A human body fall risk early warning method, system, device and storage medium

Abstract

The application provides a human body fall risk early warning method and system, equipment and a storage medium, the method comprising: acquiring point cloud data of a human body in a monitoring area and preprocessing to form an original point cloud sequence; processing the original point cloud sequence to enhance the point cloud density; extracting spatial and temporal features of the point cloud enhanced frame; splicing and processing the space-time features to obtain shared features, and performing dangerous action behavior classification and fall risk assessment based on the shared features. The application relies on non-contact radar sensing and edge computing capabilities to achieve low latency and high privacy. Through the fusion of intensity neural networks and space-time attention mechanisms, efficient mining of space-time context information of risk actions and key frame focusing are realized, the features are more targeted, a double-branch parallel architecture and an adaptive decision module are used to combine action recognition and risk quantification, solving the problem of insufficient reliability of a single judgment dimension, and realizing intelligent early warning with high accuracy and good flexibility.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method, system, device, and storage medium for early warning of human fall risk. Background Technology

[0002] Patients with chronic diseases and the elderly often experience a decline in motor skills, which may be accompanied by abnormal behaviors such as imbalance, staggering, prolonged sitting, and wandering. These are common early signs of impending falls. Early detection of these abnormal behaviors in daily life and timely rehabilitation intervention can effectively improve the elderly's inherent abilities and thus prevent injuries from falls. Therefore, developing an effective home-based fall warning method that does not significantly interfere with personal living spaces or infringe on privacy has become a current research hotspot and challenge.

[0003] Traditional fall warning technologies have significant limitations. For example, visual warning solutions rely on cameras to capture images, but their recognition accuracy drops sharply in low light (such as at night) or when the target is obstructed (such as by furniture). Furthermore, they pose ethical risks that infringe on user privacy, making them difficult to apply in private settings such as bedrooms and bathrooms.

[0004] Wearable device warning solution: Users need to wear the device containing gyroscope and accelerometer at all times. The elderly or people with mobility difficulties may experience discomfort when wearing it. In addition, issues such as device battery life and charging lead to a high "missed wear rate" in actual use and insufficient warning coverage.

[0005] Traditional millimeter-wave radar solutions, while able to circumvent privacy issues and adapt to complex lighting / occlusion environments, are limited by the characteristics of radar point clouds—low point cloud information density, ghosting noise caused by multipath effects, sparse point clouds of human targets, and existing methods mostly focus on "how to accurately identify fall events," lacking the ability to predict "dangerous actions before a fall (such as imbalance, staggering)," resulting in poor early warning timeliness. At the same time, existing point cloud processing models do not fully utilize point cloud intensity (SNR) and temporal information, resulting in insufficient differentiation of dangerous actions before a fall, which easily leads to false alarms or missed alarms.

[0006] For example, Chinese patent ZL202510869104.4 discloses an activity recognition system and method based on millimeter-wave radar Doppler features. Through dynamic feature filtering and cross-channel information fusion technology, it significantly improves the accuracy and robustness of activity recognition in scenarios with insufficient or imbalanced data. While this patent uses radar Doppler features for action recognition and improves the accuracy of activity recognition in specific scenarios through algorithms, it fundamentally fails to address the challenges of missing spatial details and environmental perception. Doppler features are essentially one-dimensional time series or two-dimensional time spectrum maps, discarding the target's geometric shape, size, and orientation information in space. Furthermore, it can only measure the target's velocity in the radar radial direction (i.e., towards or away from the radar). This means that the same action will generate different Doppler features depending on the target's azimuth angle relative to the radar, making the model difficult to generalize. Environmental occlusion, multi-target discrimination, and aliasing also present significant challenges.

[0007] For example, Chinese patent ZL202410757589.3 discloses a fall warning method and system based on deep learning, which combines outdoor images to train indoor radar, achieving personalized model training and fall detection throughout the indoor and outdoor process. The "joint training model of outdoor images and indoor radar" proposed in this patent faces fundamental challenges at the data level. The core issue is that the model relies on multimodal paired data of the same person in different scenarios. However, in reality, both outdoor conditions (such as weather, lighting, ground conditions, and shooting angle) and indoor conditions (such as layout, occlusion, radar angle, and personnel density) are highly uncertain and complex, making it difficult to systematically acquire and accurately label large-scale, high-quality image-radar paired fall datasets. Without sufficient training and testing data, the model's generalization ability is difficult to guarantee, let alone achieve effective personalized training. Therefore, this solution is not practical enough under current technological conditions, and its implementation feasibility is low.

[0008] To address the above issues, there is an urgent need for a technical solution that balances privacy protection, environmental adaptability, and timely early warning. Summary of the Invention

[0009] In order to achieve the above-mentioned objectives and other advantages of the present invention, the first objective of the present invention is to provide a method for early warning of human fall risk, comprising the following steps:

[0010] Acquire point cloud data of the human body within the monitoring area, and process the point cloud data to form an original point cloud sequence;

[0011] The original point cloud sequence is processed to enhance the point cloud density;

[0012] Spatial and temporal features are extracted from the point cloud augmentation frame;

[0013] The spatial features and temporal features are concatenated and processed to obtain shared features. Based on the shared features, dangerous actions and behaviors are classified and fall risk is assessed.

[0014] Furthermore, the step of processing the point cloud data to form an original point cloud sequence includes:

[0015] Perform Fast Fourier Transform and Constant False Alarm Rate (CFAR) detection operations on the collected raw data to extract the 3D coordinates and intensity of the target point cloud, forming the original point cloud sequence:

[0016]

[0017] in, Total number of frames Indicates the first Frame points The feature dimension of a point.

[0018] Further, the step of processing the original point cloud sequence to enhance the point cloud density includes:

[0019] A multi-objective clustering algorithm was used to remove noise and ghosting from the original point cloud sequence.

[0020] A dual sliding window strategy, employing frame stacking sliding window and temporal sampling sliding window, is used to enhance point cloud density.

[0021] Furthermore, the step of using a multi-objective clustering algorithm to remove noise and ghosting from the original point cloud sequence specifically involves predicting the target trajectory using Kalman filtering, associating the new frame point cloud with the historical trajectory, allocating valid point clouds to the target trajectory, removing detached noise points and ghosting points caused by multipath effects, and outputting the purified point cloud sequence.

[0022] Furthermore, the step of enhancing point cloud density using a dual sliding window strategy employing frame stacking sliding window and temporal sampling sliding window includes:

[0023] Based on the current frame, the previous frame... The frame and the current frame are stored in the edge memory;

[0024] Merge consecutive frame point clouds within a frame stacking sliding window into an enhanced frame using stacking operations:

[0025]

[0026] in, Indicates the length of the frame stacking sliding window. Indicates the frame stacking sliding window step size. This represents a stack operation of multiple point cloud frames within a sliding window, performed on the augmented frame. The point count is fixed. If the point count exceeds a preset value, random downsampling is performed. If the point count does not exceed the preset value, zero padding is performed to obtain the enhanced frame sequence.

[0027]

[0028] Extract consecutive enhanced frames from the enhanced frame sequence as temporal samples, the first... Each video clip is represented as:

[0029]

[0030] in, Indicates the length of the timing sampling sliding window. Indicates the step size of the timing sampling sliding window;

[0031] The final constructed dataset:

[0032] .

[0033] Furthermore, the spatial feature extraction steps for the point cloud augmentation frame include:

[0034] For a given set of point clouds At the processing point First, the KNN clustering algorithm is used in the feature space to find its... Let there be 3 adjacent points, denoted as _ . ;

[0035] Using the currently selected point and the found For each neighboring point, the EdgeConv convolution method is used to find the edge features belonging to the current point. The update formula for EdgeConv is expressed as:

[0036]

[0037] in, For point In the Layer feature representation, The relative characteristic differences between adjacent points and the center point. In EdgeConv, a multilayer perceptron is used to perform non-linear mapping on the convolution results;

[0038] Max pooling is used to optimize the convolution result in the neighborhood. Aggregation is performed to ensure local invariance between features.

[0039] Furthermore, a mapping transformation is performed before entering the EdgeConv layer for KNN computation:

[0040]

[0041] Use the flipped Replace the original strength information Perform KNN calculations to satisfy the requirement that smaller values ​​are more likely to be selected as neighboring points;

[0042] In the feature construction stage, the original intensity information is used. To ensure that the data content is not distorted: for the coordinate dimension, adopt... The difference between convolutions is constructed using a specific method; for the intensity dimension, the difference is calculated using... The convolutional results are constructed using a method that emphasizes the intensity dimension of features by averaging. After calculation, the features are then concatenated. And finally update:

[0043] .

[0044] Furthermore, during data preprocessing, a Boolean mask is used to mark all points with zero dimensions. And the masked points are ignored during KNN construction and convolution. This is done to remove the effect of all zeros on the convolution result.

[0045] Furthermore, the spatial feature extraction step for the point cloud enhancement frame also includes:

[0046] Construct an attention module based on the intensity dimension, using intensity values ​​as adjustment factors to dynamically adjust the weights between channels at each point:

[0047]

[0048] in, Indicates the intensity value. Indicates channel weight, Use the Sigmoid activation function; For learnable scaling factors, This represents a normalization operation, used to scale the intensity of different value ranges to a relatively fixed interval. The calculation formula is as follows:

[0049]

[0050] in, The pre-set pre-scaling factor;

[0051] Finally, the channel attention weights are obtained. The input features are processed through bitwise multiplication. Weighted average:

[0052] .

[0053] Furthermore, the temporal feature extraction steps for the point cloud enhancement frames include:

[0054] For action recognition tasks, we use contextual information in the temporal space to perform temporal modeling. Through temporal modeling, we obtain the correlation characteristics of actions in the time axis dimension and deduce the possible execution of the next action.

[0055] Furthermore, the time series modeling is implemented using a BiGRU based on GRU module augmentation: for the GRU module, the input time series feature vector is... ,in Its recursive update process is represented as:

[0056]

[0057] Among them, the results Represents the hidden state at time t, input This represents the hidden state at the previous time step; the GRU module consists of two parts: the update gate and the reset gate. The calculation processes for the two parts are as follows:

[0058]

[0059] in, For the Sigmoid activation process, This represents element-wise multiplication. and All of these are learnable parameter matrices.

[0060] Furthermore, the step of extracting temporal features from the point cloud augmentation frame also includes:

[0061] Adding a temporal attention structure as a feature enhancement at the temporal feature level:

[0062] Suppose a query is performed where the key and value are obtained through linear transformations:

[0063]

[0064] in, The feature sequence output by BiGRU Given a learnable parameter matrix, the attention weights are calculated using the scaled dot product attention formula:

[0065]

[0066] in, For vectors Dimension size, These are the channel weights for the time series.

[0067] use The final weighted time series features are calculated.

[0068] Furthermore, the step of concatenating and processing the spatial features and temporal features to obtain shared features includes:

[0069] The spatial features and the temporal features are concatenated and then input into a fully connected layer. The model is trained using the GeLU nonlinear loss function and processed by a Dropout layer to obtain shared features.

[0070] Furthermore, the steps for classifying dangerous actions and behaviors and assessing fall risk based on the shared features include:

[0071] The shared features are input into a classification network, which includes fully connected layers and a softmax activation function. The output is transformed into a class probability distribution, and the final classification result is output.

[0072] The shared features are input into a risk assessment network, which sequentially includes a fully connected layer, a Gel nonlinear activation function, a Dropout layer, a fully connected layer, and a Sigmoid activation function. The output is compressed to between 0 and 1, and the final output is a risk value.

[0073] The classification results and the risk values ​​are combined to generate a final decision.

[0074] The second objective of this invention is to provide a human fall risk early warning system, which utilizes the aforementioned method and includes a radar data acquisition module, a point cloud processing enhancement module, a spatiotemporal feature extraction module, and a fall risk decision-making module; wherein,

[0075] The radar data acquisition module is used to acquire point cloud data of the human body within the monitoring area and process the point cloud data to form an original point cloud sequence.

[0076] The point cloud processing enhancement module is used to process the original point cloud sequence to enhance the point cloud density;

[0077] The spatiotemporal feature extraction module is used to extract spatial and temporal features from the point cloud enhancement frame;

[0078] The fall risk decision-making module is used to concatenate and process the spatial features and temporal features to obtain shared features, and to classify dangerous actions and behaviors and assess fall risks based on the shared features.

[0079] A third objective of the present invention is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0080] A fourth objective of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0081] Compared with the prior art, the beneficial effects of the present invention are:

[0082] This invention provides a method, system, device, and storage medium for human fall risk early warning. The perception layer relies on non-contact radar perception and edge computing capabilities to achieve low latency and high privacy. The algorithm layer, through the fusion of intensity neural networks and spatiotemporal attention mechanisms, overcomes the limitations of traditional methods in feature extraction, achieving efficient mining of spatiotemporal context information of risky actions and key frame focusing, resulting in more targeted features. The decision layer's innovative dual-branch parallel architecture and adaptive decision module combine "action recognition" with "risk quantification," solving the problem of insufficient reliability of a single judgment dimension and achieving intelligent early warning with high accuracy and flexibility.

[0083] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0084] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0085] Figure 1 Flowchart of a method for early warning of human fall risks;

[0086] Figure 2 A framework diagram of a human fall risk early warning system;

[0087] Figure 3 A schematic diagram of millimeter-wave radar sensor deployment;

[0088] Figure 4 Flowchart for spatial and temporal feature extraction;

[0089] Figure 5 A flowchart for decision-making regarding human fall risk;

[0090] Figure 6 To validate the effectiveness of this method in action classification on a public dataset. Figure 1 ;

[0091] Figure 7 To validate the effectiveness of this method in action classification on a public dataset.Figure 2 ;

[0092] Figure 8 To validate the effectiveness of this method in action classification on a public dataset. Figure 3 ;

[0093] Figure 9 A block diagram of a human fall risk warning system;

[0094] Figure 10 This is a schematic diagram of a computer device.

[0095] Figure 11 This is a schematic diagram of a computer-readable storage medium. Detailed Implementation

[0096] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0097] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0098] The drawing numbers in this application are only used to distinguish the steps in the scheme and are not used to limit the execution order of the steps. The specific execution order is as described in the specification.

[0099] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0100] This invention provides a method and system for early warning of human fall risk based on real-time point cloud enhancement of millimeter-wave radar and spatiotemporal attention mechanism. By optimizing real-time point cloud enhancement, multi-attention feature fusion and risk classification decision update, it can achieve accurate identification and timely warning of dangerous actions before a human fall. It breaks through the accuracy bottleneck of existing single radar point cloud algorithms in dangerous action identification and fall risk prediction, and can be widely used in scenarios such as home-based elderly care and medical monitoring.

[0101] Example 1

[0102] A method for early warning of human fall risk, such as Figures 1-2 As shown, it includes the following steps:

[0103] S100: Acquire point cloud data of the human body within the monitoring area, and process the point cloud data to form an original point cloud sequence;

[0104] like Figure 3 As shown, this embodiment uses a millimeter-wave radar sensor 1 (such as TI's IWR6843AOP) to collect point cloud data of the human body in the monitoring area in real time. The radar working mode is FMCW (Frequency Modulated Continuous Wave), and the azimuth resolution is improved by MIMO technology.

[0105] In some embodiments, the step of processing the point cloud data to form an original point cloud sequence includes:

[0106] The collected raw data is subjected to Fast Fourier Transform (FFT) and Constant False Alarm Rate (CFAR) operations to extract the three-dimensional coordinates and intensity of the target point cloud, forming the original point cloud sequence:

[0107]

[0108] in, Total number of frames Indicates the first The number of points in a frame (not fixed). Represents the feature dimension of a point. Optionally, The feature dimension of a point is represented (e.g., x, y, z, snr), where snr is the signal strength of the point cloud.

[0109] This embodiment identifies the real target by combining two-dimensional FFT processing with CFAR detection algorithm, calculates the target angle based on antenna array, and finally integrates the point cloud information containing three-dimensional coordinates (x, y, z) and signal-to-noise ratio (SNR) into TLV format data packets for output.

[0110] S200. Process the original point cloud sequence to enhance the point cloud density;

[0111] In some embodiments, the step of processing the original point cloud sequence to enhance the point cloud density includes:

[0112] A multi-objective clustering algorithm is used to remove noise and ghosting from the original point cloud sequence. Specifically, the target trajectory is predicted by Kalman filtering, the new frame point cloud is associated with the historical trajectory, the effective point cloud is assigned to the target trajectory, and the wandering noise points and ghosting points caused by multipath effects are removed, resulting in a purified point cloud sequence.

[0113] To address the sparse point cloud problem in a single frame, a dual sliding window strategy combining frame stacking sliding window and temporal sampling sliding window is employed to enhance point cloud density. Specifically, the following steps are included:

[0114] Based on the current frame, the previous frame... The frame and the current frame are stored in the edge memory;

[0115] Frame stacking sliding window: Merges consecutive frame point clouds within a frame stacking sliding window into an enhanced frame through stacking operations.

[0116]

[0117] in, Indicates the length of the frame stacking sliding window. Indicates the frame stacking sliding window step size. This represents a stack operation of multiple point cloud frames within a sliding window, performed on the augmented frame. The point count is fixed. If the point count exceeds a preset value, random downsampling is performed. If the point count does not exceed the preset value, zero padding is performed to obtain the enhanced frame sequence.

[0118]

[0119] Assume the point cloud video sequence before processing is The length of the sliding window The step size is 3. If the value is 2, then the enhanced new point cloud video sequence is: Here, the length of the sliding window... and sliding window movement step These can all be set individually, increasing the flexibility of stacking. However, considering that a frame before point cloud augmentation appearing in two frames after point cloud augmentation would cause information redundancy, the sliding window length and stride in this sliding window are set to the same value. In addition, as mentioned before, the number of points in each frame should be the same. Therefore, after stacking, downsampling / zero-point padding operations are required depending on whether there are too many or too few points.

[0120] Extract consecutive enhanced frames from the enhanced frame sequence as temporal samples, the first... Each video clip is represented as:

[0121]

[0122] in, Indicates the length of the timing sampling sliding window. Indicates the step size of the timing sampling sliding window;

[0123] The final constructed dataset:

[0124] .

[0125] The second sliding window also has two parameters that can be set: length and stride. The window length determines the length of the final generated data samples, while the window stride determines how many frames are between sample constructions. Appropriately reducing the stride can make the samples include action slices from various time points, which helps to enhance the model's generalization performance and real-time classification function. However, too short a stride will lead to a large number of similar samples, affecting the training effect.

[0126] S300: Extract spatial and temporal features from the point cloud enhancement frame;

[0127] The convolutional structure of the EdgeConv edge convolutional layer is implemented as follows: KNN algorithm is used to construct clustering relationships between point clouds, and neighborhood information is imported to improve the representation ability of local structures. In some embodiments, such as... Figure 4 As shown, the steps for spatial feature extraction of the point cloud augmentation frame include:

[0128] For a given set of point clouds The algorithm at the processing point First, the KNN clustering algorithm is used in the feature space to find its... Let there be 3 adjacent points, denoted as _ . ;

[0129] Using the currently selected point and the found For each neighboring point, the EdgeConv convolution method is used to find the edge features belonging to the current point. The update formula for EdgeConv is expressed as:

[0130]

[0131] in, For point At point number Layer feature representation, The relative characteristic differences between adjacent points and the center point. In EdgeConv, a multilayer perceptron is used to perform non-linear mapping on the convolution results; the convolution result is composed of two parts: the original features of the points and the relative difference features between adjacent points.

[0132] After nonlinear mapping, the model uses max pooling to apply the convolution result to the neighborhood. Aggregation is performed to ensure local invariance between features.

[0133] By using EdgeConv, a point-to-point convolution design that leverages inter-point differences, the model can capture the relative geometric relationships between neighboring points in the point cloud. This property is utilized to enhance the model's ability to perceive local structure.

[0134] Furthermore, considering the characteristic that the main body in the millimeter-wave radar point cloud may be located in different positions, a design is made to use relative differences to construct eigenvalues ​​to ensure invariance when the input signal is translated as a whole.

[0135] For intensity information in millimeter-wave point clouds, considering the difference in its representation between information dimension and coordinate information, a mapping transformation is performed before entering the EdgeConv layer for KNN calculation:

[0136]

[0137] Even after flipping Replace the original strength information KNN calculations are performed to ensure that smaller values ​​are more likely to be selected as neighboring points. During the feature construction stage, the original intensity information is used. This is to ensure that the data content is not distorted.

[0138] In the feature construction stage, for the coordinate dimension, the following is adopted: The difference between convolutions is constructed using a specific method; for the intensity dimension, the difference is calculated using... The convolutional results are constructed using a method that emphasizes the intensity dimension of features through averaging. After calculation, the features are then concatenated. And finally update:

[0139] .

[0140] Furthermore, considering the requirement in model design that the number of input points be the same for each frame, a Boolean mask is used to mark all points with zero dimensions during data preprocessing to address the zero-padding issue. And the masked points are ignored during KNN construction and convolution. This is done to remove the effect of all zeros on the convolution result.

[0141] To improve model classification performance, a spatial attention mechanism is added to the spatial feature dimension to enhance the model's focus on specific feature channels. Specifically, an intensity-based attention module is constructed, using the intensity (SNR) value as a modulating factor to dynamically adjust the weights between channels at each point. Constructing channel weights Channel weight The calculation formula is as follows:

[0142]

[0143] in, Indicates the intensity value. Indicates channel weight, Use the Sigmoid activation function; For learnable scaling factors, This represents a normalization operation, used to scale the intensity of different value ranges to a relatively fixed interval. The calculation formula is as follows:

[0144]

[0145] in, The pre-set pre-scaling factor;

[0146] Finally, the channel attention weights are obtained. The input features are processed through bitwise multiplication. Weighted average:

[0147]

[0148] This enables an attention mechanism based on point intensity in spatial features.

[0149] Figure 4 In the first layer of edge convolution (EdgeConv-Layer1), the feature difference vector is calculated for the 8 neighborhoods of each point, and the 4-dimensional difference features are mapped to 32 dimensions through a 1*1 convolution kernel;

[0150] Intensity-based attention layer (SNR Attention) incorporates SNR attention and calculates weights based on the signal-to-noise ratio of the point cloud. Element-wise weighting of the 32-dimensional features;

[0151] The second edge convolutional layer (EdgeConv-Layer2) maps the 32-dimensional difference features to 64 dimensions using a 1*1 convolutional kernel;

[0152] Channel Attention 1 layer incorporates channel attention and learns feature channel weights. Channel-by-channel weighting of the 64-dimensional features;

[0153] EdgeConv-Layer3 expands the neighborhood range to 12, captures global spatial relationships, and outputs 64-dimensional deep spatial features through a 1*1 convolutional kernel;

[0154] Channel Attention 2 is introduced to further optimize the feature channel weights.

[0155] The feature concatenation layer concatenates the original point cloud features (4D), the output of the first edge convolutional layer (EdgeConv-Layer1) + intensity attention layer (SNR Attention) (32D), the output of the second edge convolutional layer (EdgeConv-Layer2) + channel attention layer (Channel Attention 1) (64D), and the output of the third edge convolutional layer (EdgeConv-Layer3) + channel attention layer (Channel Attention 2) (64D), resulting in 164-dimensional multi-scale features.

[0156] A fully connected layer (Linear) + ReLU activation function + Dropout layer maps 164-dimensional features to 512 dimensions;

[0157] The Global Max Pooling layer, after global max pooling, reduces the dimensionality of the 512-dimensional features to 256-dimensional features through a linear layer, which serves as the final output of the spatial features for a single frame.

[0158] It captures the temporal correlation of spatial features of point clouds in consecutive frames, including a bidirectional gating unit (BiGRU) module and a temporal attention module.

[0159] For action recognition tasks, we use contextual information in the temporal space to perform temporal modeling. Through temporal modeling, we obtain the correlation characteristics of actions in the time axis dimension and deduce the possible execution of the next action.

[0160] The timing modeling part in this embodiment is implemented using a BiGRU based on GRU module augmentation. For the GRU module, the input timing feature vector is... ,in Its recursive update process is represented as:

[0161]

[0162] Among them, the results Represents the hidden state at time t, input This represents the hidden state at the previous time step; the GRU module consists of two parts: the update gate and the reset gate. The calculation processes for the two parts are as follows:

[0163]

[0164] in, For the Sigmoid activation process, This represents element-wise multiplication. and All are learnable parameter matrices. The GRU is expanded to BiGRU, increasing the ability to simultaneously learn patterns from both forward and reverse sequences, thus extracting more feature representations from the point cloud.

[0165] To address the varying importance of different channels in a temporal feature sequence, a temporal attention structure is designed to enhance the temporal features. The specific implementation is as follows:

[0166] Suppose a query is performed, where the key and value are obtained through linear transformations:

[0167]

[0168] in, The feature sequence output by BiGRU Given a learnable parameter matrix, the attention weights are calculated using the scaled dot product attention formula below:

[0169]

[0170] in, For vectors Dimension size, These are the channel weights for the time series.

[0171] use The final weighted time series features are calculated.

[0172] like Figure 4 As shown, the BiGRU module performs bidirectional temporal modeling on the 256-dimensional features of consecutive frames and outputs the 256-dimensional temporal features after splicing the bidirectional hidden states.

[0173] The temporal attention module uses a multi-head self-attention mechanism to weight the dependencies between frame features, outputting 256-dimensional enhanced temporal features and strengthening the weight of key action frames before the fall.

[0174] S400. The spatial features and temporal features are concatenated and processed to obtain shared features. Based on the shared features, dangerous action behaviors are classified and fall risk is assessed.

[0175] In some embodiments, such as Figure 5 As shown, the step of concatenating and processing the spatial features and temporal features to obtain shared features includes:

[0176] The spatial features and temporal features are concatenated and then input into a fully connected layer. The model is trained using the GeLU nonlinear loss function and processed by a Dropout layer. Some neurons are randomly "deactivated" to prevent overfitting and improve generalization ability. This part is the model's "basic feature processing unit" and provides shared features for the two subsequent tasks (classification and risk assessment).

[0177] After sharing features, the model splits into two branches to complete the dangerous action behavior classification task and the fall risk assessment task, respectively. The shared features are input into the classification network, which includes fully connected layers and a softmax activation function. The output is transformed into a category probability distribution (e.g., determining whether the current human action behavior belongs to the categories of imbalance, staggering, prolonged sitting, wandering, etc.), and finally outputs the classification result (Out class).

[0178] The shared features are input into the risk assessment network, which sequentially includes a fully connected layer, a Gel non-linear activation function, a Dropout layer to prevent overfitting, a fully connected layer to further refine the features, and a Sigmoid activation function. The output is compressed to between 0 and 1 (representing the "continuous probability value" of the fall risk, for example, 0.8 represents an 80% risk level), and finally outputs the risk value (Out risk).

[0179] The classification results and risk values ​​are combined to generate a final decision. Risk calibration integrates and adjusts the classification results (Outclass) and risk values ​​(Outrisk) to make the risk corresponding to the classification more accurate. Adaptive decision-making outputs the final decision based on the calibrated results.

[0180] like Figure 5 As shown, the basic feature processing adaptation layer performs dimensionality reduction and nonlinear adaptation calculations on the input 256-dimensional spatiotemporal fusion features to reduce redundant feature interference, thereby improving the computational efficiency of subsequent tasks and outputting 128 dimensions.

[0181] The learning process employs a two-branch layer in parallel, balancing action category recognition with fall risk intensity quantification. Branch A outputs the probability distribution of various dangerous actions, while branch B calculates the fall risk probability (0-1).

[0182] The risk calibration layer combines the probability of dangerous action categories to correct the original risk value, improving the accuracy of quantification. If the probability of actions like "unbalanced backward lean / rapid forward lean" in branch A is >0.8, indicating a high-risk action, the risk probability is calibrated upward (original probability * 1.2), amplifying the risk weight. If medium-risk actions such as "body swaying / staggering gait" are identified, the risk probability is finely calibrated (original probability + 0.2). When low-risk actions such as "slow bending / small turning" are identified, the risk probability is calibrated downward (original probability * 0.8), reducing the probability of misjudgment. The calibration coefficients can be obtained through training with historical data, such as statistically analyzing the "probability of actually falling after a certain action," establishing a mapping table between actions and probability corrections.

[0183] The adaptive decision-making layer employs a low-threshold trigger strategy for high-risk actions (e.g., an alert is triggered when the calibrated probability is ≥0.6); for medium-risk actions, a medium threshold is used (e.g., an alert is triggered when the probability is ≥0.75), while also incorporating the "action duration" condition (e.g., shaking for more than 2 seconds); and for low-risk actions, a high threshold is used (e.g., an alert is triggered when the probability is ≥0.9). Scene-adaptive logic can also be introduced, such as further lowering the threshold in a "nighttime + high-risk action" scenario and raising the threshold in a "still body + low-risk action" scenario. The thresholds and probability calibration coefficients are continuously updated using "time + state + behavior" data from real-world scenarios.

[0184] The method provided by this invention has been validated in action classification on multiple public datasets, as follows: Figures 6-8 As shown.

[0185] To overcome the shortcomings of existing fall warning technologies, this invention provides a fall risk warning method based on millimeter-wave radar real-time point cloud enhancement and spatiotemporal attention mechanism. It realizes a three-level processing flow of "real-time point cloud enhancement - spatiotemporal feature fusion - dangerous action prediction", and achieves non-contact, all-weather, and highly accurate fall risk warning.

[0186] Example 2

[0187] A human fall risk warning system applies the method described above. For a detailed description of the method, please refer to the corresponding description in the above method embodiments; it will not be repeated here. Figure 9 As shown, the system 500 includes a radar data acquisition module 510, a point cloud processing enhancement module 520, a spatiotemporal feature extraction module 530, and a fall risk decision-making module 540; among which,

[0188] The radar data acquisition module is used to acquire point cloud data of the human body within the monitoring area and process the point cloud data to form an original point cloud sequence.

[0189] like Figure 3As shown, this embodiment uses a millimeter-wave radar sensor 1 (such as TI's IWR6843AOP). A single millimeter-wave radar sensor can effectively cover an area of ​​approximately 16 square meters. Its built-in antenna is responsible for transmitting FMCW signals and receiving target echoes. The echoes are mixed to obtain intermediate frequency signals, which are then sampled and converted into digital signals by an ADC. Subsequently, the module identifies the real target through two-dimensional FFT processing combined with the CFAR detection algorithm, and calculates the target angle based on the antenna array. Finally, the point cloud information containing three-dimensional coordinates (x, y, z) and signal-to-noise ratio (SNR) is integrated into a TLV format data packet for output.

[0190] The point cloud processing enhancement module is used to process the original point cloud sequence to enhance the point cloud density. Optionally, the point cloud processing enhancement module receives data packets sent by the radar data acquisition module via wired or wireless means and performs protocol parsing, and achieves point cloud density enhancement image through clustering noise removal and dual sliding window frame stacking densification operation.

[0191] The spatiotemporal feature extraction module is used to extract spatial and temporal features from the point cloud enhancement frame;

[0192] The fall risk decision-making module is used to concatenate and process the spatial features and temporal features to obtain shared features, and to classify dangerous actions and behaviors and assess fall risks based on the shared features.

[0193] Based on the technical solutions of the above embodiments, optionally, the step of processing the point cloud data to form an original point cloud sequence includes:

[0194] Perform Fast Fourier Transform and Constant False Alarm Rate (CFAR) detection operations on the collected raw data to extract the 3D coordinates and intensity of the target point cloud, forming the original point cloud sequence:

[0195]

[0196] in, Total number of frames Indicates the first Frame points The feature dimension of a point.

[0197] Based on the technical solutions of the above embodiments, optionally, the step of processing the original point cloud sequence to enhance the point cloud density includes:

[0198] A multi-objective clustering algorithm was used to remove noise and ghosting from the original point cloud sequence.

[0199] A dual sliding window strategy, employing frame stacking sliding window and temporal sampling sliding window, is used to enhance point cloud density.

[0200] Based on the technical solution of the above embodiments, optionally, the step of using a multi-objective clustering algorithm to remove noise and ghosting from the original point cloud sequence specifically involves predicting the target trajectory through Kalman filtering, associating the new frame point cloud with the historical trajectory, allocating effective point clouds to the target trajectory, removing free noise points and ghosting points caused by multipath effects, and outputting the purified point cloud sequence.

[0201] Based on the technical solutions of the above embodiments, optionally, the step of enhancing point cloud density using a dual sliding window strategy of frame stacking sliding window and temporal sampling sliding window includes:

[0202] Based on the current frame, the previous frame... The frame and the current frame are stored in the edge memory;

[0203] Merge consecutive frame point clouds within a frame stacking sliding window into an enhanced frame using stacking operations:

[0204]

[0205] in, Indicates the length of the frame stacking sliding window. Indicates the frame stacking sliding window step size. This represents a stack operation of multiple point cloud frames within a sliding window, performed on the augmented frame. The point count is fixed. If the point count exceeds a preset value, random downsampling is performed. If the point count does not exceed the preset value, zero padding is performed to obtain the enhanced frame sequence.

[0206]

[0207] Extract consecutive enhanced frames from the enhanced frame sequence as temporal samples, the first... Each video clip is represented as:

[0208]

[0209] in, Indicates the length of the timing sampling sliding window. Indicates the step size of the timing sampling sliding window;

[0210] The final constructed dataset:

[0211] .

[0212] Based on the technical solution of the above embodiments, optionally, the step of extracting spatial features from the point cloud enhancement frame includes:

[0213] For a given set of point clouds At the processing point First, the KNN clustering algorithm is used in the feature space to find its... Let there be 3 adjacent points, denoted as _ . ;

[0214] Using the currently selected point and the found For each neighboring point, the EdgeConv convolution method is used to find the edge features belonging to the current point. The update formula for EdgeConv is expressed as:

[0215]

[0216] in, For point In the Layer feature representation, The relative characteristic differences between adjacent points and the center point. In EdgeConv, a multilayer perceptron is used to perform non-linear mapping on the convolution results;

[0217] Max pooling is used to optimize the convolution result in the neighborhood. Aggregation is performed to ensure local invariance between features.

[0218] Based on the technical solution of the above embodiments, optionally, a mapping transformation is performed before entering the EdgeConv layer for KNN calculation:

[0219]

[0220] Use the flipped Replace the original strength information Perform KNN calculations to satisfy the requirement that smaller values ​​are more likely to be selected as neighboring points;

[0221] In the feature construction stage, the original intensity information is used. To ensure that the data content is not distorted: for the coordinate dimension, adopt... The method constructs the difference of convolutions; for the intensity dimension, it adopts... The convolution results are constructed using a method that emphasizes the intensity dimension of features by averaging. After calculation, the features are then concatenated. And finally update:

[0222] .

[0223] Based on the technical solution of the above embodiments, optionally, in the data preprocessing construction, a Boolean mask is used to mark all points with zero dimensions. And the masked points are ignored during KNN construction and convolution. This is done to remove the effect of all zeros on the convolution result.

[0224] Based on the technical solution of the above embodiments, optionally, the step of extracting spatial features from the point cloud enhancement frame further includes:

[0225] Construct an attention module based on the intensity dimension, using intensity values ​​as adjustment factors to dynamically adjust the weights between channels at each point:

[0226]

[0227] in, Indicates the intensity value. Indicates channel weight, Use the Sigmoid activation function; For learnable scaling factors, This represents a normalization operation, used to scale the intensity of different value ranges to a relatively fixed interval. The calculation formula is as follows:

[0228]

[0229] in, The pre-set pre-scaling factor;

[0230] Finally, the channel attention weights are obtained. The input features are processed through bitwise multiplication. Weighted average:

[0231] .

[0232] Based on the technical solution of the above embodiments, optionally, the step of extracting temporal features from the point cloud enhancement frame includes:

[0233] For action recognition tasks, we use contextual information in the temporal space to perform temporal modeling. Through temporal modeling, we obtain the correlation characteristics of actions in the time axis dimension and deduce the possible execution of the next action.

[0234] Based on the technical solutions of the above embodiments, optionally, the time series modeling is implemented using BiGRU based on GRU module augmentation: for the GRU module, the input time series feature vector is... ,in Its recursive update process is represented as:

[0235]

[0236] Among them, the results Represents the hidden state at time t, input This represents the hidden state at the previous time step; the GRU module consists of two parts: the update gate and the reset gate. The calculation processes for the two parts are as follows:

[0237]

[0238] in, For the Sigmoid activation process, This represents element-wise multiplication. and All of these are learnable parameter matrices.

[0239] Based on the technical solution of the above embodiments, optionally, the step of extracting temporal features from the point cloud enhancement frame further includes:

[0240] Adding a temporal attention structure as a feature enhancement at the temporal feature level:

[0241] Suppose a query is performed where the key and value are obtained through linear transformations:

[0242]

[0243] in, The feature sequence output by BiGRU Given a learnable parameter matrix, the attention weights are calculated using the scaled dot product attention formula:

[0244]

[0245] in, For vectors Dimension size, These are the channel weights for the time series.

[0246] use The final weighted time series features are calculated.

[0247] Based on the technical solution of the above embodiments, optionally, the step of concatenating and processing the spatial features and the temporal features to obtain shared features includes:

[0248] The spatial features and the temporal features are concatenated and then input into a fully connected layer. The model is trained using the GeLU nonlinear loss function and processed by a Dropout layer to obtain shared features.

[0249] Based on the technical solution of the above embodiments, optionally, the step of classifying dangerous actions and behaviors and assessing fall risk based on the shared features includes:

[0250] The shared features are input into a classification network, which includes fully connected layers and a softmax activation function. The output is transformed into a class probability distribution, and the final classification result is output.

[0251] The shared features are input into a risk assessment network, which sequentially includes a fully connected layer, a Gel nonlinear activation function, a Dropout layer, a fully connected layer, and a Sigmoid activation function. The output is compressed to between 0 and 1, and the final output is a risk value.

[0252] The classification results and the risk values ​​are combined to generate a final decision.

[0253] To overcome the shortcomings of existing fall warning technologies, this invention provides a human fall risk warning system based on millimeter-wave radar real-time point cloud enhancement and spatiotemporal attention mechanism. It realizes a three-level processing flow of "real-time point cloud enhancement - spatiotemporal feature fusion - dangerous action prediction", and achieves non-contact, all-weather, and highly accurate fall risk warning.

[0254] Example 3

[0255] A computer device 600, such as Figure 10 As shown, the system includes a memory 610, a processor 620, and a computer program 630 stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a method for early warning of human fall risks. For a detailed description of the method, please refer to the corresponding description in the above method embodiments, which will not be repeated here.

[0256] Example 4

[0257] A computer-readable storage medium, such as Figure 11 As shown, a computer program is stored thereon, which, when executed by a processor, implements the steps of a method for early warning of human fall risks. For a detailed description of the method, please refer to the corresponding description in the above method embodiments, and will not be repeated here.

[0258] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0259] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

[0260] The apparatus, computer device, and non-volatile computer storage medium and method provided in the embodiments of this specification are corresponding. Therefore, the apparatus, computer device, and non-volatile computer storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding apparatus, computer device, and non-volatile computer storage medium will not be repeated here.

[0261] Those skilled in the art will also know that, besides implementing the controller in the form of purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller take the form of logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices included within it for implementing various functions can also be considered structures within that hardware component. Alternatively, the devices for implementing various functions can be considered as both software units implementing the method and structures within a hardware component.

[0262] The systems, apparatuses, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above apparatuses are described separately as various units based on their functions. Of course, when implementing one or more embodiments of this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0263] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0264] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0265] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0266] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0267] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0268] This specification may be described in the general context of computer-executable instructions, such as program units, that are executed by a computer. Generally, program units include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification may also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program units may reside in local and remote computer storage media, including storage devices.

[0269] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0270] The above description is merely an embodiment of this specification and is not intended to limit the scope of one or more embodiments of this specification. Various modifications and variations can be made to one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of one or more embodiments of this specification.

Claims

1. A method for early warning of human fall risk, characterized in that, Includes the following steps: Acquire point cloud data of the human body within the monitoring area, and process the point cloud data to form an original point cloud sequence; The original point cloud sequence is processed to enhance the point cloud density; Spatial and temporal features are extracted from the point cloud augmentation frame; The spatial features and temporal features are concatenated and processed to obtain shared features, and dangerous action behaviors are classified and fall risk is assessed based on the shared features; The step of processing the original point cloud sequence to enhance the point cloud density includes: A multi-objective clustering algorithm was used to remove noise and ghosting from the original point cloud sequence. A dual sliding window strategy, employing frame stacking sliding window and temporal sampling sliding window, is used to enhance point cloud density; The steps for enhancing point cloud density using a dual sliding window strategy that combines frame stacking sliding window and temporal sampling sliding window include: Based on the current frame, the previous frame... The frame and the current frame are stored in the edge memory; Merge consecutive frame point clouds within a frame stacking sliding window into an enhanced frame using stacking operations: in, Indicates the length of the frame stacking sliding window. Indicates the frame stacking sliding window step size. This represents a stack operation of multiple point cloud frames within a sliding window, performed on the augmented frame. The point count is fixed. If the point count exceeds a preset value, random downsampling is performed. If the point count does not exceed the preset value, zero padding is performed to obtain the enhanced frame sequence. Extract consecutive enhanced frames from the enhanced frame sequence as temporal samples, the first... Each video clip is represented as: in, Indicates the length of the timing sampling sliding window. Indicates the window step size of the timing sampling sliding window; The final constructed dataset: ； The steps for extracting temporal features from point cloud augmented frames include: For action recognition tasks, we use contextual information in the temporal space to perform temporal modeling, obtain the correlation characteristics of actions in the time axis dimension through temporal modeling, and deduce the possible execution of the next action. The time series modeling is implemented using a BiGRU based on GRU module augmentation: for the GRU module, the input time series feature vector is... ,in Its recursive update process is represented as: Among them, the results Represents the hidden state at time t, input This represents the hidden state at the previous time step; the GRU module consists of two parts: the update gate and the reset gate. The calculation processes for the two parts are as follows: in, For the Sigmoid activation process, This represents element-wise multiplication. and All are learnable parameter matrices; The steps for extracting temporal features from point cloud augmented frames also include: Adding a temporal attention structure as a feature enhancement at the temporal feature level: Suppose a query is performed where the key and value are obtained through linear transformations: in, The feature sequence output by BiGRU Given a learnable parameter matrix, the attention weights are calculated using the scaled dot product attention formula: in, For vectors Dimension size, These are the channel weights for the time series. use The final weighted time series features are calculated.

2. The method for early warning of human fall risk as described in claim 1, characterized in that, The step of processing the point cloud data to form the original point cloud sequence includes: Perform Fast Fourier Transform and Constant False Alarm Rate (CFAR) detection operations on the collected raw data to extract the 3D coordinates and intensity of the target point cloud, forming the original point cloud sequence: in, Total number of frames Indicates the first Frame points The feature dimension of a point.

3. The method for early warning of human fall risk as described in claim 1, characterized in that, The step of using a multi-objective clustering algorithm to remove noise and ghosting from the original point cloud sequence specifically involves predicting the target trajectory using Kalman filtering, associating the new frame point cloud with the historical trajectory, allocating valid point clouds to the target trajectory, removing detached noise points and ghosting points caused by multipath effects, and outputting the purified point cloud sequence.

4. The method for early warning of human fall risk as described in claim 1, characterized in that, The steps for spatial feature extraction from point cloud augmented frames include: For a given set of point clouds At the processing point First, the KNN clustering algorithm is used in the feature space to find its... Let there be 3 adjacent points, denoted as _ . ; Using the currently selected point and the found For each neighboring point, the EdgeConv convolution method is used to find the edge features belonging to the current point. The update formula for EdgeConv is expressed as: in, For point In the Layer feature representation, The relative characteristic differences between adjacent points and the center point. In EdgeConv, a multilayer perceptron is used to perform non-linear mapping on the convolution results; Max pooling is used to optimize the convolution result in the neighborhood. Aggregation is performed to ensure local invariance between features.

5. A method for early warning of human fall risk as described in claim 4, characterized in that, Before performing KNN computation in the EdgeConv layer, a mapping transformation is performed: Use the flipped Replace the original strength information Perform KNN calculations to satisfy the requirement that smaller values ​​are more likely to be selected as neighboring points; In the feature construction stage, the original intensity information is used. To ensure that the data content is not distorted: for the coordinate dimension, adopt... The difference between convolutions is constructed using a specific method; for the intensity dimension, the difference is calculated using... The convolutional results are constructed using a method that emphasizes the intensity dimension of features by averaging. After calculation, the features are then concatenated. And finally update: 。 6. A method for early warning of human fall risk as described in claim 5, characterized in that, In the data preprocessing construction, a Boolean mask is used to mark all points with a dimension of zero. And the masked points are ignored during KNN construction and convolution. This is done to remove the effect of all zeros on the convolution result.

7. A method for early warning of human fall risk as described in claim 5, characterized in that, The spatial feature extraction steps for point cloud augmentation frames also include: Construct an attention module based on the intensity dimension, using intensity values ​​as adjustment factors to dynamically adjust the weights between channels at each point: in, Indicates the intensity value. Indicates channel weight, Use the Sigmoid activation function; For learnable scaling factors, This represents a normalization operation, used to scale the intensity of different value ranges to a relatively fixed interval. The calculation formula is as follows: in, The pre-set pre-scaling factor; Finally, the channel attention weights are obtained. The input features are processed through bitwise multiplication. Weighted average: 。 8. A method for early warning of human fall risk as described in claim 1, characterized in that, The step of concatenating and processing the spatial features and temporal features to obtain shared features includes: The spatial features and the temporal features are concatenated and then input into a fully connected layer. The model is trained using the GeLU nonlinear loss function and processed by a Dropout layer to obtain shared features.

9. A method for early warning of human fall risk as described in claim 8, characterized in that, The steps for classifying dangerous actions and behaviors and assessing fall risk based on the shared features include: The shared features are input into a classification network, which includes fully connected layers and a softmax activation function. The output is transformed into a class probability distribution, and the final classification result is output. The shared features are input into a risk assessment network, which sequentially includes a fully connected layer, a Gel nonlinear activation function, a Dropout layer, a fully connected layer, and a Sigmoid activation function. The output is compressed to between 0 and 1, and the final output is a risk value. The classification results and the risk values ​​are combined to generate a final decision.

10. A human fall risk warning system, using the method as described in any one of claims 1 to 9, characterized in that: It includes a radar data acquisition module, a point cloud processing enhancement module, a spatiotemporal feature extraction module, and a fall risk decision-making module; among which, The radar data acquisition module is used to acquire point cloud data of the human body within the monitoring area and process the point cloud data to form an original point cloud sequence. The point cloud processing enhancement module is used to process the original point cloud sequence to enhance the point cloud density; The spatiotemporal feature extraction module is used to extract spatial and temporal features from the point cloud enhancement frame; The fall risk decision-making module is used to concatenate and process the spatial features and temporal features to obtain shared features, and to classify dangerous actions and behaviors and assess fall risks based on the shared features.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 9.

Images