A Smart Monitoring Method for Children's Falls Based on Motion Capture

Abstract

This invention discloses a child fall monitoring method based on motion capture, belonging to the field of intelligent protection technology. The method includes: acquiring video of a child's activities and performing posture estimation to form a continuous sequence of activity frames; performing dual-channel parallel processing on the continuous activity frame sequence to obtain skeletal motion features and trunk motion features; performing spectral analysis and instantaneous velocity calculation on the skeletal motion features to obtain the velocity of lower limb bones, and simultaneously performing dense optical flow calculation on the trunk motion features to generate trunk optical flow velocity; based on the child's fall risk value, performing real-time posture tracking and surrounding environment scanning of the child's activity state, and collecting motion trajectory data; and performing a hazard classification response based on the motion trajectory data to output an intelligent protection plan. This invention, through a dual-channel parallel processing architecture and a spatiotemporal behavior analysis model, achieves collaborative perception of multi-dimensional motion features and accurate capture of the imbalance process of fall precursors, improving the timeliness of fall warnings.

Description

Technical Field

[0001] This invention relates to the field of intelligent protection technology, and in particular to an intelligent monitoring method for children's falls based on motion capture. Background Technology

[0002] With the rapid development of computer vision, motion capture-based intelligent monitoring methods have been widely applied in the field of intelligent protection, especially in child safety monitoring scenarios, where falls, as a common accidental injury, require accurate identification and real-time intervention. In recent years, deep learning-based motion capture methods, dense optical flow methods, and spectral analysis have been applied to the motion dynamics analysis of fall warning systems, enabling intelligent monitoring to leap from "passive recording" to "active early warning," thus improving the real-time performance and accuracy of the monitoring process.

[0003] However, existing fall monitoring technologies for children still have some limitations. First, current methods rely on single-modal features (such as only skeletal point trajectories or only optical flow information) for judgment, ignoring the coupling relationship between the skeleton and trunk during human movement. This makes it difficult to distinguish between high-risk imbalance movements and normal activities (such as rapid squatting or jumping), resulting in a high false alarm rate in fall monitoring. Second, existing behavior recognition models only analyze motion features from a single time or space dimension, lacking the ability to fuse spatiotemporal features and model dynamic correlations. This leads to insufficient capture of risk features in continuous movements, affecting the timeliness of fall risk warnings. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a motion capture-based intelligent monitoring method for children's falls to solve the problems of high false alarm rates and insufficient timeliness of fall risk warnings.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a method for intelligent monitoring of children's falls based on motion capture, which includes: acquiring videos of children's activities and performing posture estimation to form a continuous sequence of activity frames; and performing dual-channel parallel processing on the continuous sequence of activity frames to obtain skeletal motion features and trunk motion features.

[0008] Spectral analysis and instantaneous velocity calculation are performed on the skeletal motion characteristics to obtain the point velocities of the lower limb bones. Simultaneously, dense optical flow calculation is performed on the trunk motion characteristics to generate the trunk optical flow velocity.

[0009] The optical flow velocity of the torso and the point velocity of the lower limb bones are input into the spatiotemporal behavior analysis model. The spatial convolutional layer quantifies the directional differences and assesses the limb coordination, while the temporal convolutional layer performs continuous state tracking and trend analysis, and outputs the child's fall risk value.

[0010] Based on the child's fall risk value, the system performs real-time posture tracking and environmental scanning of the child's activity status, and collects movement trajectory data; based on the movement trajectory data, it performs emergency classification response and outputs intelligent protection solutions.

[0011] As a preferred embodiment of the intelligent monitoring method for children's falls based on motion capture described in this invention, the formation of a continuous activity frame sequence refers to extracting image frames from the child's activity video, performing pose estimation and temporal alignment to form a continuous activity frame sequence.

[0012] As a preferred embodiment of the motion capture-based intelligent monitoring method for children's falls according to the present invention, the acquisition of skeletal motion features and trunk motion features specifically includes the following steps.

[0013] Motion vector analysis is performed on a continuous sequence of active frames to obtain skeletal joint displacement parameters and trunk region motion parameters.

[0014] In the first channel, the least squares method is used to fit the bone joint displacement parameters to a smooth trajectory to form a bone motion curve; the curvature extreme points of the bone motion curve are extracted and dynamic time warped to generate bone motion features.

[0015] In the second channel, energy accumulation and integration are performed on the motion parameters of the torso region to generate the torso motion energy distribution. Wavelet transform and low-pass filtering are then applied to the torso motion energy distribution to obtain the torso motion features.

[0016] As a preferred embodiment of the intelligent monitoring method for child falls based on motion capture described in this invention, the step of acquiring the velocity of lower limb skeletal points specifically includes the following steps.

[0017] Fast Fourier Transform is applied to perform spectral analysis on skeletal motion features to generate frequency domain motion features, and inverse transform is performed to reconstruct the time domain motion vector.

[0018] The instantaneous velocity of the motion vector in the time domain is calculated using the central difference formula to obtain the velocity of the lower limb bones.

[0019] As a preferred embodiment of the intelligent monitoring method for children's falls based on motion capture described in this invention, the generation of trunk optical flow velocity refers to generating trunk optical flow velocity by performing amplitude normalization and dense optical flow calculation on trunk motion features using the Farneback algorithm.

[0020] As a preferred embodiment of the motion capture-based intelligent monitoring method for children's falls according to the present invention, the specific construction process of the spatiotemporal behavior analysis model is as follows:

[0021] Spatial convolutional layers are constructed using graph convolutional networks, and temporal convolutional layers are constructed using temporal convolutional networks.

[0022] An attention mechanism is applied to perform cross-layer feature fusion and residual stacking on spatial and temporal convolutional layers to construct a spatiotemporal behavior analysis model.

[0023] As a preferred embodiment of the motion capture-based intelligent monitoring method for children's falls according to the present invention, the spatial convolutional layer performs orientation difference quantification and limb coordination assessment, specifically including the following steps.

[0024] The optical flow velocity of the torso and the point velocity of the lower limb bones were subjected to dimensional compression and then input into the spatiotemporal behavior analysis model.

[0025] The spatial convolutional layer quantifies the directional difference between the optical flow velocity of the torso and the velocity of the lower limb bones through a graph convolutional network, forming a motion direction difference matrix. It then evaluates the limb coordination based on the motion direction difference matrix and generates a spatial imbalance vector.

[0026] As a preferred embodiment of the motion capture-based intelligent monitoring method for children's falls according to the present invention, the temporal convolutional layer performs continuous state tracking and trend analysis, specifically including the following steps.

[0027] The temporal convolutional layer continuously tracks the optical flow velocity of the torso and the point velocity of the lower limb bones through dilated causal convolution, generating a spatiotemporal feature sequence.

[0028] Temporal dependency modeling and trend analysis are performed on spatiotemporal feature sequences to form a time evolution vector.

[0029] As a preferred embodiment of the intelligent monitoring method for child falls based on motion capture described in this invention, the output of the child fall risk value refers to the feature weighted fusion of the spatial imbalance vector and the temporal evolution vector, and the application of the Sigmoid function for probability normalization to output the child fall risk value.

[0030] As a preferred embodiment of the motion capture-based intelligent monitoring method for child falls described in this invention, the output intelligent protection scheme specifically includes the following steps.

[0031] Match decision parameters to children's fall risk values ​​to obtain activity monitoring strategies;

[0032] According to the activity monitoring strategy, the child's activity status is tracked in real time and the surrounding environment is scanned using depth cameras and LiDAR, while motion trajectory data is collected simultaneously using inertial measurement instruments.

[0033] Based on a preset three-level threshold, the motion trajectory data is classified into different levels of urgency and a smart protection plan is output.

[0034] The beneficial effects of this invention are as follows: By extracting skeletal and trunk motion features through a dual-channel parallel processing architecture and performing instantaneous velocity and dense optical flow calculations, collaborative perception and fusion analysis of multi-dimensional motion features are achieved, thereby reducing the false alarm rate caused by motion confusion. Employing a spatiotemporal behavior analysis model for quantifying motion direction differences and analyzing continuous state changes enables precise capture of the gradual imbalance evolution process preceding a fall, thus improving the timeliness of early warnings and securing a critical window for protective intervention. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of a motion capture-based intelligent monitoring method for children's falls.

[0037] Figure 2 A flowchart illustrating the operation of a dual-channel parallel processing architecture.

[0038] Figure 3 A flowchart for generating a child's fall risk value.

[0039] Figure 4 This is a flowchart for outputting an intelligent protection solution. Detailed Implementation

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0043] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a child fall monitoring method based on motion capture, comprising the following steps:

[0044] S1. Collect videos of children's activities, perform motion capture and temporal segmentation to form a continuous sequence of activity frames; process the continuous sequence of activity frames in parallel through dual channels to obtain skeletal motion features and trunk motion features.

[0045] S1.1 Collect and preprocess videos of children's activities.

[0046] High-definition grid cameras are used to capture videos of children's activities, which are then preprocessed. Gamma correction is used to normalize the brightness of the videos to mitigate pixel value deviations caused by non-uniform lighting. Histogram equalization is applied to enhance contrast and improve the visibility of human body contours. Median filtering is used simultaneously to denoise the videos, suppressing noise and background interference and improving the stability of subsequent motion analysis.

[0047] S1.2 Extract image frames from the preprocessed children's activity video, perform pose estimation and temporal alignment, and obtain a continuous activity frame sequence.

[0048] Read the preprocessed video stream of children's activities and extract RGB image frames at a fixed frame rate (e.g., 30fps); normalize the size and color space of the RGB image frames, for example, adjust the size to 256×256 pixels, convert the color values ​​to [0,1], and output standard image frames.

[0049] Using the Rigify plugin in Blender software, the initial positions of key human joints (such as hip, knee, and ankle joints) are identified in standard image frames, and perspective transformation is performed to achieve the conversion from two-dimensional image space to three-dimensional coordinate system, obtaining pixel-level joint coordinates.

[0050] Pixel-level joint coordinates are linearly interpolated and timestamped to form motion posture parameters. The motion posture parameters are then time-aligned using the PTP clock synchronization protocol to obtain an aligned motion posture sequence. Kalman filtering is used to smooth the data and suppress noise in the aligned motion posture sequence, outputting a continuous activity frame sequence. Each frame of the continuous activity frame sequence contains a timestamp and pixel-level joint coordinates, providing a structured data foundation for subsequent time series analysis.

[0051] S1.3 Perform motion vector analysis on the continuous active frame sequence to obtain the bone joint displacement parameters and trunk region motion parameters.

[0052] Background removal is performed on the continuous active frame sequence, and dilation and erosion operations are performed through morphological closing operations to ensure that hole noise is eliminated while maintaining regional connectivity, forming a complete foreground image. Gray-level thresholds are applied to divide the complete foreground image into regions. For example, the gray values ​​of the complete foreground image are counted. When the gray value is within the typical bone imaging range (e.g., 200, 255), it is divided into a bone highlight region. When the gray value is within the trunk soft tissue imaging range (e.g., 50, 199), it is divided into a trunk mid-gray region.

[0053] It should be noted that the development interval is based on the quantile definition of the cumulative distribution of historical gray values.

[0054] Motion vector analysis is performed on the highlighted areas of the skeleton and the gray areas of the torso. Furthermore, the motion gradient features of the highlighted areas of the skeleton are extracted by the Sobel operator and vector synthesis is performed to obtain the skeleton motion vector. The skeleton motion vector is integrated continuously over time using a system of first-order ordinary differential equations to obtain the cumulative displacement. The cumulative displacement is then discretized and sampled to output the skeleton joint displacement parameters.

[0055] It should be noted that the Sobel operator, through the Sobel function call in the OpenCV library, extracts motion gradient features (including sudden stops, rapid turns, etc.) by performing horizontal and vertical convolutions on the highlighted areas of the skeleton and calculating gradient magnitudes; vector synthesis refers to the process of combining the direction and magnitude of motion gradient features through a vector superposition algorithm; the first-order ordinary differential equation system is a set of mathematical equations describing changes in motion state, which is directly called through the odeint function in the SciPy library.

[0056] The motion amplitude parameters in the gray area of ​​the torso are statistically analyzed using a sliding window, and vector recombination is performed to obtain the motion amplitude vector. Singular value decomposition is applied to orthogonally decompose the motion amplitude vector, and the first k motion amplitude vectors are extracted and linearly combined to obtain the main motion components. The main motion components are then vector synthesized to generate the torso motion vector. According to the human kinematics constraint rules, the torso motion vector is mapped to spatial coordinates and mapped to the coordinate system of the human spine center to form the torso region motion parameters.

[0057] It should be noted that the human kinematic constraint rules are defined based on the anatomical characteristics of the physiological curvature of the spine.

[0058] S1.4 In the first channel, the least squares method is used to smooth the trajectory of the bone joint displacement parameters to form the bone motion curve; the curvature extreme points of the bone motion curve are extracted and dynamic time warped to generate bone motion features.

[0059] The least squares method is used to fit the smooth trajectory of the skeletal joint displacement parameters. Furthermore, cubic spline interpolation is performed on the skeletal joint displacement parameters to obtain a smooth displacement parameter sequence. The smooth displacement parameter sequence is then processed by piecewise linear approximation to obtain the initial motion trajectory curve. Simultaneously, kinematic constraint optimization is performed on the initial motion trajectory curve to output the skeletal motion curve.

[0060] It should be noted that piecewise linear approximation refers to the process of segmenting and linearly fitting the smooth displacement parameter sequence according to a fixed time window (e.g., 1 second / 30 frames); kinematic constraint optimization refers to the process of correcting the angle of the initial motion trajectory curve according to the human joint range of motion limits (e.g., hip flexion and extension 0-120°, knee flexion 0-130°).

[0061] The maximum curvature is obtained by calculating the first and second derivatives of the skeletal motion curve using the curvature formula. The specific mathematical formula is as follows.

[0062] ;

[0063] in, Indicates the maximum curvature. Indicates the skeletal motion curve at The first derivative in the axial direction, Indicates the skeletal motion curve at The second derivative along the axial direction, Indicates the skeletal motion curve at The first derivative in the axial direction, Indicates the skeletal motion curve at Second derivative in the axial direction;

[0064] It should be noted that, First derivative in the axial direction and The first derivative along the axial direction passes through... The directional difference operation is performed on the skeletal motion curve using the y-axis or the x-axis to obtain the curve. Second derivative in the axial direction and The second derivative along the axial direction passes through... The skeletal motion curve is obtained by performing a second-order central difference operation on the y-axis or the x-axis.

[0065] The coordinates of the point where the curvature is at its maximum in the skeletal motion curve are taken as the curvature extremum point; the curvature extremum point not only reflects the intensity of the transition of joint motion, but also characterizes the suddenness of the action.

[0066] Dynamic time warping is performed on the curvature extrema. Furthermore, the timestamps of the curvature extrema are extracted and compared with the standard timestamps to obtain the time offset of the extrema. The Sigmoid function is applied to nonlinearly scale the time offset of the extrema to obtain time-aligned curvature extrema.

[0067] It should be noted that the standard timestamp is based on the baseline time interval definition of the human movement cycle.

[0068] Principal component analysis was used to project the time-aligned curvature extrema into the feature space to form low-dimensional dense features. These low-dimensional dense features were then weighted, fused, and dimensionally compressed to obtain skeletal motion features. Skeletal motion features characterize the spatiotemporal variation patterns of joint motion and are used for subsequent motion capture and fall detection analysis.

[0069] S1.5 In the second channel, energy accumulation integration is performed on the motion parameters of the torso region to generate the torso motion energy distribution, and wavelet transform and low-pass filtering are performed on the torso motion energy distribution to obtain the torso motion features.

[0070] The Leibniz integral method is used to perform energy accumulation integration on the motion parameters of the trunk region, generating an energy accumulation value. The specific mathematical formula is as follows.

[0071] ;

[0072] in, Indicates the cumulative energy value. Indicates the current time. Indicates the start time of integration. Represents the instant of integration. Indicates the instant of integration The regional velocity vector, Indicates the instant of integration The square of the magnitude of the velocity vector of the region's motion;

[0073] It should be noted that the regional motion velocity vector is obtained by performing time difference and spatial coordinate transformation on the torso regional motion parameters; the magnitude of the regional motion velocity vector is obtained by performing Euclidean norm integration on the regional motion velocity vector.

[0074] Gaussian filtering and spatial interpolation are applied to the accumulated energy value to obtain a continuous energy distribution. The amplitude of the continuous energy distribution is then normalized to obtain the energy distribution of trunk motion.

[0075] Wavelet transform is applied to the trunk motion energy distribution. Furthermore, Daubechies wavelet basis functions are used to perform multi-resolution decomposition of the trunk motion energy distribution, separating it into high-frequency detail components and low-frequency approximate components. A Butterworth low-pass filter with a fixed cutoff frequency (e.g., 2Hz) is used to filter out the high-frequency detail components to eliminate high-frequency noise components (such as muscle tremors), while retaining the low-frequency approximate components. The low-frequency approximate components are then weighted and aggregated to form trunk motion features. Trunk motion features not only characterize the overall intensity of trunk motion but also reflect the stability and suddenness of the motion pattern.

[0076] S2. Perform spectral analysis and instantaneous velocity calculation on the skeletal motion characteristics to obtain the point velocity of the lower limb bones. Simultaneously, perform vector field modeling and dense optical flow calculation on the trunk motion characteristics to generate the trunk optical flow velocity.

[0077] S2.1 Apply Fast Fourier Transform to perform spectral analysis on the skeletal motion features, generate frequency domain motion features, and perform inverse transform reconstruction to obtain time domain motion vectors.

[0078] Z-score normalization is used to perform zero-mean processing on the skeletal motion features to eliminate baseline drift and obtain normalized skeletal motion features. The Hanning window function is applied to window the normalized skeletal motion features to reduce spectral leakage. Fast Fourier transform is used to perform frequency domain transformation to transform the normalized skeletal motion features from the time domain to the frequency domain, resulting in a complex frequency domain spectrum. The complex frequency domain spectrum characterizes the frequency distribution characteristics of the skeletal motion features.

[0079] The frequency domain spectrum in complex form is numerically normalized to obtain the normalized spectrum value. When the normalized spectrum value is greater than the energy threshold, the corresponding frequency domain spectrum in complex form is taken as the dominant frequency component. The typical human motion frequency band (such as 0.5-5Hz) in the dominant frequency component is extracted using a Butterworth bandpass filter, and the frequency band energy is integrated to obtain the frequency domain motion characteristics.

[0080] It should be noted that the energy threshold is defined based on the cumulative distribution quantile of historical normalized spectral values, with an exemplary range of [0.7, 0.9].

[0081] Inverse Fourier Transform (IFFT) is used to reconstruct the frequency domain motion features. Furthermore, phase preservation and frame overlap processing are applied to the frequency domain motion features to improve temporal continuity and obtain a preliminary reconstructed signal. The preliminary reconstructed signal is then weighted and smoothed at the edges to generate a reconstructed temporal signal. According to the kinematic projection rules, the reconstructed temporal signal is mapped to a three-dimensional spatial coordinate system through affine transformation to output a temporal motion vector.

[0082] It should be noted that frame overlap processing refers to the process of segmenting and overlapping frequency domain motion features using a Hamming window function; the kinematic projection rule is based on the definition of the motion axis of the human joint rotation center.

[0083] S2.2. The instantaneous velocity of the time-domain motion vector is calculated using the central difference formula to obtain the velocity of the lower limb bones. The specific mathematical formula is as follows.

[0084] ;

[0085] in, Indicates a time index. Indicates time The velocity of the lower limb bones. Indicates time interval, Indicates time position vector, Indicates time Position vector;

[0086] It should be noted that the position vector is obtained by extracting spatial coordinates and synthesizing vectors from the temporal motion vector.

[0087] S2.3. The Farneback algorithm is used to normalize the amplitude of the torso motion features and calculate the dense optical flow to generate the torso optical flow velocity. The specific mathematical formula is as follows.

[0088] ;

[0089] in, Indicates the optical flow velocity of the torso. Indicates the amplitude of trunk motion characteristics. The horizontal component representing the amplitude of trunk motion characteristics. The vertical component representing the amplitude of trunk motion characteristics. Represents a very small constant (used to avoid division by zero errors);

[0090] It should be noted that the amplitude of trunk motion features was obtained by normalizing the amplitude of trunk motion features using the L2 norm standardization method; the vertical and horizontal components were obtained by decomposing the amplitude of trunk motion features into directions.

[0091] S3. Input the optical flow velocity of the torso and the point velocity of the lower limb bones into the spatiotemporal behavior analysis model. The spatial convolutional layer performs directional difference quantification and limb coordination assessment, and the temporal convolutional layer performs continuous state tracking and trend analysis, outputting the child's fall risk value.

[0092] S3.1 Construct and train a spatiotemporal behavior analysis model.

[0093] In the PyTorch framework, the graph convolutional network is invoked through the torch_geometric parameter and initialized. For example, the learning rate is set to 0.001, the number of hidden layer nodes is set to 64, and the regularization coefficient is set to 0.01. After the graph convolutional network, the ReLU activation function is applied to perform non-linear transformation to enhance the feature representation capability, and dimensionality reduction projection is performed through a fully connected layer to complete the construction of the spatial convolutional layer.

[0094] The temporal convolutional network is invoked using the torch.nn.Conv1d parameter and initialized, for example, by setting the kernel size to 3, the stride to 1, and the number of convolutional layers to 3. Dilated causal convolutions are then added after the temporal convolutional network to capture long-term dependencies, and batch normalization is used to accelerate convergence, thus completing the construction of the temporal convolutional layers.

[0095] It should be noted that dilated causal convolution is a convolutional structure with a time-delayed expansion mechanism. It expands the receptive field through an exponentially increasing dilation rate, thereby capturing long-term dependencies.

[0096] An attention mechanism is used to perform cross-layer feature fusion between spatial and temporal convolutional layers to obtain spatiotemporal behavior feature vectors. The spatiotemporal behavior feature vectors are probabilistically normalized using the softmax function to obtain the feature importance weights at each time step. The feature importance weights at each time step are then weighted and backpropagated to generate residual stacked weights. Based on the residual stacked weights, the spatial and temporal convolutional layers are stacked to complete the construction of the spatiotemporal behavior analysis model.

[0097] Next, the spatiotemporal behavior analysis model is trained. Further, the historical trunk optical flow velocity and lower limb skeletal point velocity are divided into a sample set, a training set, and a validation set. On the sample set, Z-score normalization is used for feature scaling to form a normalized sample set. On the training set, the Adam optimizer is used to update the gradient of the normalized sample set, and the cross-entropy loss function is applied simultaneously to minimize the loss and obtain the optimal weight parameters. On the validation set, an early stopping mechanism is used to monitor the performance of the optimal weight parameters and obtain the validation set accuracy. When the validation set accuracy exceeds the convergence threshold for several consecutive rounds (e.g., 3 times), training terminates, and the trained spatiotemporal behavior analysis model is output simultaneously.

[0098] It should be noted that the convergence threshold is defined based on the magnitude of the change in the accuracy of the validation set, with an exemplary range of 0.001 to 0.01.

[0099] S3.2 The spatial convolutional layer performs directional difference quantification and limb coordination assessment to generate a spatial imbalance vector.

[0100] Dimensional compression is performed by globally averaging the optical flow velocity of the torso and the point velocities of the lower limb bones, and the spatiotemporal behavior analysis model is input using the PyTorch API interface.

[0101] The spatial convolutional layer quantifies the directional difference between the optical flow velocity of the torso and the velocities of the lower limb bones. Further, it concatenates the features of the torso optical flow velocity and the lower limb bone velocities to obtain fused motion features. A graph convolutional network is then used to perform graph convolution on the fused motion features to generate a spatial feature map. Local receptive fields (the neighborhood range centered on the joints in the spatial feature map) are extracted from this spatial feature map to obtain local spatial features. Finally, the directional difference measure of the local spatial features is calculated using the cosine similarity formula to obtain the difference matrix. The specific mathematical formula is as follows.

[0102] ;

[0103] in, Represents the difference matrix, The total number of dimensions representing local spatial features. Dimension index representing local spatial features and Both represent indices of spatial locations. Indicates the first The spatial location is at the first Local spatial eigenvalues ​​of dimension, Indicates the first The spatial location is at the first Local spatial eigenvalues ​​of dimension ;

[0104] It should be noted that spatial location refers to the position of local spatial features in the spatial feature map; local spatial feature values ​​are obtained by performing weighted averaging and normalization operations on the local spatial features.

[0105] The ReLU function is used to perform nonlinear activation on the heterogeneity matrix to form the motion direction difference matrix. The motion direction difference matrix contains information on the differences in motion direction between different parts, providing a mathematical basis for subsequent quantification of limb motor coordination.

[0106] The movement direction difference matrix is ​​used to assess limb coordination. Furthermore, principal component analysis is used to decompose the movement direction difference matrix into eigenvalues, and the first m eigenvalues ​​are integrated to generate a coordination feature vector. The coordination feature vector is weighted and fused, and then dimensionality reduction projection is performed through a fully connected layer to obtain a spatial imbalance vector. The spatial imbalance vector represents the degree of movement imbalance between different parts of the child's body during movement, which can help assess whether there is a potential risk of falling.

[0107] S3.3 The temporal convolutional layer performs continuous state tracking and trend analysis to form a temporal evolution vector.

[0108] Dilated causal convolution is applied to expand the receptive field of the torso optical flow velocity and the lower limb skeletal point velocity to capture the spatiotemporal correlation between the torso optical flow velocity and the lower limb skeletal point velocity, thus obtaining long-range dependency features. On the time axis, a temporal convolutional network is used to perform dilated convolution on the long-range dependency features to obtain multi-scale temporal features. The ReLU function is used to perform nonlinear transformation on the multi-scale temporal features, and batch normalization is applied to stabilize the gradient and accelerate convergence, gradually constructing a spatiotemporal feature sequence.

[0109] A gated loop mechanism is applied to perform temporal modeling of the temporal dependence of spatiotemporal feature sequences. Furthermore, an update gate is used to gating and adjust the spatiotemporal feature sequences to generate temporary hidden states. A reset gate is used to reset the temporary state vector and control forgetting, outputting the current hidden state. The temporary hidden state and the current hidden state are weighted and fused to form a temporal evolution vector. The temporal evolution vector reflects the dynamic motion characteristics and trends that change over time and is an important indicator of changes in the motion process.

[0110] S3.4. Perform feature-weighted fusion of the spatial imbalance vector and the temporal evolution vector, and apply the Sigmoid function for probability normalization to output the child's fall risk value. The specific mathematical formula is as follows:

[0111] ;

[0112] in, This indicates the risk value of a child falling. Indicates activator. The weight coefficients represent the spatial imbalance vector. Represents the spatial imbalance vector. The weight coefficients represent the time evolution vector. Represents the time evolution vector. Indicates the bias term;

[0113] It should be noted that the activation factor is defined based on the nonlinear mapping characteristics of the Sigmoid function, and the exemplary value range is [0,1]; the weight coefficient of the spatial imbalance vector is defined based on the variance contribution rate of the spatial imbalance vector, and the exemplary value range is [0.3,0.7]; the weight coefficient of the time evolution vector is defined based on the temporal correlation of the time evolution vector, and the exemplary value range is [0.3,0.7]; the bias term is used to adjust the coupling degree between the spatial imbalance vector and the time evolution vector, and the typical value is [-0.5,0.5].

[0114] S4. Based on the child's fall risk value, perform real-time posture tracking and environmental scanning of the child's activity status, and collect movement trajectory data; based on the movement trajectory data, perform emergency classification response and output intelligent protection plan.

[0115] S4.1 Match decision parameters to the risk value of children's falls and obtain activity monitoring strategies.

[0116] Based on a pre-defined activity safety strategy library, decision parameters are matched to the child's fall risk value. For example, a low range (0-0.3), a medium range (0.3-0.7), and a high range (0.7-1.0) are set for the child's fall risk value. When the child's fall risk value is in the low range, the "low risk" decision parameters (such as appropriate monitoring and activity restrictions) in the activity safety strategy library are matched. When the child's fall risk value is in the medium range, the "medium risk" decision parameters (such as appropriate activity guidance and early warning operations) in the activity safety strategy library are matched. When the child's fall risk value is in the high range, the "high risk" decision parameters (such as strict activity restrictions and monitoring) in the activity safety strategy library are matched. The decision parameters are then structured and encoded to obtain the activity monitoring strategy.

[0117] It should be noted that the activity safety strategy library is defined based on the risk frequency and scenario distribution of historical children's fall accidents; the range of children's fall risk values ​​is defined based on the probability distribution quantiles of historical children's fall risk values; structured coding refers to the process of mapping decision parameters to states and serializing logic through finite state automata.

[0118] S4.2 According to the activity monitoring strategy, the child's activity status is tracked in real time and the surrounding environment is scanned using depth cameras and lidar, and motion trajectory data is collected simultaneously using inertial measurement instruments.

[0119] The decision parameters of the activity monitoring strategy are extracted and imported into the depth camera and LiDAR via wireless transmission devices (such as Wi-Fi or Bluetooth). The depth camera tracks and captures the child's activity status (including skeletal joints, body orientation, and range of motion) in real time, and the LiDAR tracks the real-time distance and relative position of the child to obstacles in the surrounding environment to ensure blind-spot-free safety monitoring. At the same time, the inertial measurement instrument (IMU) collects and integrates the acceleration, angular velocity, and magnetometer data of the child's activities to obtain motion trajectory data.

[0120] S4.3 Based on preset three-level thresholds, perform emergency response classification on motion trajectory data and output intelligent protection solutions.

[0121] Noise is filtered out from real-time motion trajectory data using Kalman filtering to obtain motion trajectory parameters. These parameters are then used to determine the activity status and trigger a hazard classification response based on preset three-level thresholds. For example, if the motion trajectory parameters are within the first-level threshold, it is considered a low-risk routine activity, requiring an audio-visual alert and local log recording on the device. If the parameters are within the second-level threshold, it is considered a medium-risk pre-disorder sign, requiring an immediate alarm notification to the guardian's mobile app via wireless transmission. If the parameters are within the third-level threshold, it is considered a high-risk fall accident, requiring activation of smart protective devices deployed around the child (such as deployable safety airbags) and transmission of the accident scene's GPS location to the guardian. The hazard classification response data is then combined with the real-time motion trajectory data to output an intelligent protection plan.

[0122] It should be noted that the three thresholds are based on the statistical quantiles of historical motion trajectory parameters (e.g., the 95% confidence interval definition, with exemplary values ​​being: Level 1 threshold [0, 0.3], Level 2 threshold (0.3, 0.7], and Level 3 threshold (0.7, 1.0).

[0123] In summary, this invention achieves collaborative perception and fusion analysis of multi-dimensional motion features by extracting skeletal and trunk motion features through a dual-channel parallel processing architecture, and performing instantaneous velocity calculation and dense optical flow calculation, thereby reducing the false alarm rate caused by motion confusion. By employing a spatiotemporal behavior analysis model to quantify differences in motion direction and analyze continuous state changes, it can accurately capture the gradual imbalance evolution process of fall precursors, thus improving the timeliness of early warning and gaining a critical time window for protective intervention.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A motion capture-based intelligent monitoring method for child fall, characterized in that: The application relates to a child fall risk prediction method based on video analysis. Video of child activities is collected and posture estimation is performed to form a continuous activity frame sequence; The continuous activity frame sequence is processed in a double-channel parallel mode to obtain skeletal movement features and trunk movement features; Spectral analysis and instantaneous speed calculation are performed on the skeletal movement features to obtain lower limb skeletal point speeds, and dense optical flow calculation is simultaneously performed on the trunk movement features to generate trunk optical flow speeds; The trunk optical flow speeds and the lower limb skeletal point speeds are input into a space-time behavior analysis model, a spatial convolution layer performs direction difference quantification and limb coordination evaluation, a time convolution layer performs continuous state tracking and change trend analysis, and a child fall risk value is output; According to the child fall risk value, the activity state of the child is tracked in real time, the surrounding environment is scanned, and motion trajectory data is collected; critical classification response is performed based on the motion trajectory data, and an intelligent protection scheme is output.

2. The motion capture-based child fall smart monitoring method of claim 1, wherein: The continuous activity frame sequence is formed by extracting image frames of the child activity video and performing posture estimation and time sequence alignment.

3. The motion capture-based child fall smart monitoring method of claim 2, wherein: The skeletal movement features and the trunk movement features are obtained by the following steps, Motion vector analysis is performed on the continuous activity frame sequence to obtain skeletal joint displacement parameters and trunk region movement parameters; In the first channel, the skeletal joint displacement parameters are fitted by using a least square method to form a skeletal movement curve; curvature extreme points of the skeletal movement curve are extracted and dynamic time warping is performed to generate skeletal movement features; In the second channel, the trunk region movement parameters are subjected to energy accumulation integration to generate a trunk movement energy distribution, and the trunk movement energy distribution is subjected to wavelet transform and low-pass filtering to obtain trunk movement features.

4. The motion capture based child fall smart monitoring method of claim 1, wherein: The lower limb skeletal point speeds are obtained by the following steps, Fast Fourier transform is applied to the skeletal movement features to perform spectral analysis, generate frequency domain movement features, and perform inverse transform reconstruction to obtain time domain movement vectors; The lower limb skeletal point speeds are obtained by performing instantaneous speed calculation on the time domain movement vectors through a central difference formula.

5. The motion capture based child fall smart monitoring method of claim 4, wherein: The trunk optical flow speeds are generated by performing amplitude normalization and dense optical flow calculation on the trunk movement features through a Farneback algorithm.

6. The motion capture-based child fall smart monitoring method of claim 1, wherein: The specific construction process of the space-time behavior analysis model is as follows, A spatial convolution layer is built through a graph convolution network, and a time convolution layer is built through a time sequence convolution network; Attention mechanism is applied to cross-layer feature fusion and residual stacking of the spatial convolution layer and the time convolution layer to construct the space-time behavior analysis model.

7. The motion capture-based child fall smart monitoring method of claim 6, wherein: The spatial convolution layer performs direction difference quantification and limb coordination evaluation, and the specific steps are as follows, The trunk optical flow speeds and the lower limb skeletal point speeds are subjected to dimension compression and input into the space-time behavior analysis model; The spatial convolution layer performs direction difference quantification on the trunk optical flow speeds and the lower limb skeletal point speeds through a graph convolution network to form a movement direction difference matrix, and performs limb coordination evaluation on the movement direction difference matrix to generate a spatial imbalance vector.

8. The motion capture-based child fall smart monitoring method of claim 1, wherein: The time convolution layer performs continuous state tracking and change trend analysis, and the specific steps are as follows, The time convolution layer performs continuous state tracking on the trunk optical flow velocity and lower limb skeleton point velocity through an expanded causal convolution to generate a space-time feature sequence; The space-time feature sequence is subjected to time series dependence modeling and change trend analysis to form a time evolution vector.

9. The motion capture-based child fall smart monitoring method of claim 8, wherein: The output child fall risk value refers to feature weighted fusion of the space imbalance vector and the time evolution vector, and application of a Sigmoid function for probability normalization to output a child fall risk value.

10. The motion capture-based child fall smart monitoring method of claim 1, wherein: The output intelligent protection scheme specifically includes the following steps, The child fall risk value is subjected to decision parameter matching to obtain an activity monitoring strategy; According to the activity monitoring strategy, real-time posture tracking and surrounding environment scanning of the activity state of the child are performed through a deep camera and a laser radar, and motion trajectory data is collected synchronously using an inertial measurement instrument; Based on a preset three-level threshold, the motion trajectory data is subjected to critical classification response to output an intelligent protection scheme.

Images