Human body gravity center track estimation and balance capability detection method and device

Abstract

The invention relates to the field of human body motion detection and balance ability evaluation, and discloses a human body gravity center track estimation and balance ability detection method and device.The method comprises the steps that multi-dimensional motion data of nine-axis inertial measurement units worn on the head and the waist and multi-dimensional individual feature data of a human body are synchronously collected; embedding individual features through a self-adaptive attention mechanism, and performing deep fusion with the motion data; and inputting the fusion data into a special time convolution network model containing a differential expansion convolution structure and an adaptive activation function to realize end-to-end direct mapping from original data to three-dimensional gravity center trajectory parameters. According to the method, physical model hypothesis is abandoned, model errors are fundamentally eliminated, the calculation efficiency and the estimation precision in a dynamic scene are greatly improved, individual differences can be adaptively learned, the generalization ability is high, and the high-precision real-time balance ability detection requirement can be met.

Description

Technical Field

[0001] The present invention relates to the field of human motion detection and balance ability assessment, and in particular to a method and device for estimating the trajectory of the human center of gravity and detecting the balance ability. Background Art

[0002] The current general technology is plantar pressure plate (matrix) measurement, which is based on the physical model solution of the human body center of gravity based on single gyroscope data.

[0003] (1) Deficiencies in existing technologies Existing solutions for estimating the trajectory of the human center of gravity and assessing balance are based on traditional plantar pressure plate (matrix) measurements. This solution only measures the projection of the center of gravity on a two-dimensional plane (XY axis) and cannot capture changes in the center of gravity height in the vertical direction (Z axis). More importantly, plantar pressure signals cannot distinguish between the movement characteristics of the legs and torso. When the torso leans forward or the head turns, the pressure plate only captures changes in plantar pressure distribution, but cannot interpret the direct impact of changes in torso posture on the center of gravity.

[0004] In balance assessment scenarios, trunk motion characteristics (such as trunk inclination and angular velocity) significantly outweigh the leg's influence on center of gravity trajectory. Sports biomechanics analysis shows that changes in trunk posture contribute 73% to center of gravity control in dynamic posture control. However, traditional plantar pressure plates only collect plantar pressure data and cannot directly capture trunk motion information. Consequently, in center of gravity trajectory analysis scenarios involving the trunk (such as daily walking posture analysis and sports movement assessment), the accuracy of attributing center of gravity changes is less than 50%.

[0005] Although the solution that relies only on waist sensors can obtain some torso data, it ignores the impact of head movement on the overall center of gravity.

[0006] Head movement is a high-frequency dynamic factor in human balance regulation. When the head's angular velocity exceeds 30° / s (40-60° / s is common in everyday rapid head turns), its instantaneous impact on the center of gravity trajectory accounts for 22%-25%. Furthermore, the mechanical conduction pathway of head movement (skull → cervical spine → trunk) is asynchronous with the trunk-to-lumbar spine data collected by the waist sensor. A single waist sensor cannot capture the complete temporal characteristics of this conduction chain, from head trigger to trunk response (neural conduction delay is approximately 15-30ms, requiring synchronous acquisition of the head and waist sensors within 10ms or less to capture dynamic changes in the center of gravity trajectory). Using only the waist sensor would miss the correlation between the head trigger signal and the trunk response signal, rendering the model unable to distinguish between the effects of active head rotation and compensatory trunk sway on the center of gravity, leading to a significant increase in estimation errors.

[0007] 2. Insufficiency of existing models Physical models based on dynamic equations assume that the human body is rigid, requiring manual input of parameters such as bone length and joint angles that are difficult to measure accurately. Furthermore, they are unable to adaptively learn individual differences. In dynamic scenarios (such as running and standing on one leg), flexible joint motion is ignored: Using a traditional rigid body physical model (such as the human body dynamics model based on the Newton-Euler equation), the bone parameters converted from the subject's height and weight were manually input (with an error of approximately ±5%), and the center of gravity trajectory data was output; a Vicon optical motion capture system (10 infrared cameras, accuracy of 0.1mm) + a force platform (three-dimensional force accuracy of 0.1N) was used to obtain the actual center of gravity trajectory (gold standard data) for comparison. The root mean square error (RMSE) between the physical model output and the precise control data was calculated for 30 people, each performing five dynamic motions (a total of 150 data sets). Statistics showed that the average model error in running scenarios reached 16.2%, while the peak error in single-leg standing (eyes closed) scenarios reached 21.7%. Across all dynamic scenarios, 68% (102 of the 150 data sets) experienced model errors exceeding 15%.

[0008] Therefore, there is an urgent need for a balance ability detection method and device that can realize three-dimensional dynamic human center of gravity trajectory estimation, is adaptable to different groups of people and scenes, and has sufficient generalization capabilities to meet the requirements for improving the accuracy of human center of gravity trajectory estimation and balance ability detection. Summary of the Invention

[0009] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a method and device for estimating the trajectory of the center of gravity of a human body and detecting the balance ability.

[0010] A first implementation scheme is given below.

[0011] A method for estimating a human body's center of gravity trajectory and detecting balance ability, comprising: S1, data acquisition: synchronously acquiring multi-dimensional motion data output by a first inertial measurement unit worn on a human head and a second inertial measurement unit worn on a human waist; and obtaining multi-dimensional individual feature data of the human body; S2, feature fusion: inputting the multi-dimensional individual feature data into the embedding layer to generate an individual feature vector; and concatenating and fusing the individual feature vector with the multi-dimensional motion data to generate fused feature data; S3, model reasoning: inputting the fused feature data into an end-to-end neural network model, performing computational processing, and outputting the three-dimensional center of gravity trajectory parameters of the human body; The neural network model includes at least one dilated convolution layer, which includes a multi-layer differential dilated convolution structure using non-equidistant dilation factors to expand the temporal receptive field and capture long-distance dependencies.

[0012] Furthermore, the specific steps of S2 include: Inputting the multi-dimensional individual feature data into an adaptive attention fully connected embedding layer, dynamically adjusting the weights according to different individual features, and generating the individual feature vector; splicing the individual feature vector and the multi-dimensional motion data in the feature dimension; The concatenated data is subjected to dimensionality reduction through 1x1 convolution to generate the fused feature data.

[0013] Furthermore, the end-to-end neural network model is a temporal convolutional network model.

[0014] Furthermore, the activation function in the temporal convolutional network model is an adaptive activation function, which dynamically switches among a variety of preset activation functions according to the feature complexity of the current time step during the model training process.

[0015] Furthermore, in S1 , acquisition time synchronization of the first inertial measurement unit and the second inertial measurement unit is achieved by comparing hardware trigger pulses and high-precision hardware timestamps.

[0016] Furthermore, the method further comprises: S4, balance ability assessment: Based on the three-dimensional center of gravity trajectory parameters output by S3, calculate at least one balance ability assessment index and evaluate the balance ability of the human body according to the index, the index including the center of gravity offset amplitude and speed fluctuation rate.

[0017] Based on the first embodiment, a second embodiment is given.

[0018] A device for estimating the trajectory of a human body's center of gravity and detecting balance ability, comprising: The sensor module includes a first inertial measurement unit arranged on the human head and a second inertial measurement unit arranged on the human waist; a data acquisition module, connected to the sensor module, for synchronously acquiring multidimensional motion data output by the first inertial measurement unit and the second inertial measurement unit, and for acquiring multidimensional individual feature data of the human body; A processing module, connected to the data acquisition module, configured to: Processing the multi-dimensional individual feature data through an embedding layer to generate an individual feature vector; splicing and fusing the individual feature vector with the multi-dimensional motion data to generate fused feature data; The fused feature data is input into an end-to-end neural network model, and the three-dimensional center of gravity trajectory parameters of the human body are calculated and output.

[0019] Furthermore, in the processing module, the embedding layer used to generate individual feature vectors is an adaptive attention fully connected embedding layer, which is used to dynamically adjust weights according to different individual features, and reduce the dimension of the spliced ​​data through 1x1 convolution to generate the fused feature data. The end-to-end neural network model solidified in the processing module is a temporal convolutional network model, which includes a multi-layer differential dilated convolution structure with non-equidistant dilation factors, and an adaptive activation function that can be dynamically selected during training.

[0020] Furthermore, the device further comprises: An evaluation and display module is connected to the processing module and is used to calculate the center of gravity offset amplitude and / or speed fluctuation rate based on the output three-dimensional center of gravity trajectory parameters, evaluate the balance ability of the human body, and visualize the center of gravity trajectory curve, balance ability level or spatial distribution characteristic diagram of the center of gravity.

[0021] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: The present invention abandons the physical model assumptions, fundamentally eliminates model errors, greatly improves computing efficiency and estimation accuracy in dynamic scenarios, and can adaptively learn individual differences, has strong generalization capabilities, and can meet the needs of high-precision real-time balance ability detection.

[0022] Through hardware-level time synchronization (error <0.05ms), the timing consistency of head and waist data is ensured, avoiding center of gravity trajectory estimation errors caused by time differences (error level <0.1mm).

[0023] The dual IMUs on the head and waist capture high-frequency dynamic features of the "vestibular-neck-torso" (such as rapid head rotation angular velocity >40° / s) and low-frequency features of the "torso-lower limbs", making up for the shortcomings of a single sensor (such as only the waist IMU) and reducing the center of gravity estimation error (for example, the error in static standing scenarios is reduced from 3.2cm to 1.5cm).

[0024] The introduction of individual characteristic data provides a basis for the personalized modeling of subsequent models and improves the adaptability of the model to different populations (such as the elderly and athletes).

[0025] Normalization and standardization improve the efficiency of neural network training and reduce the gradient instability problem caused by data dimension differences.

[0026] Through an end-to-end TCN architecture, the physical model assumptions are abandoned, and the center of gravity trajectory mapping is learned directly from multimodal data, reducing the error by 35% (RMSE from 5.2cm to 3.4cm). In particular, the error is controlled within 7% in joint flexible movement scenarios (such as squats).

[0027] The attention mechanism embeds individual characteristics, enhancing the model's adaptability to different populations (such as the middle-aged and elderly), reducing error by 28%. End-to-end learning eliminates errors in physical model assumptions, significantly reducing dynamic scene error. The error for dynamic scenes using only an end-to-end neural network model is 5% to 8%, while the error for scenes using a physical model is 15% to 20%.

[0028] Visualized results (such as heat maps and curve charts) help users and professionals quickly identify balance problems (such as excessive fluctuation in a certain axis), assist in clinical diagnosis or sports training, and combine assessment results based on individual characteristics to adapt to different populations (such as a 39% reduction in errors in the middle-aged and elderly group), thereby improving the targeted nature of the assessment.

[0029] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 It is a flow chart of the method for estimating the trajectory of the center of gravity of the human body and detecting the balance ability of the present invention; Figure 2 It is a structural diagram of the human body center of gravity trajectory estimation and balance ability detection device of the present invention. DETAILED DESCRIPTION

[0032] To make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0033] In order to solve the problems existing in the prior art, the inventors have developed a variety of solutions to optimize and replace the existing technical solutions.

[0034] To address the shortcomings of the aforementioned existing technologies in terms of three-dimensional center of gravity trajectory estimation accuracy, dynamic feature capture, and individual adaptability, the inventors proposed an innovative architecture (Solution 1) that integrates physical models with deep learning. This solution utilizes multimodal inertial sensors (a nine-axis IMU mounted on the head and waist) to simultaneously collect multidimensional motion data. Combined with layered physical modeling and adaptive neural network optimization, this approach not only compensates for the simplification errors of traditional physical models regarding joint flexibility, but also enhances the ability to learn individual differences and high-frequency dynamic features through a data-driven approach. Ultimately, this approach enables highly accurate center of gravity trajectory estimation and quantitative measurement of balance ability.

[0035] The specific implementation of Plan 1 is as follows: The system collects multi-dimensional motion data (18 dimensions) from nine-axis sensors on the head and waist, as well as individual bone parameter measurements or estimates. Multi-dimensional motion data is acquired using the sensors' built-in high-precision A / D converters, achieving 16-bit accuracy to ensure data accuracy. Individual bone parameter measurements are obtained using medical imaging equipment (such as X-rays and CT scans) to obtain data such as key bone lengths and joint angles. If measurements are unavailable, estimates are made based on statistical models and individual physical data (height, weight, etc.), or manually specified by experienced professionals.

[0036] The center of gravity trajectory estimation error between a single IMU (waist only) and two IMUs (head + waist) was compared (the root mean square error (RMSE) was calculated using the Vicon optical system's accuracy of 0.1mm). Thirty subjects (aged 20-50, covering the general population and sports enthusiasts) were selected for this comparison. These subjects were tested in three typical scenarios: static standing (eyes closed), dynamic walking (normal pace), and rapid head turning with torso shaking. The results are shown in the table below.

[0037] Table 1: Comparison of center of gravity trajectory estimation error between a single IMU (waist only) and two IMUs (head + waist) Analyzing the reasons, the waist IMU only covers the "trunk-lower limb" motion chain, while the head IMU can capture high-frequency posture changes transmitted by the "vestibular-neck-trunk" (for example, when the angular velocity of the head turns rapidly is greater than 40° / s, the head IMU can sense the signal 15-30ms in advance), thus compensating for the response delay of the waist sensor to the "head-triggered center of gravity disturbance"; at the same time, the fusion of the two sets of IMU data can distinguish the contribution of "active head movement (such as turning the head)" and "trunk compensatory movement (such as bending over)" to the center of gravity - in the "head rotation + trunk leaning forward" compound action, the single-group IMU solution misjudged the "head movement contribution ratio" by 35%, while the two-group IMU solution can control the decoupling error within 8%.

[0038] Different from the traditional "single-site IMU" solution (such as only the waist / wrist), the dual-IMU collaborative mechanism of this solution has unique advantages: it does not rely on "rigid body assumption" correction (directly covers the "head-chest-waist" key motion chain, reducing model assumption errors); it is adaptable to the universality of dynamic scenes (in high-frequency scenes such as running and jumping, the error of the single-IMU solution increases linearly with the complexity of the movement (R²=0.87), while the error growth trend of the dual-IMU solution is gentle (R²=0.32)), verifying the robustness of multi-site sensing in dynamic scenes.

[0039] R² (coefficient of determination) is a measure of the goodness of fit of a regression model. Its mathematical meaning is the proportion of the variation in the dependent variable that is explained by the independent variable. Specifically, it is equal to the ratio of the regression sum of squares to the total sum of squares, or the square of the correlation coefficient.

[0040] The core improvements of Solution 1 include physical model construction and neural network calculation.

[0041] The physical model construction includes establishing a kinematic mapping relationship through the Newton-Euler equation based on a simplified head-thorax-lower skeletal model (the head and thorax, and the thorax and lower thorax are considered to be rotatable connections). A hierarchical modeling approach is used to first model the head, thorax, and lower thorax separately, and then integrate them through the connection parts. Combined with the input individual bone parameters, the preliminary center of gravity estimate (three-dimensional coordinate G x ,G y ,G z The human body is simplified into three parts: head, chest, and subthorax. The connections between these parts follow the laws of rotation. To adapt to actual conditions, the rotation range of the head relative to the chest is limited to ±60°, and the rotation range of the chest relative to the subthorax is set to ±30°. The model's range of motion and mechanical properties are also adjusted based on individual skeletal parameters.

[0042] Neural network calculations include center of gravity estimation based on the output of the physical model + multi-dimensional motion data (18 dimensions) + multi-dimensional individual characteristic data (8 dimensions, including age, height, gender, sitting height, waist circumference, chest circumference, hip circumference, and length from the sole of the foot to the knee joint) + individual bone parameters; The unit of data involving length or height is cm.

[0043] Before data input, the original sensor time series data is normalized and the minimum-maximum normalization method is used to map the data to the [0,1] interval to improve the training efficiency of the neural network.

[0044] The neural network uses a simple fully connected architecture consisting of two layers. The first layer contains 32 neurons, and the second layer contains 6 neurons. It directly processes the input data and outputs 6-dimensional center of gravity trajectory parameters (3D coordinates + 3D velocity). The model is optimized using the MSE loss function. To prevent overfitting, an L2 regularization term is added to the loss function, with a regularization coefficient of 0.001.

[0045] The implementation process of Option 1 is as follows Data collection: We simultaneously collect 18-dimensional multi-dimensional motion data (100Hz) from the head and waist, individual skeletal parameter measurements (if available), and multi-dimensional individual feature data. We use an optical capture system (with an accuracy of <2cm) to obtain the true center of gravity label. Individual skeletal parameter estimates are calculated using a pre-trained regression model combined with multi-dimensional individual feature data. Multi-threading technology is used during data collection to ensure simultaneous data collection and prevent data loss.

[0046] Physical model calculation: The sensor data is used to calculate the relative joint angles between the head and chest, and between the chest and subthorax. Combined with individual skeletal parameter measurements or estimates, the initial center of gravity position is calculated using dynamic equations. Quaternions are used to calculate joint angles, avoiding gimbal lock and improving accuracy.

[0047] Neural Network Optimization: The physical model output, multidimensional motion data, and multidimensional individual feature data are integrated with individual skeletal parameters and fed into a fully connected neural network to learn dynamic correction parameters to compensate for errors in the simplified assumptions of the physical model. During model training, a cross-validation approach is used, dividing the dataset into training, validation, and test sets in a ratio of 7:2:1.

[0048] It has been verified that the accuracy of Scheme 1 has been greatly improved compared with the traditional test scheme, but the following problems still exist.

[0049] 1. Parameter sensitivity: Physical models rely heavily on the accuracy of multidimensional motion data and individual skeletal parameters. Both measurement errors and estimation biases can lead to increased errors in center of gravity estimation. Furthermore, simple fully connected neural networks struggle to fully extract complex temporal features, impacting model accuracy.

[0050] 2. Insufficient dynamic adaptability: The physical model is based on simplified structural assumptions and is difficult to adapt to joint flexible movements in real time. It has low correction efficiency in fast-action scenarios, and the neural network structure is simple and cannot effectively process dynamically changing data.

[0051] 3. High computational complexity: Solving the dynamic equations requires matrix iteration operations, coupled with the calculation or processing of individual bone parameters. The time consumption for single-frame data processing is further increased, far exceeding 80ms, making it even more difficult to meet real-time detection requirements (required to be <50ms).

[0052] After systematic research and analysis, a pure neural network adaptive solution (Scheme 2) was proposed. This solution, centered on end-to-end deep learning, abandons the structural assumptions of traditional physical models. By deeply integrating multimodal sensor data with individual features and combining it with the enhanced capture of temporal dynamics by a dilated convolutional network, it achieves a direct mapping from raw data to center-of-gravity trajectory. Compared to Scheme 1, this solution significantly improves model generalization, computational efficiency, and accuracy in complex motion scenarios, providing a more robust, data-driven solution for balance testing.

[0053] Figure 1 The implementation process of the detection method of scheme 2 is shown.

[0054] A method for estimating the trajectory of the human body's center of gravity and detecting balance ability, S1, data acquisition: synchronously acquiring multi-dimensional motion data output by a first inertial measurement unit worn on a human head and a second inertial measurement unit worn on a human waist; and obtaining multi-dimensional individual feature data of the human body; The acquisition time synchronization of the first inertial measurement unit and the second inertial measurement unit is achieved by comparing the hardware trigger pulse and the high-precision hardware timestamp.

[0055] S2, feature fusion: inputting the multi-dimensional individual feature data into the embedding layer to generate an individual feature vector; and concatenating and fusing the individual feature vector with the multi-dimensional motion data to generate fused feature data; Inputting the multi-dimensional individual feature data into an adaptive attention fully connected embedding layer, dynamically adjusting the weights according to different individual features, and generating the individual feature vector; splicing the individual feature vector and the multi-dimensional motion data in the feature dimension; The concatenated data is subjected to dimensionality reduction through 1x1 convolution to generate the fused feature data.

[0056] S3, model reasoning: inputting the fused feature data into an end-to-end neural network model, performing computational processing, and outputting the three-dimensional center of gravity trajectory parameters of the human body; The end-to-end neural network model is a temporal convolutional network model.

[0057] The temporal convolutional network model includes at least one dilated convolutional layer, which includes a multi-layer differential dilated convolution structure using non-equidistant dilation factors to expand the temporal receptive field and capture long-distance dependencies.

[0058] The activation function in the temporal convolutional network model is an adaptive activation function. During the model training process, it dynamically switches among multiple preset activation functions according to the feature complexity of the current time step.

[0059] S4, balance ability assessment: Based on the three-dimensional center of gravity trajectory parameters output by S3, calculate at least one balance ability assessment index and evaluate the balance ability of the human body according to the index, the index including the center of gravity offset amplitude and speed fluctuation rate.

[0060] The core improvements of Solution 2 include abandoning the dependence on physical models, adopting individual feature embedding mechanism and dynamic feature enhancement.

[0061] Abandoning the dependence on physical models, the center of gravity trajectory mapping relationship is directly learned from "multi-dimensional motion data + individual feature data" through end-to-end deep learning, avoiding the systematic errors caused by simplified structural assumptions.

[0062] A specific embodiment is given according to the method for estimating the trajectory of the center of gravity of the human body and detecting the balance ability.

[0063] We selected 100 sets of dynamic motion data (running, jumping, agile direction changes, etc.) and compared the error performance of solution 1 (physical model + neural network) and solution 2 (pure end-to-end learning).

[0064] The error of Scheme 2 is reduced by 35% compared with Scheme 1 (RMSE is reduced from 5.2cm to 3.4cm). In addition, in joint flexible movement scenarios (such as squats and standing up), the error rate of Scheme 1 is as high as 18% due to the rigid body assumption, while that of Scheme 2 can be controlled within 7%.

[0065] By utilizing pre-trained model parameters from large-scale public datasets (such as Human3.6M and AMASS), and targeting the "head + waist multimodality + individual feature" scenario of this solution, the "universal human motion features" and "individual difference features" of the pre-trained model are decoupled through a feature adapter (a unique design that is different from general transfer learning). This increases the model convergence speed by 40% (reducing the number of iterations from 200 to 120), solving the training efficiency issue in small sample scenarios.

[0066] Individual feature embedding mechanism: 8-dimensional individual feature data (8 dimensions, including age, height, gender, sitting height, waist circumference, chest circumference, hip circumference, and length from the sole to the knee joint, with the larger unit of length and height in cm) is introduced. A 16-dimensional feature vector is generated through an adaptive attention fully connected embedding layer. After being spliced ​​with the data obtained by the 18-dimensional sensor, it is compressed to 18 dimensions through a 1×1 convolution to guide the network to learn individual-specific movement patterns.

[0067] A specific design has also been added. Unlike the "universal embedding + fixed weight" approach, this mechanism dynamically adjusts attention weights based on individual characteristics. For example, the correlation weight between height and center of gravity height is 32% higher for older adults (age > 60) than for younger individuals, and the correlation weight between weight and horizontal center of gravity offset is 27% higher for men than for women. By embedding an individual feature attention matrix (dimension 8×16) within the fully connected layer, a personalized "feature-weight" binding is achieved. In scenarios with significant individual differences (such as the comparison between the elderly and athletes), the center of gravity trajectory estimation error is reduced by 28%.

[0068] We then conducted ablation experiments to verify that after removing the "individual feature embedding" or "attention mechanism", the model's average error for subjects of different heights and weights increased by 25% and 19% respectively, proving that this mechanism plays an important role in learning individual-specific movement patterns.

[0069] Dynamic feature enhancement design: A three-layer differentially expanded convolutional structure (expansion factors of 1 / 3 / 5, different from the general isometric expansion of 1 / 2 / 4) is adopted to expand the receptive field to 45 time steps layer by layer. Combined with the adaptive neuron activation function (automatically switching between ReLU, LeakyReLU, and ELU during training, dynamically selected based on the feature complexity of the current time step) and the BN layer, it improves the ability to capture high-frequency dynamic features such as rapid head rotation (angular velocity > 50° / s) and forward torso leaning.

[0070] It should be noted here that isometric dilation (e.g., factors of 1 / 2 / 4) in general dilated convolutions can easily lead to "feature hole overlap". In scenes with rapid head rotation (angular velocity 60° / s), the feature loss rate can reach 22%. This solution uses a non-equidistant expansion factor design to make the receptive field coverage more uniform, and the feature loss rate is controlled within 8% in the same scenario.

[0071] Traditional fixed activation functions (such as single ReLU or Leaky ReLU) have significant limitations during model training. When processing complex dynamic features such as rapid head rotation and forward torso leaning, the model is prone to falling into local extremes due to vanishing gradients, causing training convergence to stagnate or significantly slow. For example, in training scenarios involving high-frequency motion time series, fixed activation functions can significantly increase the number of model iterations in local extremes, extending the overall training cycle.

[0072] The adaptive activation function mechanism proposed in this scheme optimizes the convergence characteristics through a dynamic learning strategy: Dynamic Selection Mechanism: During training, the model automatically switches between activation functions such as ReLU, Leaky ReLU, and ELU based on the feature complexity of the input data at the current time step (such as signal fluctuation amplitude and time series correlation). For example, when processing high-frequency acceleration signals generated by head rotation, the model adaptively selects Leaky ReLU to preserve the gradient information of negative features, avoiding the vanishing gradient caused by the "neuron death" problem of ReLU.

[0073] Local extrema avoidance: By evaluating gradient trends in real time, the system automatically switches activation functions and adjusts neuron activation patterns when it detects a slowdown in model convergence (e.g., when the loss decreases by less than a preset threshold over multiple consecutive iterations). This mechanism effectively reduces the time the model spends in local extrema, improving training stability.

[0074] Experimental verification shows that compared with a fixed activation function, this adaptive mechanism can significantly reduce the convergence stagnation caused by local extreme values ​​during the training of the model in complex dynamic scenes, optimize both the number of training rounds and the convergence time, and meet the requirements of dynamic balance detection tasks for the real-time and robustness of the model.

[0075] A specific implementation method of the second solution is described as follows.

[0076] 1. Multimodal data fusion: The nine-axis IMU on the head (forehead) and the nine-axis IMU on the waist (lumbar spine) synchronously collect 18-dimensional motion data, and achieve high-precision time synchronization through a hardware-level synchronous sampling control architecture. The specific solution is as follows: Hardware collaboration: Select an IMU sensor with a built-in synchronous sampling control module. The main control SOC outputs a hardware trigger pulse (matching the sensor sampling rate) through the GPIO pin to ensure that after the IMU receives the trigger signal, it starts synchronous sampling and ensures that the sampling time of the dual IMUs is consistent.

[0077] Timestamp and error verification: Each IMU embeds a 64-bit hardware timestamp when outputting data (generated by a high-precision timer inside the sensor with a timing accuracy better than 1μs). The data processing end calculates the synchronization error by comparing the timestamps of the two IMUs: Verified by 100,000 sets of continuous sampling data, The maximum value is 0.8μs, and the time fluctuation of 99.9% of the data is less than 0.05ms, which meets the synchronization requirements in high-frequency motion scenarios. The hardware timestamp embedded when the head IMU collects data. The hardware timestamp embedded when the waist IMU collects data.

[0078] Quantification of error impact: The center of gravity trajectory error caused by time synchronization fluctuation can be derived through the kinematic model: Assuming that the center of gravity of the human body is about 1.5m high, the head and waist IMUs have a time difference. The spatial displacement error caused by for: (a is the acceleration of human motion, taking the typical value of 5m / s for running scenes 2 ) ,have to , which is much smaller than the sensor measurement noise (about 0.1 mm), verifying that the error of this synchronization scheme is negligible and ensuring the accuracy of data fusion.

[0079] 2. TCN Model Architecture Input layer: T×34-dimensional data (T=100 time steps, 34 dimensions = 18-dimensional sensor data + 16-dimensional individual feature vector).

[0080] The input data is a time series-feature two-dimensional matrix, defined as: in: T = 100, time step, each step corresponds to a 10ms sampling interval, covering continuous motion data within 1 second; 34-dimensional features = 18-dimensional motion data + 16-dimensional individual feature vector; 18-dimensional motion data: 9-axis IMU on the head (3D acceleration + 3D angular velocity + 3D magnetic field) + 9-axis IMU on the waist (3D acceleration + 3D angular velocity + 3D magnetic field), ,Right now ; : head acceleration [x, y, z] 3D row vector; : head angular velocity [x, y, z] 3D row vector; : head magnetic field [x, y, z] 3D row vector; : waist acceleration [x, y, z] 3D row vector; : waist angular velocity [x, y, z] 3D row vector; : [x, y, z] 3D row vector of the waist magnetic field.

[0081] 16-dimensional individual feature vector: generated by converting 8-dimensional original individual features (height, weight, sitting height, age, etc.) through the adaptive attention fully connected embedding layer, that is, (The same individual feature vector is reused at each time step to ensure temporal consistency).

[0082] The input data is processed in blocks, and each block contains data of a certain time step, which facilitates parallel calculation of the model and improves calculation efficiency.

[0083] Feature fusion layer: A 1×1 convolution compresses 34 dimensions to 18 dimensions, enabling cross-modal feature interaction. During feature fusion, a channel attention mechanism is employed to assign different weights to features of different channels based on their importance, enhancing the model's ability to extract key features.

[0084] The specific instructions are as follows: The feature fusion layer achieves cross-modal feature compression and fusion through 1×1 convolution operation. The specific process is as follows: Input matrix: . Convolution kernel parameters: Use 1×1 convolution kernel (size 1×1, input channels 34, output channels 18).

[0085] Dimensionality reduction calculation: in Represents the convolution operation, and the output matrix is .

[0086] Dilated convolutional layer: 3 layers (dilation factors 1 / 3 / 5) with a convolution kernel size of 5 capture long-range dependencies in time series. Residual connections are used in the dilated convolution layers to avoid vanishing gradients and improve model training stability.

[0087] The dilated convolution layer contains three layers of differentiated dilated convolution structures. The matrix form of input and output of each layer and the logic of receptive field expansion are as follows: Table 2. 3-layer differential dilated convolution structure of dilated convolution layer K1, K2, and K3 are the convolution kernel parameters corresponding to the three convolution layers. Indicates that its characteristics are dimensional matrix; X1, X2, and X3 are the output matrices corresponding to the three convolutional layers. Indicates that its characteristics are dimensional matrix.

[0088] Note: The convolution kernel size is fixed to 5×1 (5 steps in the time dimension and 1 step in the feature dimension) to ensure the continuity of the temporal features; the non-equidistant expansion factor (1→3→5) makes the receptive field expand nonlinearly to avoid the "feature holes" caused by the equidistant factor (for example, the traditional factor 1→2→4 will cause 12% feature loss at T=100, while this solution reduces it to less than 5%); the output matrix maintains dimensionality, ensuring temporal integrity (the features of each time step correspond one-to-one to the original samples), and preserving temporal association for the subsequent dynamic estimation of the center of gravity trajectory.

[0089] Output layer: The 6-dimensional center of gravity trajectory parameters and the loss function are MSE + individual feature regularization term (λ = 0.01), forcing the model to learn individual difference characteristics. During training, the regularization coefficient is dynamically adjusted based on the model's training status to balance the model's fitting ability and generalization ability.

[0090] After extracting the time series features through 3 layers of dilated convolution, The convolution + fully connected layer outputs the 6-dimensional center of gravity trajectory parameters for each time step in vector form: Among them, the 6-dimensional parameters are defined as , corresponding to the user Cartesian coordinate system (unit: m) and three-dimensional velocity (unit: m / s) based on the world coordinate point and orientation at the initial moment, respectively.

[0091] is the coordinate of the center of gravity; is the velocity vector relative to the previous point; Y out It is the full link output vector, with T dimensions for rows and 6 dimensions for columns.

[0092] The core basis of dynamic adjustment is to define two key indicators as adjustment trigger conditions: Loss difference ratio: in, is the MSE loss of the t-th round training set, is the MSE loss of the validation set in round t. Reflects the risk of model overfitting: The larger it is, the higher the risk of overfitting (validation loss is much larger than training loss).

[0093] Loss change rate: Reflecting the trend of validation loss: When , the validation loss increases and the model generalization ability decreases.

[0094] The update formula of the regularization coefficient λ is: Adjustment conditions: Condition 1: Increase λ (strengthen regularization): When (loss difference ratio ≥ 30%) and When the validation loss increases, λ is increased by 20%, forcing the model to reduce overfitting to individual features (for example, avoiding overfitting extreme bone parameters of a certain group of people).

[0095] Condition 2: Reduce λ (weaken regularization): When (loss difference ratio ≤ 10%) and When the validation loss decreases or stabilizes, λ is reduced by 20%, allowing the model to more fully learn individual differences (such as the differences in bone parameters between the elderly and athletes).

[0096] In other cases where condition 1 or condition 2 is not met, λ remains unchanged.

[0097] The following is an example of dynamic adjustment.

[0098] Initial For example, simulate the adjustment process of 5 rounds of training: Table 3 Adjustment process of simulation training Balance ability assessment indicators Center of gravity offset: D: Usually represents a certain distance (such as the distance between two points in space, etc., determined in combination with the specific application scenario, and the formula form is similar to the spatial distance formula).

[0099] 、 、 : Represents the coordinate value of a point in the x, y, and z dimensions (such as the three coordinate axis directions of the spatial rectangular coordinate system) or related physical quantities, characteristic quantities, etc.

[0100] 、 、 :Usually 、 、 The mean of the corresponding dimension is the mean of the center of gravity in static standing. By calculating the center of gravity deviation amplitude, the stability of the human body during movement can be evaluated.

[0101] Speed ​​Fluctuation: Reflects the stability of the center of gravity movement. The smaller the speed fluctuation rate, the more stable the center of gravity movement and the better the body's balance ability.

[0102] in: : Usually represents the standard deviation associated with the variable V, used to measure The degree of dispersion of the data.

[0103] N: represents the total number of items to be summed, which is a positive integer representing the amount of data.

[0104] : is the nth observation or data point, where n ranges from 1 to N.

[0105] :yes The mean is used to measure the central tendency of the data.

[0106] The following comparison between Scheme 1 and Scheme 2 further illustrates the advantages of Scheme 2 in test accuracy and generalization ability.

[0107] Table 4 Comparison of scheme advantages 5 Through actual analysis and comparison, Solution 1 has achieved a significant improvement in accuracy compared to the existing technology of solving the human center of gravity physical model based on single gyroscope data, which uses plantar pressure plate (matrix) measurement. Solution 2, through the optimization of the convolutional network, not only maintains the accuracy but also further improves it, while also improving the computing speed and generalization ability.

[0108] Specifically, the model assumption error is fundamentally eliminated, and the model convergence speed is increased by 40%; the computing efficiency prompt function is more than 3 times, and the dynamic error is reduced by more than 50%. At the same time, the network structure is designed for time series dynamic features and the dynamic selection of activation functions has more generalization capabilities.

[0109] Figure 2 The structure of the human body center of gravity trajectory estimation and balance ability detection device that implements the detection method of scheme 2 is shown.

[0110] The following provides a human body center of gravity trajectory estimation and balance ability detection device that implements the human body center of gravity trajectory estimation and balance ability detection method.

[0111] The configuration of the device for estimating the trajectory of the human center of gravity and detecting the balance ability is described below, and the part of the device's corresponding component modules that execute the method for estimating the trajectory of the human center of gravity and detecting the balance ability is not repeated here.

[0112] (1) Sensor module: Nine-axis head IMU (M1): Integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. It is fixed on the forehead and its coordinate system is consistent with the sagittal plane of the human body.

[0113] Waist nine-axis IMU (M2): The same specification model, installed in the middle of the lumbar spine, the coordinate system is aligned with the torso axis.

[0114] (2) Data acquisition module: Synchronously collect 9-dimensional data (acceleration, angular velocity, and magnetic field strength) from the head and waist to form an 18-dimensional raw input vector. Simultaneously retrieve pre-entered individual features.

[0115] The data acquisition module includes a microcontroller unit: a high-performance microcontroller with an integrated floating-point unit (FPU) that supports 100Hz high-speed data processing.

[0116] Data processing flow: Synchronous calibration: Dual sensor time synchronization (error < 1 μs) is achieved through a hardware timer, and spatial alignment is achieved based on the anatomical coordinate system.

[0117] Preprocessing: Data filtering removes high-frequency noise such as muscle vibration; extended Kalman filtering fuses multi-sensor data and outputs calibrated acceleration, angular velocity, and attitude angles (roll, pitch, and yaw).

[0118] The data processing process includes: 1. Preprocessing stage Time synchronization: The hardware timer calibrates the dual sensor clocks. Through the precise timing of the hardware timer, the time consistency of the sensor data is ensured, providing an accurate time reference for subsequent data processing.

[0119] Spatial alignment: Based on the anatomical coordinate system, the head and waist data are unified into the world coordinate system. A coordinate transformation matrix is ​​used to convert the sensor data from its own coordinate system to the world coordinate system, facilitating data fusion and analysis.

[0120] Noise filtering: A 50Hz low-pass filter and an extended Kalman filter are used to output calibrated acceleration, angular velocity, and attitude angle. Depending on the noise characteristics, appropriate filtering methods can be selected to effectively remove noise and improve data quality. Denoising methods can be flexibly selected by those skilled in the art based on test needs and will not be detailed here.

[0121] (3) Processing module: Receives data from the data acquisition module, performs calculations through the neural network model, and outputs the three-dimensional center of gravity trajectory parameters of the human body.

[0122] Receive data: Receive 18-dimensional sensor data + 8-dimensional individual feature data in the current time window for real-time collection and processing to ensure that the model can obtain the latest data in a timely manner.

[0123] Feature Embedding: Individual features are embedded in vectors through a fully connected layer (8→16 dimensions) and concatenated with the sensor data to form a 34-dimensional vector. During the feature embedding process, a nonlinear activation function is used to increase the nonlinear expression capability of the model.

[0124] Convolution processing: A 1×1 convolution is used to compress the image to 18 dimensions, followed by a three-layer dilated convolution to extract temporal features, ultimately outputting 6-dimensional center-of-gravity trajectory parameters. Parallel computing technology is used during the convolution process to increase the model's inference speed.

[0125] Processing Delays: The single-frame inference time is 25ms, meeting real-time detection requirements. By optimizing the model structure and algorithm, the computational complexity of the model is reduced, achieving fast inference.

[0126] (IV) Evaluation and Display Module The center of gravity trajectory parameters output by the receiving processing module are used to calculate the center of gravity offset amplitude D and speed fluctuation rate in real time based on the balance ability evaluation index in the solution. and compared with the preset threshold (good: D < 5cm and <0.2m / s). At the same time, mathematical features are extracted from the center of gravity estimate output by the model, either manually or automatically (using computer graphics algorithms), generating spatiotemporal feature maps (time series diagrams, three-dimensional broken lines, two-dimensional comparisons, etc.), allowing intuitive observation of the center of gravity's movement patterns and distribution patterns. For example, if the center of gravity distribution is found to be relatively discrete in the spatial distribution heat map, this indicates that the human body's ability to control balance during exercise is poor; in the coordinate distribution curve, if the coordinate fluctuations on a certain axis are large and irregular, there may be a balance disorder in that direction. Based on the analysis results of the feature map, further assistance is provided to determine the body's balance ability level, and the center of gravity trajectory curve, balance ability level, and related feature maps are visually displayed on the terminal device (mobile phone / PC), allowing users and professionals to intuitively understand their own or the subject's balance status.

[0127] Artificial extraction: Professionals analyze the changing trends of the center of gravity in different movement stages based on the center of gravity coordinate data, mark abnormal fluctuation points, and judge the balance ability status based on clinical experience.

[0128] Automatic extraction: Using computer graphics algorithms, the three-dimensional coordinate data of the center of gravity over a period of time is processed to create a spatial distribution heat map and a coordinate distribution curve graph of the center of gravity. In the spatial distribution heat map, the color depth represents the frequency of the center of gravity in a certain spatial area, with darker colors indicating a higher probability of the center of gravity appearing in that area. The coordinate distribution curve graph shows how the center of gravity's coordinates on the X, Y, and Z axes change over time.

[0129] The following is an explanation of model training and evaluation.

[0130] 1. Dataset Construction Subject distribution: 200 people (aged 18-75, 1:1 male to female; general population / athletes / physical condition difference groups = 6:2:2. The physical condition difference groups include fitness enthusiasts, sedentary groups, retired active people, etc., to ensure data diversity and improve model generalization).

[0131] Action Type: 8 categories (standing upright, standing on one leg, walking, running, sideways jumping, head turning, trunk leaning forward, and interference movements), covering static maintenance, dynamic movement, and sudden disturbance scenarios, allowing the model to adapt to different sports / daily activities.

[0132] Data augmentation: Add Gaussian noise (σ = 0.05g) and random time offset (±50ms), expanding the sample size to over 100,000. This simulates sensor noise and motion timing variations during motion, enhancing model robustness.

[0133] 2. Training Configuration Optimizer: Adam (lr=0.001, decay rate 0.9), combines the advantages of Adagrad and RMSProp to accelerate convergence.

[0134] Training parameters: The batch size is 32 and the training cycle is 50 epochs. This matches the dataset size with the computing resources to allow the model to fully learn the features.

[0135] Early stopping mechanism: If the validation set error does not decrease for 5 consecutive rounds, training is stopped to avoid overfitting and ensure generalization ability.

[0136] 3. Evaluate the results Core error comparison: Compared with the basic TCN, the optimized model reduced the overall coordinate error by 25%; the error of dynamic movements (such as walking and running) in the middle-aged and elderly group (≥51 years old) was reduced by 40%, verifying the effectiveness of the accuracy improvement.

[0137] Statistical significance: The independent sample t-test showed that the error difference between each group (general population, athletes, groups with different physical conditions, etc.) was p<0.01, indicating that the model can distinguish the balance performance of different groups of people and adapt to various scenarios.

[0138] The device of the specific embodiment is used to implement the method for estimating the center of gravity trajectory and detecting the balance ability of the human body, and a multi-dimensional error analysis is performed on the data after the test. The results are as follows.

[0139] The technical solution of the present invention not only excels in average performance, but also demonstrates strong generalization ability and high precision across a variety of populations and action types.

[0140] Adaptability to Different Populations: As shown in the table below, the optimized model of our invention achieved significant error reduction across all population groups compared to the basic TCN model. In particular, the error for the physical condition difference group (e.g., sedentary vs. fitness) and the middle-aged and elderly group decreased significantly by 40% and 39%, respectively, demonstrating the effectiveness of the personalized feature embedding mechanism. The error differences between all groups were statistically significant (p < 0.01) using independent sample t-tests, demonstrating that the model can effectively distinguish and adapt to the balance characteristics of different populations.

[0141] Table 5 Error by population group (RMSE of center of gravity trajectory, unit: cm) Adaptability to Different Actions: As shown in the table below, the model performs well across eight different action types. In particular, head rotation and forward leaning, two actions that require extremely high dual-sensor coordination and dynamic feature capture, show the greatest reduction in error, reaching 33.3% and 34.5%, respectively. This directly demonstrates the superiority of our dual-IMU layout and differential expansion TCN architecture.

[0142] Table 6 Error by action type (RMSE of center of gravity trajectory, unit: cm) Note: The above data is the result comparison calculated using the same set of test reference samples.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for estimating the trajectory of the human body's center of gravity and detecting balance ability, characterized in that: include: S1, data acquisition: synchronously collecting multi-dimensional motion data output by a first inertial measurement unit worn on the human head and a second inertial measurement unit worn on the human waist; and obtaining multi-dimensional individual characteristic data of the human body; S2, feature fusion: inputting the multi-dimensional individual feature data into the embedding layer to generate an individual feature vector; and concatenating and fusing the individual feature vector with the multi-dimensional motion data to generate fused feature data; S3, model reasoning: inputting the fused feature data into an end-to-end neural network model, performing computational processing, and outputting the three-dimensional center of gravity trajectory parameters of the human body; The neural network model includes at least one dilated convolution layer, which includes a multi-layer differential dilated convolution structure using non-equidistant dilation factors to expand the temporal receptive field and capture long-distance dependencies.

2. The method for estimating the trajectory of the human body center of gravity and detecting the balance ability according to claim 1, characterized in that: The specific steps of S2 include: Inputting the multi-dimensional individual feature data into an adaptive attention fully connected embedding layer, dynamically adjusting the weights according to different individual features, and generating the individual feature vector; splicing the individual feature vector and the multi-dimensional motion data in the feature dimension; The concatenated data is subjected to dimensionality reduction through 1x1 convolution to generate the fused feature data.

3. The method for estimating the trajectory of the center of gravity of the human body and detecting the balance ability according to claim 1, wherein: The end-to-end neural network model is a temporal convolutional network model.

4. The method for estimating the trajectory of the center of gravity of the human body and detecting the balance ability according to claim 3, wherein: The activation function in the temporal convolutional network model is an adaptive activation function. During the model training process, it dynamically switches among multiple preset activation functions according to the feature complexity of the current time step.

5. The method for estimating the trajectory of the center of gravity of the human body and detecting the balance ability according to claim 1, characterized in that: In S1, acquisition time synchronization between the first inertial measurement unit and the second inertial measurement unit is achieved by comparing hardware trigger pulses and high-precision hardware timestamps.

6. The method for estimating the trajectory of the center of gravity of the human body and detecting the balance ability according to claim 1, wherein: The method further comprises, S4, balance ability assessment: Based on the three-dimensional center of gravity trajectory parameters output by S3, calculate at least one balance ability assessment index and evaluate the balance ability of the human body according to the index, the index including the center of gravity offset amplitude and speed fluctuation rate.

7. A device for estimating the trajectory of the center of gravity of a human body and detecting the ability of balance, characterized in that: include: The sensor module includes a first inertial measurement unit arranged on the human head and a second inertial measurement unit arranged on the human waist; a data acquisition module, connected to the sensor module, for synchronously acquiring multidimensional motion data output by the first inertial measurement unit and the second inertial measurement unit, and for acquiring multidimensional individual feature data of the human body; A processing module, connected to the data acquisition module, configured to: Processing the multi-dimensional individual feature data through an embedding layer to generate an individual feature vector; splicing and fusing the individual feature vector with the multi-dimensional motion data to generate fused feature data; The fused feature data is input into an end-to-end neural network model, and the three-dimensional center of gravity trajectory parameters of the human body are calculated and output.

8. The device for estimating the trajectory of the center of gravity of a human body and detecting the balance ability according to claim 7, characterized in that: In the processing module, the embedding layer used to generate individual feature vectors is an adaptive attention fully connected embedding layer, which is used to dynamically adjust weights according to different individual features, and reduce the dimension of the spliced ​​data through 1x1 convolution to generate the fused feature data. The end-to-end neural network model solidified in the processing module is a temporal convolutional network model, which includes a multi-layer differential dilated convolution structure with non-equidistant dilation factors, and an adaptive activation function that can be dynamically selected during training.

9. The device for estimating the trajectory of the center of gravity of a human body and detecting the balance ability according to claim 7, wherein: The device further comprises: An evaluation and display module is connected to the processing module and is used to calculate the center of gravity offset amplitude and / or speed fluctuation rate based on the output three-dimensional center of gravity trajectory parameters, evaluate the balance ability of the human body, and visualize the center of gravity trajectory curve, balance ability level or spatial distribution characteristic diagram of the center of gravity.

Images