A future community management system and method based on the Internet of Things

Abstract

The application discloses a future community management system and method based on Internet of Things, and relates to the field of intelligent management.The system is composed of a plurality of function modules, including: an identity recognition module, which acquires facial images and audio data; facial recognition and age interval judgment are performed according to the facial images; an audio data extraction module, which utilizes an attention mechanism to fuse time-frequency dual-channel features, constructs anti-interference footstep sound features, and obtains footstep data; audio energy values, zero-crossing rates and moving speeds are extracted from the footstep data; footstep intervals are calculated according to the zero-crossing rates; feature vectors are synthesized based on the audio energy values and the age intervals, and are used as inputs of a model, are divided into three groups of sets, and are randomly sampled; a plurality of points in a hyperparameter space are selected for evaluation; a supervised learning model is used to judge the physical state of the personnel entering and exiting; the physical state, the moving speed and the footstep interval are acquired.

Description

Technical Field

[0001] This invention relates to the field of intelligent management, specifically to a future community management system and method based on the Internet of Things. Background Technology

[0002] The origins of modern property management systems can be traced back to the Industrial Revolution, when rapid urbanization brought about a surge in housing demand and management challenges. To address these issues, the earliest rent collection and facility maintenance processes emerged. While simple, these early management systems laid the foundation for later development, helping cities maintain a degree of order and efficiency during rapid expansion. Over time, technological advancements and societal needs have continuously driven the evolution of property management systems, transitioning from initial paper records and manual operations to computerized management and automated systems. The application of information technology has made rent collection, facility maintenance, and resident management more efficient and transparent, significantly improving the overall level of property management. As people's demands for quality of life increase, property management is no longer limited to the maintenance and management of hardware facilities, but pays more attention to the concept of "people-oriented". The community of the future will not only be a place to live, but also a living space full of humanistic care and harmonious coexistence. In order to better serve residents, the future property management system will also integrate more humanized functions, and is committed to creating an efficient, comfortable and sustainable community environment. In the future, with the further integration of technology and humanism, the community will become the core unit of urban sustainable development, not only improving the quality of life of residents, but also contributing to the overall development of the city. This evolution process is not only the result of technological progress, but also a concrete manifestation of society's pursuit of a better life.

[0003] Traditional methods rely on manual input and intervention, failing to achieve true intelligent management. Data sharing between different subsystems is difficult, leading to incomplete information and decision-making difficulties. Monitoring and property management systems often operate independently, unable to share real-time data, impacting overall management efficiency. With the increasing aging population and the growing number of elderly people living alone, existing community management pays little attention to their lives and their inconvenience in accessing and moving around. This is partly due to high labor costs and partly due to a lack of technological support. Therefore, a future community management system is needed to provide effective assistance measures for people with mobility impairments. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a future community management system and method based on the Internet of Things, solving the problems mentioned in the background.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A future community management system based on the Internet of Things includes:

[0009] The identity recognition module acquires facial images and audio data; based on the facial images, it performs facial recognition and age range determination.

[0010] The data processing module extracts audio data, fuses time-frequency dual-path features using an attention mechanism, constructs anti-interference footstep sound features, and obtains footstep data. From the footstep data, it extracts audio energy value, zero-crossing rate, and movement speed; calculates footstep distance based on the zero-crossing rate; synthesizes feature vectors based on audio energy value and age range, and uses them as input to the model, dividing them into three sets for random sampling; selects several points in the hyperparameter space for evaluation; uses a supervised learning model to determine the physical state of people entering and exiting; and calculates the care index by acquiring physical state, movement speed, and footstep distance.

[0011] The dynamic strategy response module compares the care index with the preset care threshold based on a preset care threshold. If the care threshold is less than the care index, it determines that the current person entering or leaving is in good health and continues to identify them; otherwise, it executes the corresponding dynamic strategy.

[0012] Further, facial image and audio data:

[0013] The system uses cameras to capture the frontal facial features of people entering and exiting, and performs facial recognition and age range determination.

[0014] Sound in the stairwell is captured by acoustic sensors and then processed for noise reduction.

[0015] Furthermore, construct anti-interference footstep sound features:

[0016] Time-domain and frequency-domain features are extracted from the audio data to form two independent feature representations. Short-time energy combined with a recurrent neural network is used to capture the temporal dynamic changes of footstep data, while frequency-domain features are generated into a spectrogram through short-time Fourier transform and key frequency components are extracted using an attention module.

[0017] Feature fusion is achieved by introducing an attention mechanism. In the cross-modal cross-attention algorithm, temporal features are used as query vectors and frequency domain features are used as key-value pairs. By calculating the correlation between temporal and frequency domain features, an attention weight matrix is ​​generated, and the contribution ratio of frequency domain features is dynamically adjusted. At the same time, reverse attention is calculated symmetrically with frequency domain features as queries and temporal features as keys, and the bidirectional attention results are concatenated. A multi-footstep suppression strategy is embedded in the fusion process.

[0018] Global information modeling is performed on the fused features. A Transformer encoder is used to capture long-distance dependencies. The features are compressed into fixed-dimensional feature vectors through an attention pooling layer. Footstep features are fused with time and frequency. The classification loss and feature discrimination loss are trained and jointly optimized.

[0019] Furthermore, audio energy value:

[0020] The audio signal is divided into multiple short time periods with a fixed duration, and each time period is called a window. The choice of window length needs to balance time resolution and computational stability. Within each window, the square value of the signal sample is calculated point by point, and the square values ​​of all samples are summed and averaged to obtain the audio energy value Ep.

[0021] Furthermore, the zero-crossing rate:

[0022] Within the time window, count the number of times the signal crosses zero and normalize it to a unit time to calculate the zero-crossing rate ZCR;

[0023] Movement speed:

[0024] After identity authentication is completed, the system enters a timer-ready state, monitoring the footstep sound field in the stairwell. When the first valid acoustic event matching the preset footstep sound characteristics is detected, the timer is started and marked as the initial moment. At this time, the system synchronously records the initial sound pressure level parameters. This serves as a baseline reference value for distance calculation; and historical step frequency data is used to verify the continuity of movement, marking the termination time after confirming the absence of continuous movement. The total distance traveled, D, is obtained through distance inversion calculation, and the speed of movement, v, is calculated.

[0025] Furthermore, assess the physical condition of people entering and exiting:

[0026] The model acquires audio energy values ​​and age ranges by collecting and preprocessing a large dataset of labeled data. The audio energy values ​​and age ranges are combined into a feature vector, which serves as the model input. The dataset is divided into training, validation, and test sets. The training set is further divided into K subsets. K-1 subsets are used for training each time, with the remaining subset used for validation. This process is repeated K times, selecting a different subset of the validation set each time. The average of the K validation results is then calculated as the model performance evaluation metric. A random sampling method is used to randomly select several points in the hyperparameter space for evaluation. Several different hyperparameter combinations are selected from the adjusted hyperparameters and their value ranges. The model is trained using the training set, and its performance is evaluated on the validation set. The supervised learning model outputs a body state score S.

[0027] Furthermore, foot spacing:

[0028] Based on acoustic propagation theory, the foot spacing is obtained according to the ratio of the total moving distance D to the zero-crossing rate.

[0029] Furthermore, the care index:

[0030] The formula, which integrates age correction factors, physical condition scores, and environmental response indicators, is as follows:

[0031]

[0032] In the formula, For the care index; S is the age-adjusted factor; S is the physical condition score. v is the historical average movement speed; v is the current speed; M is the total number of detection points. The measured value at the detection point; The mean; This is the velocity decay sensitivity coefficient; For frequency variation weights; This represents the step spacing.

[0033] Furthermore, whether it exceeds the care index is used to determine:

[0034] When the care threshold is less than the care index, it is determined that the current person entering or leaving is in good physical condition and identification continues.

[0035] When the care threshold is greater than or equal to the care index, it is determined that the physical condition of the person entering or leaving is poor, and the corresponding dynamic strategy is executed.

[0036] A future community management method based on the Internet of Things includes the following steps:

[0037] Step 1: Acquire facial images and audio data; perform facial recognition and age range determination based on the facial images; extract footstep data from the audio data;

[0038] Step 2: Extract audio data, fuse time-frequency dual-path features using an attention mechanism to construct anti-interference footstep sound features, and obtain footstep data; extract audio energy value, zero-crossing rate, and movement speed from the footstep data; calculate footstep distance based on the zero-crossing rate; synthesize feature vectors based on audio energy value and age range, and use them as input to the model, dividing them into three sets for random sampling; select several points in the hyperparameter space for evaluation; use a supervised learning model to determine the physical state of people entering and exiting; calculate the care index by acquiring physical state, movement speed, and footstep distance.

[0039] Step 3: Based on the preset care threshold, compare the care index with the care threshold. If the care threshold is less than the care index, determine that the current person entering or leaving is in good physical condition and continue to identify them; otherwise, if the person is in poor physical condition, execute the corresponding dynamic strategy.

[0040] (III) Beneficial Effects

[0041] This invention provides a future community management system and method based on the Internet of Things, which has the following beneficial effects:

[0042] (1) This invention can adapt to environmental changes through a dynamic fusion mechanism. It obtains facial images and audio data through the identity recognition module and uses facial recognition technology and age range to determine the identity of people entering and leaving. The data processing module uses the attention mechanism to fuse time and frequency dual-path features and constructs anti-interference footstep features, thereby accurately extracting footstep data. It uses an end-to-end training method to optimize feature extraction and anti-interference module in synergy, which is suitable for intelligent identity recognition and detection of abnormal physical conditions.

[0043] (2) This invention realizes data interconnection between different subsystems through Internet of Things technology, and can centrally store and analyze the identity information, footstep characteristics, and movement speed of people entering and exiting; it uses cameras to capture the frontal facial features of people entering and exiting, and performs face recognition and age range judgment to ensure safety while improving user experience; at the same time, it dynamically adjusts the lighting parameters based on environmental perception and personnel behavior data to balance energy saving and safety. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the system flow of the present invention;

[0045] Figure 2 This is a schematic diagram of the overall method of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1:

[0048] Please see Figure 1 This embodiment provides a future community management system based on the Internet of Things. The detection system includes:

[0049] The identity recognition module acquires facial images and audio data; based on the facial images, it performs facial recognition and age range determination.

[0050] Acquisition of facial images:

[0051] Select cameras with high resolution, low light performance, and support for wide dynamic range technology, and install them in appropriate positions above or on both sides of the entrance door of each unit to ensure that the frontal facial features of all people entering and exiting can be captured; the optimal viewing angle should cover the entire entrance area and avoid image distortion caused by angle issues. Adjust parameters in real time according to changes in ambient light to obtain clearer and more stable facial images.

[0052] Facial recognition and age range determination:

[0053] Establish a database containing facial data of all legitimate residents for rapid matching and verification; when the system successfully identifies a legitimate user, it automatically sends a signal to the electronic lock to open the building's main door; for unauthorized or unidentifiable individuals, the system will prompt them to undergo further identity verification or contact the resident of the building to remotely open the building's entrance door.

[0054] Age range determination:

[0055] Facial features are extracted from facial images, and an age recognition algorithm is used to determine the age range of people entering and exiting the area. The determination result is then correlated with the facial recognition result.

[0056] It should be noted that the system strictly adheres to relevant laws and regulations when determining age ranges to ensure that personal privacy is not violated and that information is not disclosed to third parties.

[0057] Acquisition of audio data:

[0058] High-sensitivity acoustic sensors are installed at the corners of the stairs to effectively capture the sound in the stairwell. Automatic gain control and noise reduction technology are used to adapt to sound changes in different time periods and environmental conditions, ensuring the stability and clarity of the audio signal.

[0059] Automatic Gain Control (AGC): Used to adjust the amplitude of the audio signal. When the input signal is weak, AGC increases the amplification factor to make the signal stronger; when the input signal is strong, AGC decreases the amplification factor to prevent signal overload or distortion, keeping the output signal at a relatively stable level and changing the intensity of the input signal.

[0060] For example, in environments such as stairwells, sound intensity may vary depending on the time and type of activity; AGC can receive clear sound signals whether it is a quiet morning or a noisy evening;

[0061] Noise reduction technology: used to analyze the spectrum of audio signals, identify and reduce frequency components that belong to background noise, adjust filter parameters in real time according to changes in ambient noise to suppress background noise in stairwells to the maximum extent, remove background noise from the original audio signal, thereby improving the quality and intelligibility of the useful audio signal;

[0062] The data processing module, based on the audio data, uses an attention mechanism to fuse time-frequency dual-path features to construct anti-interference footstep sound features and obtain footstep data; from the footstep data, it extracts audio energy value, zero-crossing rate, and movement speed; based on the audio energy value and age range, it uses a supervised learning model to determine the physical condition of people entering and exiting; based on the zero-crossing rate, it calculates the footstep distance; and by acquiring physical condition, movement speed, and footstep distance, it calculates the care index.

[0063] Constructing anti-interference footstep sound features:

[0064] By setting a footstep energy threshold, the system detects when footsteps occur. When the energy of the audio signal exceeds the preset footstep energy threshold, footsteps are considered to have occurred. For various footstep sounds in stairwells (such as pet footsteps or footsteps of multiple people walking side-by-side), an attention mechanism is used to fuse time-frequency dual-path features to construct an anti-interference footstep feature profile. This includes feature extraction, attention fusion, anti-interference optimization, and feature modeling, extracting time-domain and frequency-domain features in parallel from the audio data. Time-domain features typically include statistics such as short-time energy, which are combined with convolutional or recurrent neural networks to capture the temporal dynamics of footstep data. Frequency-domain features are generated using short-time Fourier transform or Mel-frequency spectrum analysis, and key frequency components are extracted using convolutional layers or frequency band attention modules. This characterizes the essential properties of footsteps from both time and frequency dimensions.

[0065] An attention mechanism is introduced to achieve feature fusion. In the cross-modal cross-attention algorithm, temporal features are used as query vectors and frequency domain features are used as key-value pairs. By calculating the correlation between temporal and frequency domain features, an attention weight matrix is ​​generated, and the contribution ratio of frequency domain features is dynamically adjusted. At the same time, reverse attention is symmetrically calculated with the frequency domain as the query and the temporal domain as the key. Finally, the bidirectional attention results are concatenated, so that the time and frequency features strengthen complementary information through mutual guidance. A multi-footstep suppression strategy is embedded in the fusion process to enhance anti-interference ability.

[0066] For example, a noise perception branch is added when calculating attention weights to analyze the spectral distribution of environmental noise in real time and apply weight penalties to noise-dominant frequency bands or time periods. At the same time, a contrastive learning training strategy is adopted to construct triplet samples containing clean footsteps, noisy footsteps, and interference sounds. The contrastive loss function forces the model to narrow the feature distance of the same footstep under different noise levels, thus expanding its difference from interference sounds. Meanwhile, various background noises (such as wind noise and human voice) and simulated reverberation effects are mixed in the training set to improve the model's generalization ability to complex environments.

[0067] Global information modeling is performed on the fused features. A Transformer encoder is used to capture long-distance dependencies, or a residual network is used to stack deep nonlinear transformations. Then, the features are compressed into fixed-dimensional feature vectors through attention pooling layers. At this point, the footstep features have integrated the advantages of time and frequency and have noise robustness. During training, the classification loss and feature discrimination loss are jointly optimized so that the features can distinguish the footsteps of different individuals while maintaining intra-class compactness. The model's anti-interference performance is verified on a low signal-to-noise ratio test set by indicators such as recognition accuracy and error rate, realizing the extraction of stable and distinguishable footstep feature representations from noisy frequencies.

[0068] Extracting audio energy values ​​from anti-interference footstep characteristics:

[0069] Footsteps have high energy because of the significant impact they generate.

[0070] Continuous audio data is divided into multiple short time intervals (e.g., 20 milliseconds) of fixed duration, each interval being called a window. The choice of window length requires a trade-off between temporal resolution and computational stability; shorter windows can respond quickly to signal abrupt changes, while longer windows can smooth out random noise. Within each window, the square of each signal sample is calculated point by point to eliminate the influence of negative values. Then, the squares of all samples are summed and averaged to obtain the audio energy value Ep, as shown in the formula:

[0071]

[0072] in, denoted as the sample points of the audio signal; n is the current time point; N is the number of samples within the window; where high energy usually indicates a stronger audio signal; when a person walks, the contact between the foot and the ground generates a large impact force, which will be manifested as a higher amplitude or energy in the audio signal; therefore, the audio signal of footsteps usually has high energy.

[0073] Specifically, it reflects the amplitude of the signal; a higher energy signal means a larger amplitude and sounds "louder," thus distinguishing footsteps from other sounds.

[0074] Zero crossing rate:

[0075] Within a short time window, the number of times the signal crosses zero is counted and normalized to a unit time, using the following formula:

[0076]

[0077] In the formula, ZCR is the zero-crossing rate; For signal The sign function (returns +1 or -1); N is the number of samples in the window;

[0078] For footsteps, since they are mainly low-frequency components, their zero-crossing rate is relatively low. A low zero-crossing rate usually indicates that the signal frequency is low and changes slowly, while a high zero-crossing rate indicates that the signal frequency is high and changes quickly. The zero-crossing rate is the number of times an audio signal crosses the zero point per unit time, i.e., from positive to negative or from negative to positive. It is an indicator of the rate of change of an audio signal and reflects the frequency characteristics of the signal. However, footsteps are mainly composed of low-frequency components, and the changes of these low-frequency sounds are relatively slow, so they do not frequently cross the zero point. Therefore, the zero-crossing rate of footsteps is low.

[0079] By combining energy and zero-crossing rate as two features, footsteps can be effectively distinguished from other background noise, improving the accuracy of footstep detection. Combining MFCC for feature extraction further enhances the system's performance, enabling more accurate footstep recognition and behavior analysis.

[0080] The following is a specific code example for calculating MFCC using Python and the `librosa` library:

[0081] Python

[0082] import librosa

[0083] import numpy as np

[0084] import matplotlib.pyplot as plt

[0085] # Read audio file

[0086] audio_path = 'path_to_audio_file.wav'

[0087] y, sr = librosa.load(audio_path, sr=None)

[0088] # Extracting MFCC features

[0089] mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)

[0090] # Print the shape of the MFCC feature matrix

[0091] print("MFCC shape:", mfccs.shape)

[0092] # Draw MFCC feature map

[0093] plt.figure(figsize=(10, 4))

[0094] librosa.display.specshow(mfccs, sr=sr, x_axis='time')

[0095] plt.colorbar()

[0096] plt.title('MFCC')

[0097] plt.tight_layout()

[0098] plt.show()

[0099] ```

[0100] Movement speed:

[0101] After successful facial recognition authentication, the timing function is activated, and the program enters the preparation state. Timing begins after the first footstep sound is captured, and the sound of approaching footsteps is acquired. In a stable state, the movement speed of an individual is uniform. As the individual moves away from the acoustic sensor, the audio data intensity gradually decreases. Timing ends when the audio data intensity drops to 20% or the sound interval exceeds three seconds, indicating that the movement has stopped; otherwise, timing continues. The distance the individual moved is obtained by comparing the percentage of the audio data intensity at the end of the movement to the initial footstep audio data intensity with the ratio of distance to audio data intensity in the experimental data. Based on the movement time, the movement speed is calculated.

[0102] Once the facial recognition system completes identity authentication, it automatically activates the high-precision timing unit and enters a ready-to-trigger state. Acoustic sensors monitor the footstep sound field in the stairwell in real time. When the first valid acoustic event matching the preset footstep sound characteristics is detected, the timer is immediately started and marked as the initial moment. At this time, the system synchronously records the initial sound pressure level parameters. , as a benchmark reference value for distance calculation;

[0103] Preset footstep characteristics: time domain pulse width 50-200ms, frequency domain energy concentrated at 200-500Hz;

[0104] During the timing duration phase, a multi-dimensional signal processing strategy is employed. Dynamic range compression technology is applied to eliminate the interference of sound pressure fluctuations caused by distance changes on feature extraction. The sound source movement path is constructed based on the time difference of arrival algorithm, and the direction of motion is calculated by combining the topology of the acoustic sensor. Non-target sound sources, such as falling objects and running pets, are filtered out in real time. By loading a pre-trained footsteps classifier, the average sound pressure level of three consecutive sampling windows is obtained. And make a judgment:

[0105] When the average sound pressure level of three consecutive sampling windows satisfy When this occurs, it is determined to be outside the effective monitoring range;

[0106] When the average sound pressure level of three consecutive sampling windows satisfy When the movement is deemed to be within the effective monitoring range, a secondary verification mechanism is triggered, which retrieves historical step frequency data to verify the continuity of the movement. Once it is confirmed that there is no continuous movement, the termination time is marked. ;

[0107] Based on acoustic propagation theory, the total distance traveled, D, is obtained through distance inversion calculation, using the following formula:

[0108]

[0109] In the formula, This is the pre-calibrated reference distance; For the corresponding reference sound pressure level;

[0110] Movement speed is calculated using the following formula:

[0111]

[0112] In the formula, v is the moving speed; dynamic compensation factor For everyday items, the dynamic compensation factor typically ranges from 0.8075 to 1.122. For example, the attenuation coefficient for floor materials is 0.85 for carpets and 1.02 for floor tiles, and the correction value for shoe sole type is 1.1 for sneakers and 0.95 for leather shoes.

[0113] Assess the physical condition of people entering and exiting:

[0114] The energy value of each step is extracted from the audio signal, and the age range of people entering and exiting is obtained after identity authentication, which is used for feature preparation; a large number of labeled datasets are collected, which include different categories of physical condition, such as healthy, fatigued, weak, etc.; the data sources can be obtained through experiments or historical records, as well as from data standard websites.

[0115] The labeled dataset is cleaned to remove noise and outliers, and missing values ​​are handled using mean imputation, interpolation, or other methods. Audio energy values ​​and age range information are combined into a feature vector, which serves as the model input. The dataset is divided into training, validation, and test sets. The training set is further divided into K subsets, with K-1 subsets used for training each time, and the remaining subset used for validation. This process is repeated K times, selecting a different subset of the validation set each time. Finally, the average of the K validation results is calculated as the model performance evaluation metric. A random sampling method is used to randomly select several points in the hyperparameter space for evaluation. The process involves: 1) Selecting adjusted hyperparameters and their possible value ranges; 2) Randomly selecting several points (e.g., 100) from the hyperparameter space to form hyperparameter combinations; 3) Training a model on the training set for each randomly sampled hyperparameter combination and evaluating its performance on the validation set; 4) Selecting the best-performing hyperparameter combination based on the evaluation results on the validation set; 5) Training a model on the training set for each hyperparameter combination and evaluating its performance on the validation set; 6) Recording the evaluation results for each combination, and obtaining the predicted physical state score S for the personnel entering and exiting the area from the probability distribution of the physical state output by the supervised learning model: "healthy," "fatigued," or "weak."

[0116] For example:

[0117] Given an audio dataset, the audio energy value, zero-crossing rate, and age range information have been extracted. The audio energy value for each step is extracted from the audio data. The user's age range is obtained through a facial recognition system. The time interval for each step is calculated, the movement speed is estimated, and the zero-crossing rate for each step is calculated to estimate the stride distance. The audio energy value and age range are used as input features and fed into a pre-trained supervised learning model to obtain a prediction of the person's physical condition, such as "healthy," "fatigued," or "weak."

[0118] Foot spacing:

[0119] A low zero-crossing rate typically indicates a higher proportion of low-frequency components, meaning the footstep sound has a lower frequency and corresponds to a larger stride. The lower the zero-crossing rate, the larger the stride. Based on acoustic propagation theory, the stride distance is obtained by using the ratio of the total distance traveled (D) to the zero-crossing rate. ;

[0120] For example: Given that the total moving distance D is 10 meters and the zero-crossing rate is 20; calculate the average step distance: 10 meters / 20 steps = 0.5 meters;

[0121] Attention Index:

[0122] By integrating multi-dimensional parameters such as age correction factors, physical condition scores, and environmental responses, a quantifiable comprehensive evaluation index is formed. Based on the multi-source data fusion-based care index calculation framework, the formula is as follows:

[0123]

[0124] In the formula, The index is the susceptibility index; i represents the current detection point. The age correction factor, set according to a piecewise function, reflects the principle of social care that requires greater attention to the elderly population, and its value range is:

[0125] ;

[0126] Where age is age; S is the physical condition score; The historical average movement speed is represented by v; the current speed is represented by M; and the total number of detection points is represented by M (20 points for a 200-meter section). The measured value at the detection point; The mean; The velocity decay sensitivity coefficient (typically set to 0.5-1.2) is dynamically adjusted using the Logistic function.

[0127]

[0128] When physical condition score hour, A rapid rise amplifies the effects of speed anomalies; k is an adjustment parameter, a larger k value will... It rises from near 0 to near 1 much faster, thus amplifying the effects of speed anomalies;

[0129] The frequency variation weight (usually set to 0.3-0.6) is negatively correlated with the audio energy value.

[0130]

[0131] The lower the energy value (weak state), the greater the contribution of frequency variation to CI; E represents the audio energy value;

[0132] The dynamic strategy response module determines whether the value exceeds the attention index based on a preset attention threshold.

[0133] Preset attention threshold:

[0134] Based on historical behavioral data analysis, the system collects behavioral data of residents in the community under normal conditions, such as footstep energy value, movement speed, and noise level, and statistically analyzes the distribution patterns. At the same time, the footstep energy threshold is reduced by 20% at night (22:00-6:00), and residents are more sensitive to faint abnormal noises in quiet environments. On rainy or snowy days, the walking speed generally decreases due to slippery ground, so the system automatically lowers the movement speed threshold.

[0135] For example, for walking speed data of healthy adults, the 15th percentile of historical data is taken as the threshold for frailty. Below this value may indicate a risk of falling. Because the risk of falling is high, the walking speed threshold is lowered, such as from 0.8 m / s to 0.5 m / s, while the weight of audio energy value is increased. The lighter footsteps of frail elderly people are associated with medical record data, which indirectly infers their behavioral deviation, i.e., the care threshold.

[0136] Whether it exceeds the care index is the judgment:

[0137] When the care threshold is less than the care index, it is determined that the current person entering or leaving is in good physical condition and identification continues.

[0138] When the care threshold is greater than or equal to the care index, it is determined that the physical condition of the person entering or leaving is poor, and the corresponding dynamic strategy is executed.

[0139] Dynamic strategies include intelligent lighting control, linkage with elevator systems, and voice prompts in stairwells;

[0140] Intelligent lighting control:

[0141] By using environmental perception and human behavior data, the system dynamically adjusts lighting parameters (brightness, color temperature, and on / off status) to balance energy conservation and safety. A light sensor monitors the corridor lighting intensity in real time, automatically supplementing the light to a preset threshold (e.g., ≥200 lumens at night). A sound sensor detects movement in the corridor; when no one is around, the system switches to energy-saving mode, reducing brightness to 30%. If a slow-moving person, such as an elderly person, is detected, the brightness is increased to 500 lumens to reduce the risk of falls. Simultaneously, voice reminders are triggered, such as "The floor is slippery, please hold onto the handrail." At night, when people pass by, a voice prompt says "Walk slowly and carefully" to reduce noise from neighbors.

[0142] When the system detects that a person requiring care has entered the building's entrance door, the elevator automatically locks onto that floor and, in conjunction with the elevator system, prioritizes stopping at that floor to reduce waiting time.

[0143] By capturing facial information of people entering and exiting in real time and automatically matching it with the resident database, the system records the resident's fields, including: name, room number, entry and exit time, frequency, and movement speed. When a legitimate resident has no entry or exit records for 3 consecutive days and the registration information is "living alone", the system marks it as "abnormal activity" and arranges for community staff to visit the resident or contact the resident's relatives and friends to understand the resident's situation.

[0144] Example 2:

[0145] Please see Figure 2 Based on Example 1, this embodiment also provides a future community management method based on the Internet of Things, including the following specific steps:

[0146] Step 1: Acquire facial images and audio data; perform facial recognition and age range determination based on the facial images; extract footstep data from the audio data;

[0147] Step 2: Extract audio data, fuse time-frequency dual-path features using an attention mechanism to construct anti-interference footstep sound features, and obtain footstep data; extract audio energy value, zero-crossing rate, and movement speed from the footstep data; calculate footstep distance based on the zero-crossing rate; synthesize feature vectors based on audio energy value and age range, and use them as input to the model, dividing them into three sets for random sampling; select several points in the hyperparameter space for evaluation; use a supervised learning model to determine the physical state of people entering and exiting; calculate the care index by acquiring physical state, movement speed, and footstep distance.

[0148] Step 3: Based on the preset care threshold, compare the care index with the care threshold. If the care threshold is less than the care index, determine that the current person entering or leaving is in good physical condition and continue to identify them; otherwise, if the person is in poor physical condition, execute the corresponding dynamic strategy.

[0149] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0150] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A future community management system based on the Internet of Things, characterized in that: The system includes: The identity recognition module acquires facial images and audio data; Based on facial images, perform facial recognition and age range determination; The data processing module extracts audio data, fuses time-frequency dual-path features using an attention mechanism, constructs anti-interference footstep sound features, and obtains footstep data. From the footstep data, it extracts audio energy value, zero-crossing rate, and movement speed; calculates footstep distance based on the zero-crossing rate; synthesizes feature vectors based on audio energy value and age range, and uses them as input to the model, dividing them into three sets for random sampling; selects several points in the hyperparameter space for evaluation; uses a supervised learning model to determine the physical state of people entering and exiting; and calculates the care index by acquiring physical state, movement speed, and footstep distance. The dynamic strategy response module compares the care index with the preset care threshold based on a preset care threshold. If the care threshold is less than the care index, it determines that the current person entering or leaving is in good physical condition and continues to identify them; otherwise, it executes the corresponding dynamic strategy. Among them, constructing anti-interference footstep sound features includes: Time-domain and frequency-domain features are extracted from the audio data to form two independent feature representations. Short-time energy combined with a recurrent neural network is used to capture the temporal dynamic changes of footstep data, while frequency-domain features are generated into a spectrogram through short-time Fourier transform and key frequency components are extracted using an attention module. Feature fusion is achieved by introducing an attention mechanism. In the cross-modal cross-attention algorithm, temporal features are used as query vectors and frequency domain features are used as key-value pairs. By calculating the correlation between temporal and frequency domain features, an attention weight matrix is ​​generated, and the contribution ratio of frequency domain features is dynamically adjusted. At the same time, reverse attention is calculated symmetrically with frequency domain features as queries and temporal features as keys, and the bidirectional attention results are concatenated. A multi-footstep suppression strategy is embedded in the fusion process. Global information modeling is performed on the fused features. A Transformer encoder is used to capture long-distance dependencies. The features are compressed into fixed-dimensional feature vectors through an attention pooling layer. Footstep features are fused with time and frequency. The classification loss and feature discrimination loss are jointly optimized by training.

2. The future community management system based on the Internet of Things according to claim 1, characterized in that: Facial image and audio data: The system uses cameras to capture the frontal facial features of people entering and exiting, and performs facial recognition and age range determination. Sound in the stairwell is captured by acoustic sensors and then processed for noise reduction.

3. The future community management system based on the Internet of Things according to claim 2, characterized in that: Audio energy value: The audio signal is divided into multiple short time periods with a fixed duration, and each time period is called a window. The choice of window length needs to balance time resolution and computational stability. Within each window, the square value of the signal sample is calculated point by point, and the square values ​​of all samples are summed and averaged to obtain the audio energy value Ep.

4. The future community management system based on the Internet of Things according to claim 1, characterized in that: Zero crossing rate: Within the time window, count the number of times the signal crosses zero and normalize it to a unit time to calculate the zero-crossing rate ZCR; Movement speed: After identity authentication is completed, the system enters a timer-ready state, monitoring the footstep sound field in the stairwell. When the first valid acoustic event matching the preset footstep sound characteristics is detected, the timer is started and marked as the initial moment. ; At this time, the system synchronously records the initial sound pressure level parameters. , as a benchmark reference value for distance calculation; It then retrieves historical step frequency data to verify the continuity of the movement, and marks the termination time after confirming that there is no continuous movement. The total distance traveled, D, is obtained through distance inversion calculation, and the speed of movement, v, is calculated.

5. A future community management system based on the Internet of Things according to claim 4, characterized in that: Assess the physical condition of people entering and exiting: Based on audio energy values ​​and age ranges, a large dataset of labeled data is collected and preprocessed to combine audio energy values ​​and age range information into a feature vector, which is then used as the input to the model. The dataset is divided into training, validation, and test sets. The training set is further divided into K subsets. K-1 subsets are used for training each time, and the remaining subset is used for validation. This process is repeated K times, selecting a different subset of the validation set each time. The average of the K validation results is then calculated as the evaluation metric for model performance. A random sampling method is used to randomly select several points in the hyperparameter space for evaluation. Several different combinations of hyperparameters are selected from the adjusted hyperparameters and their value ranges. The model is trained using the training set, and the performance of the trained model is evaluated on the validation set. The supervised learning model outputs a body state score S.

6. The Internet of Things-based future community management system according to claim 5, characterized in that: Foot spacing: Based on acoustic propagation theory, the foot spacing is obtained according to the ratio of the total moving distance D to the zero-crossing rate.

7. A future community management system based on the Internet of Things according to claim 6, characterized in that: Attention Index: The formula, which integrates age correction factors, physical condition scores, and environmental response indicators, is as follows: In the formula, For the care index; Age correction factor; The historical average movement speed is M; M is the total number of detection points. The measured value at the detection point; The mean; This is the velocity decay sensitivity coefficient; For frequency variation weights; This refers to the step spacing.

8. A future community management method based on the Internet of Things, characterized in that: Includes the following steps: Step 1: Acquire facial images and audio data; perform facial recognition and age range determination based on the facial images; extract footstep data from the audio data; Step 2: Extract audio data, fuse time-frequency dual-path features using an attention mechanism, construct anti-interference footstep sound features, and obtain footstep data; From the footstep data, the audio energy value, zero-crossing rate, and movement speed are extracted; the footstep distance is calculated based on the zero-crossing rate; based on the audio energy value and age range, a feature vector is synthesized and used as the input of the model, divided into three sets for random sampling; Several points are selected in the hyperparameter space for evaluation; a supervised learning model is used to determine the physical condition of people entering and exiting; and the care index is calculated by acquiring physical condition, movement speed, and stride distance. Step 3: Based on the preset care threshold, compare the care index with the care threshold. If the care threshold is less than the care index, determine that the current person entering or leaving is in good physical condition and continue to identify them. Conversely, if the physical condition is poor, a corresponding dynamic strategy is implemented; Among them, constructing anti-interference footstep sound features includes: Time-domain and frequency-domain features are extracted from the audio data to form two independent feature representations. Short-time energy combined with a recurrent neural network is used to capture the temporal dynamic changes of footstep data, while frequency-domain features are generated into a spectrogram through short-time Fourier transform and key frequency components are extracted using an attention module. Feature fusion is achieved by introducing an attention mechanism. In the cross-modal cross-attention algorithm, temporal features are used as query vectors and frequency domain features are used as key-value pairs. By calculating the correlation between temporal and frequency domain features, an attention weight matrix is ​​generated, and the contribution ratio of frequency domain features is dynamically adjusted. At the same time, reverse attention is calculated symmetrically with frequency domain features as queries and temporal features as keys, and the bidirectional attention results are concatenated. A multi-footstep suppression strategy is embedded in the fusion process. Global information modeling is performed on the fused features. A Transformer encoder is used to capture long-distance dependencies. The features are compressed into fixed-dimensional feature vectors through an attention pooling layer. Footstep features are fused with time and frequency. The classification loss and feature discrimination loss are jointly optimized by training.

Images