Real-time detection and alarm method and device for human body falling action

Abstract

The invention provides a human body tumble action real-time detection and alarm method and device, and the method comprises the steps: obtaining original point cloud data containing distance, speed and angle through a millimeter wave radar, filtering static clutter points with the speed lower than a threshold value, generating a dynamic point cloud, calculating the difference between points, generating a cross-frame semantic coherence region, and carrying out the real-time detection and alarm of the human body tumble action. And dynamically allocating weights to generate weighted fusion features, outputting a fall probability through a time sequence convolutional network and classification Head, and triggering an alarm if the fall probability exceeds a threshold value. The method improves the detection accuracy, reduces the cost and protects privacy.

Description

Technical Field

[0001] The present invention relates to the field of fall detection, and more particularly to a method and device for real-time detection and alarm of human fall movements. Background Art

[0002] Currently, fall detection technologies are primarily categorized into wearable and non-wearable approaches. While wearable devices, integrating sensors like inertial measurement units and accelerometers, can achieve a certain degree of fall detection, they suffer from common issues such as discomfort and the tendency of elderly people to forget to wear them, limiting their widespread adoption. Among non-wearable technologies, image-based detection methods are unsuitable for private spaces like bathrooms and bedrooms due to privacy concerns and are significantly less effective in low-light conditions. While Wi-Fi-based solutions are relatively low-cost, they are susceptible to interference from external signals and suffer from poor stability.

[0003] Millimeter-wave radar technology stands out due to its unique advantages. It offers excellent safety and stability, high measurement accuracy, and relatively low cost. More importantly, it effectively captures human motion information without infringing on personal privacy and is suitable for a variety of environments, including those with low lighting or complete darkness.

[0004] However, existing millimeter-wave radar fall detection algorithms generally face the problem of insufficient learning ability of global and local features when processing point cloud data, resulting in limited detection accuracy. At the same time, high computational complexity also increases the requirements for hardware performance and increases technical costs. Summary of the Invention

[0005] The purpose of this application is to overcome the defects in the above-mentioned prior art and provide a real-time detection and alarm method and device for human fall action.

[0006] This application provides a real-time detection and alarm method for human fall action, including:

[0007] Obtain the original point cloud data including distance, speed, and angle collected by the millimeter-wave radar;

[0008] filtering out static clutter points with a speed lower than a first speed threshold according to the speed component in the original point cloud data to generate a dynamic point cloud;

[0009] Based on the dynamic point cloud, calculating the spatial distance difference, speed difference and time difference between each point to generate a semantically coherent region across frames;

[0010] According to the spatial distance difference, speed difference and time difference of each point in the semantically coherent area, a corresponding weight ratio is dynamically assigned to generate a weighted fusion feature;

[0011] Inputting the weighted fusion features into a temporal convolutional network to generate temporal dependency features;

[0012] Input the temporal dependency feature into the classification head and output the probability of falling action;

[0013] When the probability of falling exceeds a preset probability threshold, a local sound and light alarm and a remote notification are triggered.

[0014] Optionally, according to the spatial distance difference, speed difference and time difference of each point in the semantically coherent area, corresponding weight ratios are dynamically assigned to generate weighted fusion features, including:

[0015] The operation of dynamically allocating the weight ratio is achieved through a multi-layer perceptron network.

[0016] Optionally, the weighted fusion features are input into a temporal convolutional network to generate temporal dependency features, wherein the temporal convolutional network comprises at least five one-dimensional convolutional layers, and the temporal dependency features are generated by global average pooling.

[0017] Optionally, the temporal dependency feature is input into the classification head to output the probability of the falling action. The classification head includes two fully connected layers, and the output dimension of the second fully connected layer is one-dimensional.

[0018] Optionally, based on the dynamic point cloud, the spatial distance difference, speed difference and time difference between each point are calculated to generate a semantically coherent area across frames, wherein the semantically coherent area includes speed mean and variance features, and the speed mean and variance features participate in the weight ratio distribution.

[0019] The present application also provides a real-time detection and alarm device for human fall action, comprising:

[0020] The acquisition module obtains the original point cloud data including distance, speed and angle collected by the millimeter wave radar;

[0021] a filtering module, configured to filter out static clutter points having a velocity lower than a first velocity threshold according to a velocity component in the original point cloud data, and generate a dynamic point cloud;

[0022] a calculation module, which calculates the spatial distance difference, speed difference and time difference between each point based on the dynamic point cloud, and generates a semantically coherent region across frames;

[0023] An allocation module dynamically allocates corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent area to generate a weighted fusion feature;

[0024] The convolution module inputs the weighted fusion features into the temporal convolutional network to generate temporal dependency features;

[0025] The classification module inputs the temporal dependency features into the classification head and outputs the probability of falling action;

[0026] The trigger module triggers a local sound and light alarm and a remote notification when the probability of the falling action exceeds a preset probability threshold.

[0027] Optionally, the allocation module dynamically allocates corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent area to generate a weighted fusion feature, including:

[0028] The operation of dynamically allocating the weight ratio is achieved through a multi-layer perceptron network.

[0029] Optionally, the convolution module inputs the weighted fusion features into a temporal convolutional network to generate temporal dependency features, wherein the temporal convolutional network includes at least five one-dimensional convolutional layers, and generates the temporal dependency features through global average pooling.

[0030] Optionally, the classification module inputs the temporal dependency feature into a classification head and outputs the probability of a fall action. The classification head includes two fully connected layers, and the output dimension of the second fully connected layer is one-dimensional.

[0031] Optionally, the calculation module calculates the spatial distance difference, speed difference and time difference between each point based on the dynamic point cloud, and generates a semantically coherent area across frames, where the semantically coherent area includes speed mean and variance features, and the speed mean and variance features participate in the weight ratio distribution.

[0032] The beneficial effects of this application are:

[0033] The present application provides a method for real-time detection and alarm of human fall actions, including: obtaining raw point cloud data containing distance, speed, and angle collected by millimeter-wave radar; filtering out static clutter points with speeds lower than a first speed threshold based on the speed component in the raw point cloud data to generate a dynamic point cloud; based on the dynamic point cloud, calculating the spatial distance difference, speed difference, and time difference between each point to generate a semantically coherent region across frames; dynamically assigning corresponding weight ratios based on the spatial distance difference, speed difference, and time difference of each point in the semantically coherent region to generate weighted fusion features; inputting the weighted fusion features into a temporal convolutional network to generate temporal dependency features; inputting the temporal dependency features into a classification head to output a fall action probability; when the fall action probability exceeds a preset probability threshold, triggering a local sound and light alarm and a remote notification. The present application collects point cloud data through millimeter-wave radar, filters out static clutter, and combines AI deep learning algorithms to dynamically assign weights to generate weighted fusion features, thereby achieving efficient fall detection and triggering alarms, improving accuracy, reducing costs, and protecting privacy.

[0034] In addition, to address the problems of insufficient local feature capture and weak global spatiotemporal modeling capabilities in traditional point cloud temporal modeling, this application proposes a dual optimization strategy: first, at the data structure modeling level, an irregular patch extraction mechanism driven by speed-space consistency is introduced to highlight key action areas through velocity gradient clustering, maintain the compactness of local geometric structures through spatial constraints, and implicitly suppress interference points, significantly improving local semantic clarity and feature expression capabilities; second, at the network design level, temporal convolution and frame-level pooling operations are combined to encode the temporal evolution of patches, integrate spatial-temporal features to achieve global dynamic modeling, and enhance action discrimination capabilities through global average pooling, thereby further improving detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of real-time detection and alarm of human fall action in this application;

[0036] Figure 2 This is a working diagram of the millimeter wave radar fall detection system in this application;

[0037] Figure 3 This is a schematic diagram of the logical architecture of the millimeter-wave radar human fall detection system in this application.

[0038] Figure 4 This is a schematic diagram of the millimeter wave radar acquisition equipment in this application.

[0039] Figure 5 This is a schematic diagram of the millimeter wave radar fall detection system in this application.

[0040] Figure 6 This is a schematic diagram of the installation of the millimeter wave radar in this application.

[0041] Figure 7 This is a schematic diagram of the fall detection algorithm in this application.

[0042] Figure 8 This is a schematic diagram of the algorithm principle of this application. DETAILED DESCRIPTION

[0043] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it is understood that various forms of implementing the present disclosure are not limited by the embodiments set forth herein. Rather, the embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0044] The present application provides a method and device for real-time detection and alarm of human fall action.

[0045] Based on the above-mentioned method and device for real-time detection and alarm of human fall, the logical architecture of a millimeter-wave radar human fall detection system is as follows:

[0046] Please refer to Figure 3 As shown in the figure, the architecture is mainly divided into five layers according to its functions, namely the collection layer, computing layer, data layer, service layer and user layer.

[0047] The user layer, data layer, and service layer are more software-oriented, the acquisition layer is more hardware-oriented, and the computing layer is both software and hardware-oriented, which is the core component of this system. The functions of each layer of the system are described below:

[0048] The acquisition layer consists of a group of millimeter-wave acquisition terminals deployed at fixed locations and an acquisition controller. The acquisition controller mainly completes functions such as data communication, main control program, task instructions, task scheduling, data reporting, and alarm reporting.

[0049] The computing layer, the core functional layer of the system, primarily performs 4D point cloud generation and motion classification. The software runs on a GPU server. The core algorithm includes processing steps such as denoising the original point cloud, correcting point clouds at different angles, sampling the point cloud, extracting feature points, generating 4D point clouds, extracting local feature information, and classifying and identifying 4D point clouds.

[0050] At the data layer, the database stores the human body radar data collected by the millimeter-wave radar terminal and the real-time four-dimensional point cloud data of the human body after noise reduction by the computing layer algorithm. The business database consists of terminal, task, scheduling, data results, interface data and other data.

[0051] The service layer is the system's resident service, mainly including task scheduling, abnormal alarm and other services; human walking motion recognition service, human sitting posture recognition service, human squatting motion recognition service, human falling motion recognition service, human four-dimensional point cloud data cropping service, and multi-source data fusion service; in addition, the service layer provides data sharing and data interface with external systems.

[0052] The user layer is the end-user interface and consists of two main parts: monitoring management and system management. Monitoring management includes terminal management, parameter management, task management, data management, and real-time detection of human motion status; system management mainly includes interface management, user management, alarm management, and log management.

[0053] Please refer to Figure 4 As shown, the hardware architecture of the system for detecting and identifying human falls using millimeter-wave radar provided in this application includes a data acquisition system, a data transmission system, a data processing system, a remote motion status monitoring system, and a fall alarm system.

[0054] The data acquisition system is responsible for collecting high-precision point cloud data of human motion in real time, including information such as distance, speed, and angle, providing the raw data foundation for subsequent analysis. The acquisition terminal equipment mainly consists of millimeter-wave radar sensors, signal processing units, power supply equipment, and communication devices.

[0055] Please refer to Figure 5 As shown, millimeter-wave radar is used for real-time collection of human motion data. It uses a frequency-modulated continuous wave (FM CW) radar chip (such as the TIIWR6843), operates in the 60GHz / 77GHz frequency bands, supports multi-antenna arrays (e.g., 4Tx / 3Rx), and offers high resolution (millimeter-level accuracy) and a wide field of view (120° horizontal / 60° vertical). Mounted between 1.2 and 1.8 meters above the ground, the radar is ideal for capturing overall human motion, especially when standing or walking. Mounting too high can miss low-level movements such as falls; too low can only detect leg movements, ignoring changes in the upper body. When mounted horizontally, the radar beam primarily covers the frontal area, while mounting at a downward angle is more suitable for monitoring activity near the ground. The mounting angle can be selected based on actual needs. The radar must be securely fixed to prevent data jitter caused by vibration or movement. Wall mounting is particularly suitable for long-term monitoring applications.

[0056] The signal processing unit is integrated into the radar chip, completing ADC sampling, fast Fourier transform (FFT), range-Doppler spectrum analysis, etc., and outputting target point cloud data.

[0057] The power supply device supports DC power supply and POE power supply. The power indicator light on the device is used to ensure that the device can work normally.

[0058] The communication device includes a 5G module and a 5G antenna, which enables low-latency, high-bandwidth real-time backhaul of data to uplink data analysis equipment.

[0059] The anti-interference design suppresses environmental noise (such as static clutter from furniture, pets, etc.) through technologies such as dynamic threshold filtering and Doppler phase compensation.

[0060] Please refer to Figure 6 As shown, the 67Ghz millimeter wave radar acquisition terminal device used in this application includes a power supply device, a millimeter wave radar sensor, a data processing unit, a memory, a processor, and a communication module.

[0061] The data transmission system is used to upload the point cloud data generated by the filtering algorithm and noise removal algorithm to the data processing system through 5G communication.

[0062] The data processing system receives the point cloud data collected by the radar, completes the point cloud according to its sparsity, normalizes the point cloud, proposes a neural network model StapNet algorithm to classify the point cloud, obtains motion classification information, and finally uploads the classification information to the remote detection system.

[0063] The remote motion status monitoring system is based on the millimeter-wave radar data collected by the data processing system and the computing layer algorithm containing the neural network model to achieve real-time grasp of the motion status of the monitored person. It aims to track, analyze and visualize the human body's motion status in real time, and provide long-term health assessment and intelligent early warning.

[0064] Human motion radar data collected by the data acquisition terminal is converted into point cloud data through filtering algorithms and the Fast Fourier Transform algorithm. The data transmission system then transmits this point cloud data to the computing layer, where a noise reduction algorithm further eliminates noise from the point cloud data to prevent radar data from objects other than the human body from affecting the detection results. The computing layer's neural network model receives the de-noised point cloud data, clusters it, extracts its features, analyzes and classifies the feature information, and ultimately outputs the motion recognition results to the data transmission system, which is then transmitted to the user terminal in real time via a 5G communication module. Multi-dimensional data processing and cross-platform interaction transform the raw point cloud information collected by millimeter-wave radar into understandable health indicators, providing round-the-clock safety monitoring for scenarios such as homes, nursing homes, and medical institutions.

[0065] Millimeter wave radar data acquisition equipment installation layout:

[0066] Please continue to refer to Figure 5 As shown, in indoor scenarios, it should be installed on the wall about 1.8 meters from the ground, tilted downward by 5° to 15°. This is suitable for monitoring low-center-of-gravity movements such as falls near low surfaces. It can also expand the limited monitoring area to cover the entire room. According to measurements, one terminal device can collect human motion data within a range of 20 square meters.

[0067] Please refer to Figure 1 and Figure 2 As shown, the present application provides a method for real-time detection and alarm of human fall action, comprising:

[0068] S101, obtaining original point cloud data including distance, speed, and angle collected by the millimeter wave radar;

[0069] Please refer to Figure 7As shown in the figure, the millimeter-wave radar terminal collects human motion data in real time, transmits radar signals through the transmitting antenna, and the returned radar signals are received by the receiving antenna after contacting the human body. The signal processing device built into the radar chip completes ADC sampling, filtering preprocessing, fast Fourier transform (FFT), range-Doppler spectrum analysis, etc., and outputs target point cloud data to the data processing layer.

[0070] The device collects reflected radar signals to generate continuous wave body motion radar data. It then mixes the high-frequency RF signal with the local oscillator (LO) signal, downconverts it to an intermediate frequency (IF) or baseband, and samples it using an ADC to convert the IF / baseband signal into a digital signal. After sampling, the signal is filtered and de-noised using clustering and CFAR algorithms.

[0071] The basic radar equation is:

[0072]

[0073] Among them, P r is the power received by the receiver, P t is the power transmitted by the transmitter, G r is the receiving antenna gain, G t is the transmitting antenna gain, λ is the wavelength of the electromagnetic wave, σ is the target radar cross section (RCS), and R is the distance between the target and the radar.

[0074] After signal mixing and ADC sampling, the sampled data is processed twice with RangeFFT and DopplerFFT (range-Doppler processing) to obtain the energy distribution of the range-Doppler two-dimensional plane. The peak value corresponds to the distance and speed of the target.

[0075] In this application, when the linear frequency modulation rate is S (Hz / s), the difference frequency f obtained after mixing is beat There is an approximate linear relationship with the target distance R:

[0076]

[0077] The Doppler frequency shift f D Relationship with target-to-target relative velocity ν:

[0078]

[0079] In millimeter-wave radar, a second FFT (DopplerFFT) is performed on the data in the slow time dimension (between chirp frames) to obtain the Doppler frequency component and infer the target speed.

[0080] Using the received signal phase difference Δφ, assuming the array element spacing is d and the incident angle is θ, the angle estimation formula is:

[0081]

[0082] The collected target points (including the distance (R), pitch angle (φ), azimuth angle (Θ), and speed information ν) are transmitted to the data processing system through the communication system.

[0083] The system uses the formula to calculate the target points collected by the data acquisition system:

[0084]

[0085] S102, filtering out static clutter points with a speed lower than a first speed threshold according to the speed component in the original point cloud data, to generate a dynamic point cloud;

[0086] Generate a point cloud, where each point contains information (x, y, z, v), and perform data preprocessing on the collected point cloud:

[0087] Normalization processing, where the coordinate normalization takes the center of the scene as the origin formula is:

[0088]

[0089] Among them, μ x is the mean value of the point cloud coordinates, σ x is the coordinate standard deviation.

[0090] The speed normalization formula is:

[0091]

[0092] Linearly map to [0,1], preserving relative motion intensity.

[0093] Dynamic background filtering, based on the velocity eigenvalue, filters out point clouds with eigenvalue v < 0.1m / s, which can reduce the data volume by 60% and lower the computational load.

[0094] S103, calculating the spatial distance difference, speed difference, and time difference between each point based on the dynamic point cloud, and generating a semantically coherent region across frames;

[0095] To overcome the shortcomings of traditional point cloud temporal modeling in local feature capture and global spatiotemporal modeling, this step adopts a dynamic patch extraction mechanism driven by speed-space consistency. The specific implementation process is as follows:

[0096] Assume that the input point cloud sequence is as follows:

[0097]

[0098] Among them, the time label of the i-th point is t i ∈{1,…,T},x i ,y i ,z i ,v i Represent the xyz axis coordinates and velocity scalar of the i-th point respectively.

[0099] First, for each pair of neighbor points (i, j) (determined by spatial KNN), this application constructs a three-element difference vector:

[0100]

[0101] in, That is, it is expressed as the spatial coordinates of a point, which reflects the degree of deviation between two points in space, speed and frame index. i, j inputs a two-layer fully connected network, and after ReLU activation and Softmax normalization, the attention coefficient is obtained:

[0102]

[0103] The Softmax operation ensures that the weights of the three are adaptive and differentiable.

[0104] The total number of parameters in the attention module is proportional to the width of the hidden layer, and standard regularization is used to ensure stable training.

[0105] S104, dynamically assigning corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent region to generate a weighted fusion feature;

[0106] Use the obtained attention coefficient α ij , construct an adaptive weighted distance:

[0107]

[0108] Compared with the fixed ratio linear combination, It can automatically focus on the most discriminative metric weight in various scenarios such as slow walking, fast running, and actors falling.

[0109] Finally, for all The similarity matrix is ​​calculated and density peak clustering is used to divide several patches. The number of cluster centers M is adaptively determined by the point density, laying the foundation for downstream multi-scale modeling.

[0110] In order to facilitate unified processing of all subsequent patches in the same feature space, this application designs a lightweight two-layer MLP embedder.

[0111] Let the Patch feature matrix extracted from the tth frame be:

[0112]

[0113] Among them, C p Including point coordinates, velocity mean, and velocity variance in the region, the total dimension is 5. MLP_em bed first passes the projection matrix Map the features to the hidden layer, then pass through the ReLU activation function and the normalized layer LayerNorm, and finally use Finally projected to dimension d.

[0114] After obtaining the patch embedding for each frame, this application performs average pooling on each frame according to the number of patches to obtain a frame-level feature sequence, and then splices the time dimension of the sequence to obtain a tensor:

[0115]

[0116] S105, inputting the weighted fusion features into a temporal convolutional network to generate temporal dependency features;

[0117] The sequence is then fed into a deep temporal convolutional network (TCN). The TCN consists of L layers of 1D convolutions, with the lth layer feature map:

[0118]

[0119] TCN final output T ' is the length after time series pooling. This application uses global average pooling:

[0120]

[0121] This operation compresses temporal information into a fixed-dimensional vector without requiring additional parameters and helps alleviate overfitting. In experiments, this application also verified that maximum pooling or attention weight pooling (generating temporal weights through small FCs) slightly improves final accuracy, but average pooling is preferred due to its simplicity and robustness.

[0122] S106: Input the temporal dependency feature into a classification head and output a fall action probability;

[0123] The classification head is responsible for converting the global feature vector g into the final fall probability prediction.

[0124] In order to take into account both model capacity and real-time performance, this application adopts a two-layer fully connected (FC) structure, supplemented by nonlinear activation, normalization and Dropout regularization, and uses the Sigmoid function to output the binary classification probability.

[0125] The specific design is as follows.

[0126] First layer (FC1):

[0127]

[0128] First layer (FC1):

[0129]

[0130] Sigmoid output:

[0131]

[0132] The action category probability output by the classification branch is then input into the decision fusion module, and finally the fall judgment result is output.

[0133] S107: When the probability of the falling action exceeds a preset probability threshold, triggering a local sound and light alarm and a remote notification.

[0134] Please refer to Figure 7 As shown in the figure, when the computing layer outputs a fall detection signal, the data transmission system transmits the alarm information to the fall alarm system. The alarm system is divided into two parts: local response and remote notification. Upon receiving the alarm information, the local response part triggers the sound and light alarm (LED flashes red and a buzzer sounds). The remote alarm part transmits the alarm information to the children's mobile phone app through the communication module, prompting them to provide medical assistance to the monitored person. When the local alarm system or the remote system receives the "aided" command, the alarm activity will cease.

[0135] The present application also provides a real-time detection and alarm device for human fall action, comprising:

[0136] The acquisition module obtains the original point cloud data including distance, speed and angle collected by the millimeter wave radar;

[0137] a filtering module, configured to filter out static clutter points having a velocity lower than a first velocity threshold according to a velocity component in the original point cloud data, and generate a dynamic point cloud;

[0138] a calculation module, which calculates the spatial distance difference, speed difference and time difference between each point based on the dynamic point cloud, and generates a semantically coherent region across frames;

[0139] An allocation module dynamically allocates corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent area to generate a weighted fusion feature;

[0140] The convolution module inputs the weighted fusion features into the temporal convolutional network to generate temporal dependency features;

[0141] The classification module inputs the temporal dependency features into the classification head and outputs the probability of falling action;

[0142] The trigger module triggers a local sound and light alarm and a remote notification when the probability of the falling action exceeds a preset probability threshold.

[0143] Furthermore, the allocation module dynamically allocates corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent region to generate weighted fusion features, including:

[0144] The operation of dynamically allocating the weight ratio is achieved through a multi-layer perceptron network.

[0145] Furthermore, the convolution module inputs the weighted fusion features into a temporal convolutional network to generate temporal dependency features. The temporal convolutional network includes at least five one-dimensional convolutional layers, and the temporal dependency features are generated by global average pooling.

[0146] Furthermore, the classification module inputs the time-dependent features into the classification head and outputs the probability of falling action. The classification head includes two fully connected layers, and the output dimension of the second fully connected layer is one-dimensional.

[0147] Optionally, the calculation module calculates the spatial distance difference, speed difference and time difference between each point based on the dynamic point cloud, and generates a semantically coherent area across frames, where the semantically coherent area includes speed mean and variance features, and the speed mean and variance features participate in the weight ratio distribution.

[0148] The above description of the embodiments is intended to facilitate understanding and application of this application by those skilled in the art. It will be readily apparent to those skilled in the art that various modifications to the above embodiments can be made, and the general principles described herein can be applied to other embodiments without requiring creative effort. Therefore, this application is not limited to the above embodiments. Any improvements or modifications made to this application by those skilled in the art based on the disclosure of this application should fall within the scope of protection of this application.

Claims

1. A real-time detection and alarm method for human falling action, characterized in that: include: Obtain the original point cloud data including distance, speed, and angle collected by the millimeter-wave radar; filtering out static clutter points with a speed lower than a first speed threshold according to the speed component in the original point cloud data to generate a dynamic point cloud; Based on the dynamic point cloud, calculating the spatial distance difference, speed difference and time difference between each point to generate a semantically coherent region across frames; According to the spatial distance difference, speed difference and time difference of each point in the semantically coherent area, a corresponding weight ratio is dynamically assigned to generate a weighted fusion feature; Inputting the weighted fusion features into a temporal convolutional network to generate temporal dependency features; Input the temporal dependency feature into the classification head and output the probability of falling action; When the probability of falling exceeds a preset probability threshold, a local sound and light alarm and a remote notification are triggered.

2. The method for real-time detection and alarm of a human falling action according to claim 1, characterized in that: According to the spatial distance difference, speed difference and time difference of each point in the semantically coherent area, the corresponding weight ratio is dynamically assigned to generate a weighted fusion feature, including: The operation of dynamically allocating the weight ratio is achieved through a multi-layer perceptron network.

3. The method for real-time detection and alarm of a human falling action according to claim 1, characterized in that: The weighted fusion features are input into a temporal convolutional network to generate temporal dependency features. The temporal convolutional network includes at least five one-dimensional convolutional layers, and the temporal dependency features are generated by global average pooling.

4. The method for real-time detection and alarm of a human fall according to claim 1, characterized in that: The temporal dependency features are input into the classification head to output the probability of falling action. The classification head includes two fully connected layers, and the output dimension of the second fully connected layer is one-dimensional.

5. The method for real-time detection and alarm of a human falling action according to claim 1, characterized in that: Based on the dynamic point cloud, the spatial distance difference, speed difference and time difference between each point are calculated to generate a semantically coherent region across frames, wherein the semantically coherent region includes speed mean and variance features, and the speed mean and variance features participate in the weight ratio distribution.

6. A real-time detection and alarm device for human falling action, characterized in that: include: The acquisition module obtains the original point cloud data including distance, speed and angle collected by the millimeter wave radar; a filtering module, configured to filter out static clutter points having a velocity lower than a first velocity threshold according to a velocity component in the original point cloud data, and generate a dynamic point cloud; a calculation module, which calculates the spatial distance difference, speed difference and time difference between each point based on the dynamic point cloud, and generates a semantically coherent region across frames; An allocation module dynamically allocates corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent area to generate a weighted fusion feature; A convolution module inputs the weighted fusion features into a temporal convolutional network to generate temporal dependency features; The classification module inputs the temporal dependency features into the classification head and outputs the probability of falling action; The trigger module triggers a local sound and light alarm and a remote notification when the probability of the falling action exceeds a preset probability threshold.

7. The real-time detection and alarm device for human falling action according to claim 6, characterized in that: The allocation module dynamically allocates corresponding weight ratios according to the spatial distance difference, speed difference, and time difference of each point in the semantically coherent area to generate weighted fusion features, including: The operation of dynamically allocating the weight ratio is achieved through a multi-layer perceptron network.

8. The real-time detection and alarm device for human falling action according to claim 6, characterized in that: The convolution module inputs the weighted fusion features into a temporal convolutional network to generate temporal dependency features. The temporal convolutional network includes at least five one-dimensional convolutional layers, and the temporal dependency features are generated by global average pooling.

9. The method for real-time detection and alarm of a human falling action according to claim 6, characterized in that: The classification module inputs the time-dependent features into the classification head and outputs the probability of falling action. The classification head includes two fully connected layers, and the output dimension of the second fully connected layer is one-dimensional.

10. The real-time detection and alarm device for human falling action according to claim 6, characterized in that: The calculation module calculates the spatial distance difference, speed difference and time difference between each point based on the dynamic point cloud, and generates a semantically coherent region across frames, wherein the semantically coherent region includes speed mean and variance features, and the speed mean and variance features participate in the weight ratio distribution.

Images