changes in the center of gravity in real time, assisting in judging the human body's balance status, especially suitable for scenarios with visual blind spots. The data preprocessing unit performs noise reduction (median filtering, wavelet denoising), timestamp synchronization, and format conversion (unified to JSON format) on the raw sensor data, outputting a standardized data stream to provide a foundation for subsequent analysis.
[0083] In this embodiment, the software and algorithm module includes a multimodal fusion unit, a lightweight analysis unit, a risk prediction unit, a policy generation unit, and a privacy computing unit, wherein:
[0084] The multimodal fusion unit is used to weightedly fuse visual, IMU, and environmental data using an attention mechanism to generate spatiotemporally consistent motion representations;
[0085] The lightweight analysis unit performs real-time pose estimation and gait period analysis based on a distilled and pruned CNN-LSTM network. The specific operation is as follows:
[0086] (1) Spatial feature extraction: Depthwise separable convolution is used to replace standard convolution. The calculation formula is as follows:
[0087]
[0088] in, The decomposed depthwise convolution kernel has its parameter count reduced to 1 / 8 of that of the standard convolution.
[0089] (2) Temporal modeling: Channel pruning is performed on the LSTM layer, and the number of hidden layer units is reduced from 256 to 128;
[0090] (3) Knowledge distillation: The teacher model outputs a pose angle heatmap as a soft label to guide student model training. The loss function is:
[0091]
[0092] Where α = 0.7 is the task weight. This represents the KL divergence distillation loss.
[0093] The risk prediction unit is used to calculate the relationship between the center of mass projection and the support surface using a biomechanical model, and outputs the probability of falling.
[0094] The strategy generation unit is used to generate corresponding tiered protection strategies based on the risk assessment results, and to guide the implementation of intervention mechanisms.
[0095] The privacy-preserving computation unit is used to perform data anonymization and encrypted aggregation at the edge using a federated learning framework. The specific operations are as follows:
[0096] (1) Data anonymization: k-anonymization is used on the local training data to ensure that at least 5 records are indistinguishable;
[0097] (2) Encryption Aggregation: The Paillier homomorphic encryption algorithm is used to aggregate model parameters. The aggregation process satisfies the following:
[0098]
[0099] in, Here, n is the encryption parameter, and n is the modulus of the public key.
[0100] (3) Differential privacy: Injecting Laplacian noise during global model updates Noise scale Privacy budget ∈ ≤ 0.5.
[0101] It should be noted that the multimodal fusion unit performs attention-based weighted fusion of visual, IMU, and environmental data to generate a unified motion representation. Real-time pose and gait analysis is then performed by the lightweight analysis unit's distilled and pruned CNN-LSTM network. Spatial feature extraction utilizes depthwise separable convolutions to reduce the number of parameters, while temporal modeling uses channel pruning to compress the LSTM unit and knowledge distillation to improve the accuracy of the small model. The analysis results are input into the risk prediction unit's biomechanical model to calculate the fall probability, which the policy generation unit then uses to formulate tiered protection strategies. Simultaneously, the privacy computation unit achieves data anonymization and secure aggregation within the federated learning framework through k-anonymization, Paillier homomorphic encryption, and differential privacy techniques.
[0102] Furthermore, it should be noted that the risk prediction unit is based on biomechanical principles and outputs a fall probability of 0-100% through a dynamic relationship model between the center of mass and the support surface.
[0103] The strategy generation unit is divided into three levels of protection strategies:
[0104] Level 1 warning (low risk, probability 20%-40%): Triggers the environmental adaptation unit to adjust the brightness of the lights or provides a voice prompt to pay attention to balance;
[0105] Level 2 warning (high risk, probability 40%-80%): The MEMS micro-jet airbag is partially inflated to provide a buffer.
[0106] Level 3 alarm (fall probability > 80%): The airbag is fully inflated and the walking aid is braked.
[0107] In this embodiment, the intervention mechanism module includes a risk prediction unit, a protection execution unit, a dynamic optimization unit, and an emergency response unit, wherein:
[0108] The risk prediction unit is based on a biomechanical simulation model to analyze the movement trends of the elderly in real time and predict the risk of imbalance.
[0109] The protection execution unit is used to trigger the MEMS micro-jet airbags to provide physical buffering protection according to the graded protection strategy. The specific operation is as follows:
[0110] (1) Gas generation mechanism: A solid sodium azide micro combustion and explosion device is used to generate ≥40L of nitrogen gas within 20ms after triggering;
[0111] (2) Multi-zone control: Based on the fall direction prediction results, the inflation pressure of the 6 airbag zones in the waist and hip is independently controlled to 20-30 kPa;
[0112] (3) Linked braking: The airbag triggers synchronously to send a PWM braking signal to the walker, and outputs a reverse torque ≥5N·m.
[0113] The dynamic tuning unit is used to dynamically adjust the risk threshold based on the user's historical data using the PPO reinforcement learning algorithm.
[0114] The emergency linkage unit is used to automatically trigger a GPS location alarm and push medical records to the emergency center after a fall is confirmed.
[0115] It should be noted that the biomechanical simulation model of the risk prediction unit monitors movement trends in real time and predicts the risk of falls. When the risk exceeds the threshold optimized by the dynamic optimization unit through the PPO reinforcement learning algorithm, the protection execution unit immediately triggers the MEMS micro-jet airbag to implement multi-zone (6 independent control areas) precise protection (pressure 20-30kPa), and at the same time, the walker outputs reverse torque to form a coordinated braking. If a fall is confirmed, the emergency linkage unit automatically starts GPS positioning and pushes medical records to the emergency center, forming a closed-loop protection system from risk warning, active protection to emergency rescue.
[0116] Furthermore, it should be noted that the MEMS micro-jet airbag uses a sodium azide combustion device to generate nitrogen gas within 20ms and complete airbag inflation (pressure 20-30kPa) within 100ms.
[0117] In this embodiment, the system integration and optimization module includes an edge computing unit, a collaborative learning unit, a resource scheduling unit, and a human-computer interaction unit, wherein:
[0118] Edge computing units deploy models via Jetson Xavier NX, supporting INT8 quantized inference;
[0119] The collaborative learning unit is used for cross-institutional model updates using differential privacy-preserving federated learning;
[0120] The resource scheduling unit is used to dynamically allocate computing and storage resources according to task priority;
[0121] The human-computer interaction unit is used to provide voice prompts, vibration feedback, and AR interface guidance.
[0122] It should be noted that the edge computing unit implements INT8 quantization deployment and real-time inference of the model on the Jetson Xavier NX platform, providing distributed computing support for the differential privacy federated learning of the collaborative learning unit, enabling cross-institutional model updates to be completed while protecting data privacy; the resource scheduling unit intelligently allocates computing power and storage resources according to the priority of each task to ensure efficient system operation; at the same time, the human-computer interaction unit realizes multimodal interaction guidance through voice, vibration and AR interface.
[0123] Furthermore, it should be noted that voice prompts (such as "Please get up slowly"), vibration feedback (low battery or high risk warnings), and AR interface guidance (displaying safe walking paths) enhance the user experience.
[0124] In this embodiment, the application scenario module includes an environment adaptation unit, an organization management unit, a public deployment unit, and a personalized configuration unit, wherein:
[0125] The environment adaptation unit is used to adjust the deployment scheme of the device according to different layouts and needs of homes, nursing homes, and public places;
[0126] The institutional management unit is used to deploy a multi-node monitoring network, and the central monitoring platform displays group risk heat maps and early warning statistics in real time;
[0127] The public deployment unit is used to achieve low-cost retrofitting of key areas such as restrooms and staircases by utilizing existing surveillance cameras and edge AI boxes;
[0128] The personalized configuration unit allows users to customize the protection level via a mobile app.
[0129] It should be noted that the environment adaptation unit enables customized deployment solutions for homes, nursing homes, and public places. The institutional management unit builds a multi-node monitoring network and generates a heat map of group risks. Meanwhile, the public deployment unit covers key areas with a lightweight solution of "surveillance cameras + edge AI boxes". The mobile APP of the personalized configuration unit enables customization of protection levels.
[0130] Furthermore, it should be noted that users can customize the protection level (such as increasing sensitivity for elderly people with limited mobility), alarm contacts, and privacy settings through the mobile app in the personalized configuration unit, thereby enhancing the system's flexibility.
[0131] Reference Figure 2 As shown, a method for assessing fall risk and implementing dynamic preventive intervention for the elderly includes the following steps:
[0132] S1. Real-time synchronous acquisition of motion posture, gait and environmental data through depth camera, IMU array and millimeter wave radar, and spatiotemporal alignment and noise reduction processing.
[0133] S2. Multimodal data is weighted and fused using an attention mechanism. Real-time pose estimation is performed based on a lightweight CNN-LSTM network. The relationship between the center of mass and the support surface is calculated using a biomechanical model, and the graded fall probability is output.
[0134] S3. Dynamically generate protection strategies based on the probability of falling. When the risk level exceeds the threshold, trigger the multi-zone inflation of the MEMS micro-jet airbag and the linkage braking of the walking aid.
[0135] S4. Aggregate encrypted model parameters through a federated learning framework, inject ε≤0.5 differential privacy noise, and dynamically adjust the risk judgment threshold using PPO reinforcement learning.
[0136] S5. Based on the environmental characteristics of homes, nursing homes, or public places, adaptively adjust the sensor deployment scheme and protection response strategy, and realize personalized configuration through a mobile APP.
[0137] Example
[0138] S1. Multimodal data acquisition and preprocessing
[0139] The human body's three-dimensional posture coordinates (accuracy ±5mm) are simultaneously acquired by a depth camera (1920×1080@30fps) and an RGB camera. Acceleration (range ±16g) and angular velocity (±2000° / s) data are obtained by combining an IMU array (100Hz sampling). Simultaneously, millimeter-wave radar detects the ground reflectivity (for slipperiness assessment) and obstacle distance (0.2-5m). All sensor data are aligned with timestamps, denoised by median filtering, and converted into a unified JSON format.
[0140] S2, Multimodal Feature Fusion
[0141] The attention mechanism is used to fuse multi-source data, and the calculation formula is as follows:
[0142]
[0143] Where Q, K, and V correspond to the query and key-value matrix of visual, IMU, and environmental features, respectively, and d k For feature dimensions;
[0144] The fused spatiotemporal features include:
[0145] ① 3D coordinates of 17 key points (COCO format);
[0146] ②6-DOF motion parameters;
[0147] ③ Environmental risk score (0-1);
[0148] S3, Real-time Attitude Analysis
[0149] Processing using a lightweight CNN-LSTM network:
[0150] (1) CNN part: Depthwise separable convolution kernel size 3×3, the number of parameters is reduced to 1 / 8 of the standard convolution;
[0151] (2) LSTM part: 128 units of hidden layer after pruning, processing time window length T = 10;
[0152] The output includes:
[0153] ① Gait period phase
[0154] ②Joint angular velocity ω i (i∈[1,17]);
[0155] S4. Fall Risk Assessment
[0156] Calculations based on biomechanical models:
[0157] R risk =1-exp(-λ‖v COM ×d BOS ‖)
[0158] Where: v COM The velocity of the center of mass is (mm / s); d BOS λ is the distance to the support surface boundary (mm); λ is the user adaptability coefficient (initial value 0.05);
[0159] S5. Generation of Tiered Protection Strategy
[0160] Execute according to risk level:
[0161]
[0162]
[0163] S6. Privacy Protection Model Update
[0164] Using a federated learning framework:
[0165] (1) Local training: Each edge node uses cross-entropy loss:
[0166]
[0167] (2) Secure aggregation: Paillier encryption parameters satisfy:
[0168]
[0169] in, Here, n is the encryption parameter, and n is the modulus of the public key.
[0170] (3) Differential privacy: Add Laplace noise
[0171] The privacy budget ε = 0.5 per round;
[0172] S7. Dynamic Parameter Optimization
[0173] Update the risk threshold using the PPO algorithm:
[0174] θ k+1 =argmax θ E[min(r(θ)A,clip(r(θ),1-ε,1+ε)A)]
[0175] Where r(θ) is the ratio of the new strategy to the old strategy; A is the advantage function; and ε = 0.2 is the pruning parameter.
[0176] In summary, this invention effectively improves the real-time performance of posture analysis and the accuracy of fall risk assessment, while reducing computational resource consumption and greatly improving the accuracy of the system's prediction of fall risk in the elderly.
[0177] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.
Claims
1. A fall risk assessment and dynamic protective intervention system for the elderly, characterized in that: It includes hardware components, software and algorithm modules, intervention mechanism modules, system integration and optimization modules, and application scenario modules, among which: The hardware module is used to collect real-time data on the elderly’s movement posture, gait and environment through multimodal sensors, providing hardware support for risk assessment and protection. The software and algorithm module is used to perform real-time posture analysis of the elderly, fall risk prediction and generation of graded protection strategies by integrating multimodal data and lightweight intelligent algorithms, and integrates privacy-preserving encrypted computing. The intervention mechanism module is used to predict imbalance trends through biomechanical simulation, and combine MEMS micro-jet airbags with reinforcement learning dynamic thresholds to carry out graded and precise protection and fall emergency linkage. The system integration and optimization module is used for efficient deployment and continuous optimization through embedded edge computing, privacy-preserving collaborative learning, and dynamic resource management; The application scenario module is designed to adapt to different environments such as homes, nursing homes, and public places, providing customized fall protection solutions. The hardware modules include a visual perception unit, an inertial sensing unit, an environmental perception unit, a wearable monitoring unit, and a data preprocessing unit, wherein: The visual perception unit is used to acquire human three-dimensional posture coordinates and environmental semantic information through a depth camera and an RGB camera; The inertial sensing unit is used to collect motion parameters such as acceleration and angular velocity through the IMU array to compensate for visual occlusion errors; The environmental sensing unit is used to detect environmental risk factors such as ground slippage and obstacle distribution using millimeter-wave radar and infrared sensors. The wearable monitoring unit is used to integrate a flexible pressure sensor to monitor the distribution of pressure on the sole of the foot and changes in the center of gravity. The data preprocessing unit is used to perform noise reduction, format conversion and synchronization processing on the raw sensor data to provide standardized data for subsequent analysis. The software and algorithm module includes a multimodal fusion unit, a lightweight analysis unit, a risk prediction unit, a policy generation unit, and a privacy computing unit, wherein: The multimodal fusion unit is used to weightedly fuse visual, IMU, and environmental data using an attention mechanism to generate a spatiotemporally consistent motion representation. The lightweight analysis unit performs real-time pose estimation and gait cycle analysis based on a distilled and pruned CNN-LSTM network. The risk prediction unit is used to calculate the relationship between the center of mass projection and the support surface through a biomechanical model, and output the probability of falling. The strategy generation unit is used to generate corresponding graded protection strategies based on the risk assessment results, and guide the execution of the intervention mechanism; The privacy computing unit is used to perform data desensitization and encrypted aggregation at the edge using a federated learning framework. The intervention mechanism module includes a risk prediction unit, a protection execution unit, a dynamic optimization unit, and an emergency response unit, wherein: The risk prediction unit is based on a biomechanical simulation model and analyzes the movement trends of the elderly in real time to predict the risk of imbalance. The protection execution unit is used to trigger the MEMS micro-jet airbag for physical buffer protection according to the graded protection strategy. The dynamic optimization unit is used to dynamically adjust the risk threshold based on the user's historical data using the PPO reinforcement learning algorithm. The emergency response unit is used to automatically trigger a GPS location alarm and push medical records to the emergency center after a fall is confirmed.
2. The fall risk assessment and dynamic prevention intervention system for the elderly according to claim 1, characterized in that: The lightweight analysis unit performs real-time pose estimation and gait period analysis based on a distilled and pruned CNN-LSTM network. The specific operations are as follows: (1) Spatial feature extraction: Depthwise separable convolution is used to replace standard convolution, and the calculation formula is as follows: in, The decomposed depthwise convolution kernel has its parameter count reduced to 1 / 8 of that of the standard convolution. (2) Temporal modeling: Channel pruning is performed on the LSTM layer, and the number of hidden layer units is reduced from 256 to 128; (3) Knowledge distillation: The teacher model outputs a pose angle heatmap as a soft label to guide student model training. The loss function is: in As task weight, This represents the KL divergence distillation loss.
3. The fall risk assessment and dynamic prevention intervention system for the elderly according to claim 1, characterized in that: The privacy computing unit employs a federated learning framework to perform data anonymization and encrypted aggregation at the edge. The specific operations are as follows: (1) Data anonymization: k-anonymization is used on the local training data to ensure that at least 5 records are indistinguishable; (2) Encryption Aggregation: The Paillier homomorphic encryption algorithm is used to aggregate model parameters. The aggregation process satisfies: in, For encrypted parameters, The modulus of the public key; (3) Differential privacy: Laplace noise is injected during global model update. Noise scale Privacy Budget .
4. The fall risk assessment and dynamic prevention intervention system for the elderly according to claim 1, characterized in that: The protection execution unit triggers the MEMS micro-jet airbag for physical buffering protection according to the hierarchical protection strategy. The specific operation is as follows: (1) Gas generation mechanism: A solid sodium azide micro combustion and explosion device is used, which generates gas within 20ms after triggering. Nitrogen; (2) Multi-zone control: Based on the fall direction prediction results, the inflation pressure of the six airbag zones in the waist and hip is independently controlled to 20-30 kPa; (3) Linked braking: The airbag triggers synchronously to send a PWM braking signal to the walking aid, outputting reverse torque. .
5. The fall risk assessment and dynamic prevention intervention system for the elderly according to claim 1, characterized in that: The system integration and optimization module includes an edge computing unit, a collaborative learning unit, a resource scheduling unit, and a human-computer interaction unit, wherein: The edge computing unit deploys the model via Jetson Xavier NX and supports INT8 quantized inference; The collaborative learning unit is used to perform cross-institutional model updates using differential privacy-preserving federated learning; The resource scheduling unit is used to dynamically allocate computing power and storage resources according to task priority; The human-computer interaction unit is used to provide voice prompts, vibration feedback, and AR interface guidance.
6. The fall risk assessment and dynamic prevention intervention system and method for the elderly according to claim 1, characterized in that: The application scenario module includes an environment adaptation unit, an organization management unit, a public deployment unit, and a personalized configuration unit, wherein: The environment adaptation unit is used to adjust the deployment scheme of the device according to different layouts and needs of homes, nursing homes, and public places; The institutional management unit is used to deploy a multi-node monitoring network, and the central monitoring platform displays the group risk heat map and early warning statistics in real time; The public deployment unit is used to achieve low-cost renovation by utilizing existing surveillance cameras and edge AI boxes to cover key areas such as restrooms and staircases; The personalized configuration unit is used to customize the protection level via a mobile app.
7. A method for assessing and dynamically preventing fall risks in the elderly, based on the fall risk assessment and dynamic prevention intervention system for the elderly as described in any one of claims 1-6, characterized in that: Includes the following steps: S1. Real-time synchronous acquisition of motion posture, gait and environmental data through depth camera, IMU array and millimeter wave radar, and spatiotemporal alignment and noise reduction processing. S2. Multimodal data is weighted and fused using an attention mechanism. Real-time pose estimation is performed based on a lightweight CNN-LSTM network. The relationship between the center of mass and the support surface is calculated using a biomechanical model, and the graded fall probability is output. S3. Dynamically generate protection strategies based on the probability of falling. When the risk level exceeds the threshold, trigger the multi-zone inflation of the MEMS micro-jet airbag and the linkage braking of the walking aid. S4. Aggregate encrypted model parameters through a federated learning framework, inject ε≤0.5 differential privacy noise, and dynamically adjust the risk judgment threshold using PPO reinforcement learning. S5. Based on the environmental characteristics of homes, nursing homes, or public places, adaptively adjust the sensor deployment scheme and protection response strategy, and realize personalized configuration through a mobile APP.