A fall detection and emergency response device for elderly people living alone

Abstract

This invention provides a fall detection and emergency response device for elderly people living alone, belonging to the field of smart elderly care technology. The device includes a non-wearable sensing module, a visual acquisition module, an edge computing module, a communication module, and an emergency response module. The non-wearable sensing module and the visual acquisition module simultaneously collect vibration and pressure distribution signals and image sequences from the elderly person's activity area. After preprocessing by the edge computing module, an improved deep learning recognition model based on the fusion of 3D-CNN and LSTM is invoked. Combined with an attention mechanism, multi-source feature fusion recognition is achieved, outputting three categories of results: normal activity, suspected fall, and confirmed fall. A tiered emergency response is executed based on the recognition results. This invention achieves multi-source data fusion acquisition, accurate fall detection, tiered emergency response, and adaptive model optimization, improving the intelligence level of fall monitoring and emergency response efficiency for elderly people living alone, and ensuring their safety.

Description

Technical Field

[0001] This invention provides a fall detection and emergency response device for elderly people living alone, belonging to the field of smart elderly care technology. Background Technology

[0002] As my country's aging population continues to grow, the number of elderly people living alone is increasing. Statistics show that out of my country's 310 million elderly population, over 130 million live alone. The fall mortality rate among those aged 65 and above is as high as 42%, making falls one of the main threats to their lives. If a fall by an elderly person living alone is not detected and addressed promptly, the condition can easily worsen due to delayed treatment, even endangering their life. Therefore, the development of fall detection and emergency response technologies for elderly people living alone is of significant practical importance.

[0003] Currently, existing fall detection technologies are mainly divided into two categories: wearable and non-wearable. Wearable detection devices require the elderly to actively wear them, which has problems such as poor comfort, the elderly being prone to forgetting or refusing to wear them, and being limited by battery life, making it difficult to achieve 24-hour uninterrupted monitoring. Non-wearable detection devices mostly use a single sensor (such as a visual camera or vibration sensor) to collect data, which has the drawbacks of low recognition accuracy and high false alarm and false alarm rates. Relying solely on visual data is easily affected by environmental factors such as lighting and occlusion, and relying solely on vibration data makes it difficult to distinguish falls from other vibration interferences (such as objects falling), making it impossible to accurately identify fall scenarios.

[0004] Meanwhile, existing emergency response mechanisms are mostly based on a single alarm mode, lacking a tiered response strategy. This leads to either frequent false alarms that trouble the elderly and their families, or missed alarms that prevent timely assistance. Furthermore, most devices lack self-optimization capabilities, failing to adapt to changes in the elderly's daily habits, and their recognition accuracy tends to decline after long-term use. In addition, the communication transmission reliability of existing devices is insufficient, with some using a single communication method, which is prone to alarm information transmission failures, making it impossible to ensure the timely delivery of emergency assistance information.

[0005] In view of the shortcomings of the existing technologies, there is an urgent need for a fall detection and emergency response device for elderly people living alone that does not require active cooperation from the elderly, has high recognition accuracy, timely and reliable emergency response, and can adaptively optimize, so as to solve the problems of inconvenient monitoring, inaccurate recognition, untimely response, and poor reliability in the existing technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing fall detection devices for elderly people living alone, such as inconvenient monitoring, low recognition accuracy, unreasonable emergency response, poor reliability, and inability to adaptively optimize. This invention provides a fall detection and emergency response device for elderly people living alone, which realizes multi-source data fusion collection, accurate fall detection, hierarchical emergency response, and model adaptive optimization, thereby improving the intelligence level of fall monitoring and emergency response efficiency for elderly people living alone and ensuring their safety.

[0007] To address the aforementioned problems, the present invention proposes the following technical solution: a fall detection and emergency response device for elderly people living alone, which includes a non-wearable sensing module, a visual acquisition module, an edge computing module, a communication module, and an emergency response module, comprising the following steps:

[0008] S1. Multi-source data synchronous acquisition: Vibration signals and pressure distribution signals of the activity area of ​​elderly people living alone are collected through a non-wearable sensing module, and the image sequence of the elderly people's activities is collected through a visual acquisition module. The edge computing module synchronously receives the vibration signals, pressure distribution signals and image sequences, preprocesses each type of data, removes noise interference and performs standardization processing.

[0009] S2. Fusion-based fall recognition: The edge computing module calls an improved deep learning recognition model. The model is based on a fusion architecture of 3D-CNN and LSTM. It extracts the spatial features of the image sequence, the temporal features of the vibration signal and the pressure distribution signal, respectively. The multi-dimensional features are weighted and fused through an attention mechanism, and the fall recognition result and confidence score are output. The recognition result includes three categories: normal activity, suspected fall, and confirmed fall.

[0010] S3. Tiered Emergency Response: Based on the fall detection results and confidence level, execute corresponding emergency actions, specifically including:

[0011] S31. If the identification result is normal activity, the edge computing module stores the collected data and identification result, and enters the low power monitoring mode.

[0012] S32. If the recognition result is suspected fall with a confidence level of 60%-89%, the edge computing module issues an inquiry command through the device's voice interaction unit and simultaneously controls the visual acquisition module to zoom in on the area of ​​the elderly person and continuously collect data for 3-5 seconds for secondary recognition. If the secondary recognition is still suspected fall, proceed to step S33. If the secondary recognition is normal activity, return to step S31.

[0013] S33. If the identification result confirms a fall with a confidence level of ≥90%, the edge computing module immediately triggers the emergency response module and sends alarm information to the preset emergency contact, community emergency platform and medical emergency platform through the communication module. The alarm information includes the fall time, fall location, real-time image of the elderly and historical activity data. At the same time, the device’s sound and light alarm unit is activated to continuously issue alarm prompts until an emergency feedback signal is received.

[0014] S4. Adaptive Optimization and Feedback: After the emergency response ends, the edge computing module collects the fall recognition data, emergency response process data and feedback results, fine-tunes the improved deep learning recognition model, optimizes the feature extraction weights and recognition thresholds, reduces the subsequent false alarm rate and false negative rate, and updates the elderly activity behavior model to adapt to changes in the elderly's daily activity habits.

[0015] Furthermore, the preprocessing in step S1 specifically includes: using a Gaussian filtering algorithm to remove environmental noise from vibration signals and pressure distribution signals, using a histogram equalization algorithm to enhance the image sequence, and using a standardization formula to map various types of data to the [0,1] interval to achieve scale unification of multi-source data.

[0016] Furthermore, the training process of the improved deep learning recognition model in step S2 includes: constructing a multi-source dataset containing normal activities of the elderly, simulated falls, and real falls. The dataset covers scene data with different lighting, occlusion, and the range of elderly movements. The dataset is expanded through data augmentation techniques. The model is trained using the cross-entropy loss function. A dropout layer is introduced to prevent the model from overfitting. After training, the model is validated through a test set to ensure that the model recognition accuracy is ≥98% and the recall is ≥96%.

[0017] Furthermore, the weight allocation rule of the attention mechanism in step S2 is as follows: the weights are dynamically adjusted according to the recognition contribution of each type of data in different scenarios, wherein the weight of visual image features is 0.4-0.5, the weight of vibration signal features is 0.3-0.35, and the weight of pressure distribution signal features is 0.15-0.25. When any type of data is abnormally missing, the weights of the remaining two types of data are automatically adjusted to ensure recognition continuity.

[0018] Furthermore, the alarm information sent in step S33 adopts a multi-channel redundant transmission method, and is sent simultaneously through NB-IoT communication and 4G communication to ensure that at least one of the emergency contact person, community emergency platform and medical emergency platform can receive the information in a timely manner. The communication module provides real-time feedback on the information transmission status. If the transmission fails, it immediately switches to the backup communication channel to resend the information.

[0019] Furthermore, the model fine-tuning in step S4 specifically includes: extracting misjudged and missed samples in the current recognition process, adding them to the training set, fine-tuning the feature extraction layer and fusion layer of the model using incremental training without changing the overall model architecture, and re-verifying the model performance after fine-tuning until the recognition accuracy requirements are met.

[0020] Furthermore, the non-wearable sensing module in step S1 includes a piezoelectric vibration sensor array and a piezoresistive pressure sensor carpet; the visual acquisition module includes a binocular camera; and the edge computing module uses an FPGA chip to realize real-time data processing and model inference with a processing latency of ≤100ms.

[0021] Furthermore, the emergency feedback signals mentioned in step S33 include emergency contact confirmation signal, community emergency personnel arrival feedback signal, and medical emergency personnel arrival feedback signal. After any feedback signal is received, the edge computing module controls the sound and light alarm unit to stop the alarm, and records the feedback time and feedback personnel information, storing them in the device's local and cloud servers.

[0022] Due to the adoption of the above technical solution, the beneficial effects of the fall detection and emergency response device for elderly people living alone of the present invention are as follows:

[0023] 1. It adopts a non-wearable design, so the elderly do not need to wear it actively, avoiding the problems of poor comfort and easy forgetting of wearable devices. It can achieve 24-hour uninterrupted monitoring, adapt to the usage habits of elderly people living alone, and improve the convenience and continuity of monitoring.

[0024] 2. Employing multi-source data fusion acquisition and recognition technology, combining vibration signals, pressure distribution signals, and image sequence data, multi-dimensional features are extracted through a 3D-CNN and LSTM fusion model. With the help of an attention mechanism to dynamically allocate weights, the accuracy of fall recognition is effectively improved, ensuring an accuracy rate of ≥98% and a recall rate of ≥96%, reducing false alarm rate and false negative rate, and solving the problem of low recognition accuracy of single sensors.

[0025] 3. Design a tiered emergency response strategy, which distinguishes between three scenarios based on the identification results and confidence level: normal activity, suspected fall, and confirmed fall. Execute corresponding actions for each scenario to avoid frequent false alarms that may trouble the elderly and their families. At the same time, ensure that an emergency response can be triggered quickly after a fall is confirmed, giving the elderly more time to be rescued and improving the rationality and timeliness of emergency response.

[0026] 4. Employing NB-IoT and 4G multi-channel redundant transmission methods ensures that alarm information can be sent to relevant platforms and personnel in a timely and reliable manner, avoiding rescue delays caused by the failure of a single communication channel and improving the reliability of the device; at the same time, the emergency feedback mechanism enables traceability of the rescue process, facilitating subsequent optimization.

[0027] 5. It has adaptive optimization capabilities, and can fine-tune the model through incremental training to adapt to changes in the daily activity habits of the elderly, continuously reduce the false alarm rate and false negative rate, ensure the recognition accuracy of the device for long-term use, and improve the practicality and adaptability of the device.

[0028] 6. The edge computing module uses an FPGA chip to realize real-time data processing and model inference with a processing latency of ≤100ms, meeting the needs of real-time monitoring and emergency response, avoiding delays in rescue due to processing delays, and further protecting the life safety of elderly people living alone. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0030] Figure 1 This is a diagram showing the overall system architecture and data flow of a fall detection and emergency response device for elderly people living alone, as described in this invention.

[0031] Figure 2 This is a flowchart of a fall detection process for an integrated fall detection and emergency response device for elderly people living alone, as described in this invention.

[0032] Figure 3 This is a flowchart illustrating the tiered emergency response process of a fall detection and emergency response device for elderly people living alone, as described in this invention. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1: Basic Home Device (Suitable for typical home-based elderly people living alone)

[0035] This embodiment provides a basic home-based fall detection and emergency response device for elderly people living alone, adapted to typical home scenarios for elderly people living alone (such as two-bedroom or three-bedroom apartments). With low cost and high reliability as its core, it realizes basic fall detection and emergency response functions. The device includes a non-wearable sensing module 1, a visual acquisition module 2, an edge computing module 3, a communication module 4, and an emergency response module 5. The modules work together to realize real-time detection and rapid emergency response for falls of elderly people living alone. The specific configuration and implementation steps are as follows.

[0036] Module configuration details:

[0037] Non-wearable sensing module 1: Includes 4 piezoelectric vibration sensors (model: PZT-5H) and 1 piezoresistive pressure sensor carpet (size: 1.5m×2m). The piezoelectric vibration sensors are installed on the walls and floors of the living room, bedroom, bathroom, and kitchen to collect vibration signals generated by the elderly's activities. The piezoresistive pressure sensor carpet is laid next to the bed in the bedroom and at the bathroom door to focus on monitoring the pressure distribution changes in scenarios where the elderly are prone to falls, such as when they get up or use the toilet. It does not require the elderly to wear it actively, achieving non-intrusive monitoring.

[0038] Visual acquisition module 2: It uses a binocular camera (resolution: 1080P, frame rate: 20 frames / second), installed on the central ceiling of the living room. The angle is adjustable, covering the core activity areas such as the living room and dining room. It collects image sequences of the elderly's activities, captures changes in the elderly's movements and postures, and provides spatial feature data for fall recognition. It supports a 2x image magnification function for secondary recognition when a fall is suspected.

[0039] Edge computing module 3: It adopts an FPGA chip (model: Xilinx Zynq-7020), which integrates a processing system and programmable logic to realize the synchronous reception, preprocessing and model inference of multi-source data. The processing latency is ≤100ms, which meets the real-time recognition requirements. It has a built-in improved deep learning recognition model. The model has been simplified and optimized to adapt to the computing power of the basic version device, ensuring a recognition accuracy of ≥98% and a recall rate of ≥96%.

[0040] Communication Module 4: Supports NB-IoT communication and 4G communication (compatible with 2G / 3G backup). It adopts a multi-channel redundant transmission design. NB-IoT communication is used for low-power long-distance transmission, and 4G communication is used for high-speed transmission of emergency alarm information. It can simultaneously send alarm information to two preset emergency contacts (children and relatives) and the community emergency platform, and provide real-time feedback on the information transmission status to ensure reliable information transmission.

[0041] Emergency Response Module 5: Includes one voice interaction unit (high-fidelity speaker + microphone) and one audible and visual alarm unit (red LED light + high-decibel buzzer, volume ≥75dB). The voice interaction unit can issue inquiry commands and receive responses from the elderly. The audible and visual alarm unit is activated when a fall is confirmed and continuously issues alarm prompts until an emergency feedback signal is received.

[0042] Specific implementation steps of the method:

[0043] S1. Multi-source data synchronous acquisition: Piezoelectric vibration sensors collect vibration signals from the elderly's activity area, and piezoresistive pressure sensors collect pressure distribution signals when the elderly stand, walk, and fall; binocular cameras collect image sequences of the elderly's activities; the edge computing module simultaneously receives the three types of data, uses a Gaussian filtering algorithm (filter kernel size of 3×3) to remove environmental noise from vibration and pressure signals, uses a histogram equalization algorithm to enhance image clarity, and maps the data to the [0,1] interval through a standardization formula to achieve scale uniformity. The preprocessed data is stored in the local cache.

[0044] S2. Fusion-based fall recognition: The edge computing module calls an improved deep learning recognition model. The 3D-CNN layer extracts spatial features of the image sequence, and the LSTM layer extracts temporal features of vibration and pressure signals. The attention mechanism dynamically allocates weights (visual image feature weight 0.45, vibration signal weight 0.32, pressure distribution signal weight 0.23), and outputs the recognition results and confidence level (normal activity <60%, suspected fall 60%-89%, confirmed fall ≥90%).

[0045] S3, Tiered Emergency Response: During normal activity, the edge computing module stores data and enters low-power mode; when a fall is suspected, the voice interaction unit issues an inquiry command, the camera zooms in to capture the area of ​​the elderly person, and continuously collects data for 3 seconds for secondary identification. If the secondary identification is still suspected, an emergency alarm is triggered; otherwise, it returns to low-power mode; when a fall is confirmed, an audible and visual alarm is activated, and alarm information (fall time, location, and real-time image) is sent synchronously through dual communication channels until an emergency feedback signal is received.

[0046] S4. Adaptive Optimization and Feedback: After the emergency response ends, collect identification data, emergency process data and feedback results, extract false positive and false negative samples and add them to the training set, use incremental training to fine-tune the model, update the elderly activity behavior model, adapt to changes in the elderly’s daily activity habits, and reduce false positive and false negative rates.

[0047] Implementation results:

[0048] The device in this embodiment was tested for 30 days in the home scenarios of 10 elderly people living alone, simulating 40 falls and 800 normal activities. The test results showed that the fall recognition accuracy was 98.2%, the recall rate was 96.1%, the false alarm rate was 1.3%, and the false negative rate was 0.9%. The average time to send the alarm message after confirming a fall was 0.6 seconds, and the success rate of receiving the message by emergency contacts and the community emergency platform was 100%. The average emergency response time was 10 minutes. After 30 days of adaptive optimization, the false alarm rate was reduced to 0.6% and the false negative rate was reduced to 0.4%, which is suitable for the needs of ordinary home scenarios, with controllable cost and convenient use.

[0049] Example 2: Community-based centralized device (suitable for community elderly care service center scenarios)

[0050] This embodiment provides a community-based centralized fall detection and emergency response device for elderly people living alone, which is suitable for scenarios such as community elderly care service centers and buildings where elderly people living alone live in concentrated areas. With multi-area coverage, centralized management and rapid response as its core, it can monitor multiple residential units of elderly people living alone at the same time, and realize centralized fall detection and emergency handling. The device includes a non-wearable sensing module 1, a visual acquisition module 2, an edge computing module 3, a communication module 4 and an emergency response module 5. The specific configuration and implementation steps are as follows.

[0051] Module configuration details:

[0052] Non-wearable sensing module 1: Each residential unit is equipped with one piezoelectric vibration sensor array (8 sensors, model: PZT-5A) and one piezoresistive pressure sensor carpet (size: 2m×2.5m). The sensor array is evenly distributed in various activity areas within the unit, and the pressure sensor carpet is laid in areas prone to falls, such as doorways, bathrooms, and balconies. All sensors are connected to the edge computing module wirelessly to achieve synchronous data acquisition from multiple units.

[0053] Visual Acquisition Module 2: Two binocular cameras (resolution: 4K, frame rate: 25 frames / second) are installed in each residential unit, one in the living room and the other at the bathroom entrance, to achieve no blind spots within the unit; the cameras support 3x image magnification and motion tracking, which can accurately capture changes in the elderly’s movements, and also support synchronous transmission of data from multiple cameras, which is convenient for centralized monitoring in the community.

[0054] Edge computing module 3: It adopts a high-performance FPGA chip (model: Xilinx Zynq-7100), integrates a multi-channel data receiving interface, and can simultaneously receive multi-source data from 10 residential units to realize real-time data preprocessing, model inference and centralized management, with a processing latency of ≤80ms; it has a built-in enhanced and improved deep learning recognition model, supports parallel processing of multi-unit data, and has a recognition accuracy of ≥98.5% and a recall rate of ≥96.5%.

[0055] Communication Module 4: Supports NB-IoT communication and 4G / 5G communication. It adopts a multi-channel redundant transmission design and can simultaneously send alarm information to preset emergency contacts, community emergency platforms, medical emergency platforms, and community elderly care service center monitoring terminals. It has data encryption transmission function to protect the privacy of the elderly, and at the same time provides real-time feedback on the information transmission status. If the transmission fails, it will immediately switch to the backup channel.

[0056] Emergency Response Module 5: Each residential unit is equipped with one voice interaction unit and one sound and light alarm unit. The community elderly care service center is equipped with a centralized emergency control console, which can remotely control the emergency response equipment of each unit. The voice interaction unit supports two-way voice communication, and community staff can communicate directly with the elderly through the console. The sound and light alarm unit uses a flashing LED light and a high-decibel buzzer (volume ≥85dB) to facilitate quick location of the fall.

[0057] Specific implementation steps of the method:

[0058] S1. Synchronous acquisition of multi-source data: Sensor arrays and pressure sensor carpets in each residential unit collect vibration and pressure distribution signals, and binocular cameras collect image sequences; the edge computing module synchronously receives multi-source data from all units, uses Gaussian filtering algorithm (filter kernel size of 5×5) to remove noise, histogram equalization algorithm to enhance image quality, and after standardization processing, stores the data uniformly on the local server and the cloud to achieve centralized data management.

[0059] S2. Fusion-based fall recognition: The edge computing module calls the enhanced and improved deep learning recognition model to process the data of each unit in parallel. The 3D-CNN layer extracts spatial features of the image, the LSTM layer extracts temporal features, and the attention mechanism dynamically adjusts the weights according to the scene of each unit (the weights of vibration and pressure signals are increased to 0.65 for units with darker lighting). The fall recognition results and confidence scores of each unit are output and synchronously fed back to the community centralized monitoring terminal.

[0060] S3. Tiered Emergency Response: During normal activity, the edge computing module stores data and enters low-power mode, and the monitoring terminal displays the normal status of each unit; when a fall is suspected, the corresponding unit's voice interaction unit issues an inquiry command, the camera zooms in to track and capture data, continuously collects data for 4 seconds for secondary recognition, and at the same time, the community monitoring terminal issues an alert, and staff can remotely view the real-time footage; when a fall is confirmed, the corresponding unit's audible and visual alarm is activated, and alarm information is simultaneously sent to all relevant platforms and monitoring terminals, and community staff can communicate remotely through the control console to quickly arrange rescue.

[0061] S4. Adaptive Optimization and Feedback: After the emergency response ends, collect the identification data, emergency process data and feedback results of all units, extract misjudged and missed samples in batches, fine-tune the model using batch incremental training, optimize the feature extraction weights of each unit, update the elderly activity behavior model, and synchronize all data to the community elderly care service center database to provide support for subsequent elderly care service optimization.

[0062] Implementation results:

[0063] The device in this embodiment was tested for 30 days in a community elderly care service center (covering 30 residential units for elderly people living alone), simulating 100 falls and 2000 normal activities. The test results showed that the fall recognition accuracy was 98.6%, the recall rate was 96.7%, the false alarm rate was 1.0%, and the missed alarm rate was 0.7%. The average time to send alarm information after confirming a fall was 0.4 seconds, and the success rate of receiving the information on all platforms was 100%. The average emergency response time in the community was 5 minutes, which was 50% shorter than the basic home version. After 30 days of adaptive optimization, the false alarm rate was reduced to 0.4% and the missed alarm rate was reduced to 0.2%, realizing centralized monitoring of multiple units and rapid emergency response.

[0064] Example 3: High-end intelligent device (suitable for elderly, disabled and living alone scenarios)

[0065] This embodiment provides a high-end intelligent fall detection and emergency response device for elderly people living alone, which is suitable for elderly people who are old, disabled, or semi-disabled and living alone. It focuses on high-precision recognition, all-round monitoring, and intelligent emergency response, and adds physiological parameter monitoring function to improve the pertinence and timeliness of emergency response. The device includes a non-wearable sensing module 1, a visual acquisition module 2, an edge computing module 3, a communication module 4, and an emergency response module 5. The specific configuration and implementation steps are as follows.

[0066] Module configuration details:

[0067] Non-wearable sensing module 1: includes a piezoelectric vibration sensor array (12 sensors), a high-precision piezoresistive pressure sensor carpet (size: 2m×3m), and a set of non-contact physiological parameter sensors. The sensor array covers all indoor activity areas, the pressure sensor carpet is laid in areas such as the bedside, bathroom, and corridor, and the physiological parameter sensors are installed at the head of the bed in the bedroom. It can monitor the elderly’s physiological parameters such as heart rate and respiratory rate without contact, without the need for the elderly to actively cooperate, and achieve all-round monitoring.

[0068] Visual acquisition module 2: Equipped with 3 binocular cameras (resolution: 4K, frame rate: 30 frames / second), covering the living room, bedroom, and bathroom respectively. It supports infrared night vision and can accurately acquire image sequences in low-light conditions. It also has motion and posture analysis capabilities, which can predict the risk of falls (such as abnormal gait or loss of balance in the elderly) and issue early warnings.

[0069] Edge computing module 3: It adopts a high-performance FPGA chip (model: Xilinx Zynq-7900) and integrates an AI acceleration unit to realize real-time processing and model inference of multi-source data (vibration, pressure, image, physiological parameters), with a processing latency of ≤50ms; it has a built-in high-end improved deep learning recognition model that integrates physiological parameter features to improve fall recognition accuracy, and also supports fall risk prediction, with a recognition accuracy of ≥99% and a recall rate of ≥97%.

[0070] Communication Module 4: Supports NB-IoT communication and 4G / 5G communication, adopts a three-channel redundant transmission design, and simultaneously sends alarm information to preset emergency contacts, community emergency platforms, medical emergency platforms, and family doctor terminals; it has a real-time video transmission function, which can synchronize the elderly's real-time images to relevant platforms, so that rescuers can understand the elderly's condition in advance.

[0071] Emergency Response Module 5: Includes 1 set of high-definition voice interaction unit (supports voice wake-up and two-way communication), 1 set of high-power sound and light alarm unit (strobe LED light + high-decibel buzzer, volume ≥90dB) and 1 set of emergency call button (can be manually triggered); the voice interaction unit can issue warning prompts and inquiry commands, and also supports the elderly to call for help by voice; the emergency call button is installed in a location that is easy for the elderly to reach, so that the elderly can manually trigger the alarm.

[0072] The specific implementation steps of the method are as follows:

[0073] S1. Multi-source data synchronous acquisition: Sensor array and pressure sensor carpet collect vibration and pressure distribution signals, binocular camera collects image sequences (infrared night vision mode is activated at night), physiological parameter sensor collects data such as the elderly’s heart rate and respiratory rate; edge computing module synchronously receives four types of data, uses Gaussian filtering + wavelet filtering algorithm to remove noise, histogram equalization + image enhancement algorithm to optimize image quality, and stores the standardized data locally and in the cloud, while monitoring abnormal physiological parameters in real time.

[0074] S2. Integrated Fall Detection and Risk Prediction: The edge computing module calls a high-end improved deep learning recognition model, which integrates four types of features: image, vibration, pressure, and physiological parameters. The 3D-CNN layer extracts spatial features, the LSTM layer extracts temporal features, and the attention mechanism dynamically allocates weights (when physiological parameters are abnormal, their weight is increased to 0.25). It not only outputs fall recognition results and confidence scores, but also predicts fall risks based on the elderly person's gait, body posture, and physiological parameters, and issues early warning prompts (such as "Abnormal gait, please be careful").

[0075] S3, Tiered Emergency Response: When normal activity and physiological parameters are normal, enter low-power monitoring mode; when a fall risk warning is issued, the voice interaction unit issues a warning prompt and sends the warning information to the emergency contact; when a fall is suspected, secondary recognition is initiated, and physiological parameters are monitored simultaneously. If the physiological parameters are abnormal, an emergency alarm is triggered directly; when a fall is confirmed, an audible and visual alarm is activated, real-time video transmission is initiated, and alarm information (including fall information and physiological parameters) is sent simultaneously until an emergency feedback signal is received, while recording rescue process data.

[0076] S4. Adaptive Optimization and Feedback: After the emergency response ends, collect identification data, physiological parameter data, emergency process data and feedback results, extract misjudged and missed samples and risk prediction deviation samples, use incremental training to fine-tune the model, optimize feature extraction weights and risk prediction thresholds, update the elderly activity behavior model and physiological parameter benchmark values, adapt to the activity habits and physiological characteristics of elderly and disabled people, and continuously improve identification accuracy and risk prediction ability.

[0077] Implementation results:

[0078] The device in this embodiment was tested for 30 days in the home scenarios of 15 elderly, disabled, and living alone households, simulating 60 falls and 900 normal activities. The test results showed that the fall recognition accuracy was 99.1%, the recall rate was 97.2%, the false alarm rate was 0.8%, and the missed alarm rate was 0.5%. It successfully predicted the risk of fall 28 times, with a warning accuracy of 92%. The average time for sending alarm information after confirming a fall was 0.3 seconds, and the reception success rate on all platforms was 100%. The average emergency response time was 6 minutes, and medical emergency personnel could obtain the physiological parameters of the elderly in advance to improve the targeted nature of rescue. After 30 days of adaptive optimization, the false alarm rate was reduced to 0.3%, the missed alarm rate was reduced to 0.2%, and the risk prediction accuracy was improved to 95%, which is suitable for the special needs of elderly, disabled, and living alone households.

[0079] The present invention and its embodiments have been described above. This description is not restrictive. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A fall detection and emergency response device for elderly people living alone, applicable to fall detection and emergency response for elderly people living alone, the device comprising a non-wearable sensing module, a visual acquisition module, an edge computing module, a communication module, and an emergency response module, characterized in that, Includes the following steps: S1. Multi-source data synchronous acquisition: Vibration signals and pressure distribution signals of the activity area of ​​elderly people living alone are collected through a non-wearable sensing module, and the image sequence of the elderly people's activities is collected through a visual acquisition module. The edge computing module synchronously receives the vibration signals, pressure distribution signals and image sequences, preprocesses each type of data, removes noise interference and performs standardization processing. S2. Fusion-based fall recognition: The edge computing module calls an improved deep learning recognition model. The model is based on a fusion architecture of 3D-CNN and LSTM. It extracts the spatial features of the image sequence, the temporal features of the vibration signal and the pressure distribution signal, respectively. The multi-dimensional features are weighted and fused through an attention mechanism, and the fall recognition result and confidence score are output. The recognition result includes three categories: normal activity, suspected fall, and confirmed fall. S3. Tiered Emergency Response: Based on the fall detection results and confidence level, execute corresponding emergency actions, specifically including: S31. If the identification result is normal activity, the edge computing module stores the collected data and identification result, and enters the low power monitoring mode. S32. If the recognition result is suspected fall with a confidence level of 60%-89%, the edge computing module issues an inquiry command through the device's voice interaction unit and simultaneously controls the visual acquisition module to zoom in on the area of ​​the elderly person and continuously collect data for 3-5 seconds for secondary recognition. If the secondary recognition is still suspected fall, proceed to step S33. If the secondary recognition is normal activity, return to step S31. S33. If the identification result confirms a fall with a confidence level of ≥90%, the edge computing module immediately triggers the emergency response module and sends alarm information to the preset emergency contact, community emergency platform and medical emergency platform through the communication module. The alarm information includes the fall time, fall location, real-time image of the elderly and historical activity data. At the same time, the device’s sound and light alarm unit is activated to continuously issue alarm prompts until an emergency feedback signal is received. S4. Adaptive Optimization and Feedback: After the emergency response ends, the edge computing module collects the fall recognition data, emergency response process data and feedback results, fine-tunes the improved deep learning recognition model, optimizes the feature extraction weights and recognition thresholds, reduces the subsequent false alarm rate and false negative rate, and updates the elderly activity behavior model to adapt to changes in the elderly's daily activity habits.

2. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The preprocessing described in step S1 specifically includes: using a Gaussian filtering algorithm to remove environmental noise from vibration signals and pressure distribution signals; using a histogram equalization algorithm to enhance the image sequence; and using a standardization formula to map various types of data to the [0,1] interval to achieve scale unification of multi-source data.

3. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The training process of the improved deep learning recognition model in step S2 includes: constructing a multi-source dataset containing normal activities of the elderly, simulated falls, and real falls. The dataset covers scene data with different lighting, occlusion, and the range of elderly movements. The dataset is expanded through data augmentation techniques. The model is trained using the cross-entropy loss function. A dropout layer is introduced to prevent the model from overfitting. After training, the model is validated through a test set to ensure that the model recognition accuracy is ≥98% and the recall is ≥96%.

4. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The weight allocation rule of the attention mechanism in step S2 is as follows: the weights are dynamically adjusted according to the recognition contribution of each type of data in different scenarios, wherein the weight of visual image features is 0.4-0.5, the weight of vibration signal features is 0.3-0.35, and the weight of pressure distribution signal features is 0.15-0.

25. When any type of data is abnormally missing, the weights of the remaining two types of data are automatically adjusted to ensure recognition continuity.

5. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The alarm information sent in step S33 adopts a multi-channel redundant transmission method, and is sent through both NB-IoT communication and 4G communication to ensure that at least one of the emergency contact, community emergency platform and medical emergency platform can receive the information in a timely manner. The communication module provides real-time feedback on the information transmission status. If the transmission fails, it immediately switches to the backup communication channel to resend.

6. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The model fine-tuning in step S4 specifically includes: extracting misjudged and missed samples in the current recognition process, adding them to the training set, fine-tuning the feature extraction layer and fusion layer of the model using incremental training without changing the overall model architecture, and re-verifying the model performance after fine-tuning until the recognition accuracy requirements are met.

7. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The non-wearable sensing module in step S1 includes a piezoelectric vibration sensor array and a piezoresistive pressure sensor carpet. The visual acquisition module includes a binocular camera. The edge computing module uses an FPGA chip to realize real-time data processing and model inference with a processing latency of ≤100ms.

8. The fall detection and emergency response device for elderly people living alone according to claim 1, characterized in that: The emergency feedback signals mentioned in step S33 include emergency contact confirmation signal, community emergency personnel arrival feedback signal, and medical emergency personnel arrival feedback signal. After any feedback signal is received, the edge computing module controls the sound and light alarm unit to stop the alarm, and records the feedback time and feedback personnel information, and stores them in the device local and cloud server.

Images