A method and device for estimating human center of gravity trajectory and detecting balance ability

Abstract

This invention relates to the field of human motion detection and balance ability assessment, and discloses a method and device for estimating the center of gravity trajectory and detecting balance ability. The method simultaneously collects multidimensional motion data from a nine-axis inertial measurement unit worn on the head and waist, as well as multidimensional individual feature data of the human body. Individual features are embedded through an adaptive attention mechanism and deeply fused with the motion data. The fused data is input into a specially designed temporal convolutional network model containing a differentially dilated convolutional structure and an adaptive activation function, achieving end-to-end direct mapping from the original data to the three-dimensional center of gravity trajectory parameters. This invention abandons physical model assumptions, fundamentally eliminates model errors, significantly improves computational efficiency and estimation accuracy in dynamic scenes, and can adaptively learn individual differences, exhibiting strong generalization ability and meeting the requirements for high-precision real-time balance ability detection.

Description

Technical Field

[0001] This invention relates to the field of human motion detection and balance ability assessment, and in particular to a method and apparatus for estimating the trajectory of the human body's center of gravity and detecting balance ability. Background Technology

[0002] Current general-purpose technology uses plantar pressure plate (matrix) measurement, which is based on the physical model of the human body's center of gravity calculated from data from a single gyroscope.

[0003] (a) Deficiencies in existing technology

[0004] Current technologies for estimating the trajectory of the human body's center of gravity and detecting balance ability are based on traditional plantar pressure plates (matrix) measurements. This approach can only measure the projection data of the body's center of gravity onto a two-dimensional plane (XY axis), failing to capture changes in the vertical (Z-axis) center of gravity height. More importantly, plantar pressure signals cannot distinguish between leg and torso movement characteristics—when the torso leans forward or the head rotates, the pressure plate can only capture changes in plantar pressure distribution, but cannot analyze the direct impact of torso posture changes on the center of gravity.

[0005] In balance assessment scenarios, trunk motion characteristics (such as trunk tilt angle and rotational angular velocity) have a significantly higher weight on the center of gravity trajectory than leg movements. Biomechanical analysis reveals that trunk posture changes account for 73% of the control of the center of gravity in human dynamic posture control. However, traditional plantar pressure plates, which only collect plantar pressure data, cannot directly obtain trunk motion information. Consequently, in center of gravity trajectory analysis scenarios involving the trunk (such as daily walking posture analysis and motion assessment), the accuracy of attributing changes in the center of gravity is less than 50%.

[0006] While solutions relying solely on waist sensors can acquire some torso data, they neglect the impact of head movements on the overall center of gravity.

[0007] Head movements are a "high-frequency dynamic factor" in human balance regulation. When the head rotation angular velocity is >30° / s (commonly reaching 40-60° / s in everyday rapid head turns), its instantaneous impact on the center of gravity trajectory accounts for 22%–25%. Furthermore, the mechanical transmission path of head movements (skull → cervical spine → trunk) is asynchronous with the "trunk-lumbar spine" data collected by the lumbar sensor. A single lumbar sensor cannot cover the complete temporal characteristics of this transmission chain, from "head triggering" to "trunk response" (neural transmission delay is approximately 15-30ms, corresponding to dynamic changes in the center of gravity trajectory requiring synchronous acquisition by both head and lumbar sensors within ≤10ms). If only lumbar sensors are used, the correlation between "head triggering signal" and "trunk response signal" will be lost, causing the model to be unable to distinguish the impact of "active head rotation" and "compensatory trunk swaying" on the center of gravity, resulting in a sharp increase in estimation error.

[0008] (ii) Inadequacy of existing models

[0009] Physical models based on dynamic equations assume the human body is rigid, requiring manual input of parameters such as bone length and joint angles, which are difficult to measure precisely, and cannot adaptively learn individual differences. In dynamic scenarios (such as running or standing on one leg), joint flexibility is ignored.

[0010] Using a traditional rigid body assumption physical model (such as a human dynamics model based on the Newton-Euler equations), the skeletal 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, 0.1mm accuracy) + force platform (3D force accuracy 0.1N) was used to acquire the actual center of gravity trajectory (gold standard data) for comparison.

[0011] For 30 individuals, each with 5 sets of dynamic movements (a total of 150 sets of data), the root mean square error (RMSE) of the physical model output and the precise control data was calculated. Statistical results show that the average error of the model in the running scenario reached 16.2%, the peak error in the single-leg standing (eyes closed) scenario reached 21.7%, and the proportion of models with errors exceeding 15% in the overall dynamic scenarios reached 68% (102 / 150 sets).

[0012] Therefore, there is an urgent need for a balance ability detection method and device that can achieve three-dimensional dynamic human body center of gravity trajectory estimation, while also being adaptable to different groups and scenarios and possessing sufficient generalization ability. This is to meet the requirements for improving the accuracy of human body center trajectory estimation and balance ability detection. Summary of the Invention

[0013] The present invention aims to at least solve one of the technical problems existing in related technologies. To this end, the present invention provides a method and apparatus for estimating the trajectory of the human body's center of gravity and detecting balance ability.

[0014] The first implementation plan is given below.

[0015] A method for estimating the trajectory of a person's center of gravity and detecting balance ability includes:

[0016] S1, Data Acquisition: Simultaneously acquire multidimensional motion data output from the first inertial measurement unit worn on the human head and the second inertial measurement unit worn on the human waist; and obtain multidimensional individual characteristic data of the human body;

[0017] S2, Feature Fusion: The multidimensional individual feature data is input into the embedding layer to generate an individual feature vector; and the individual feature vector is concatenated and fused with the multidimensional motion data to generate fused feature data;

[0018] S3, Model Inference: Input the fused feature data into the end-to-end neural network model, perform calculations, and output the three-dimensional center of gravity trajectory parameters of the human body;

[0019] The neural network model includes at least one dilated convolutional layer, which comprises a multi-layer differential dilated convolutional structure employing a non-equidistant dilation factor to expand the temporal receptive field and capture long-range dependencies.

[0020] Furthermore, the specific steps of S2 include:

[0021] The multidimensional individual feature data is input into an adaptive attention fully connected embedding layer, and the weights are dynamically adjusted according to different individual features to generate the individual feature vector;

[0022] The individual feature vector is concatenated with the multidimensional motion data along the feature dimension;

[0023] The fused feature data is generated by reducing the dimensionality of the concatenated data through 1x1 convolution.

[0024] Furthermore, the end-to-end neural network model is a temporal convolutional network model.

[0025] Furthermore, the activation function in the temporal convolutional network model is an adaptive activation function, which dynamically switches among various preset activation functions according to the feature complexity of the current time step during model training.

[0026] Furthermore, in S1, the acquisition time of the first inertial measurement unit and the second inertial measurement unit is synchronized by comparing hardware trigger pulses and high-precision hardware timestamps.

[0027] Furthermore, the method also includes,

[0028] 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 assess the human body's balance ability according to the index. The index includes the center of gravity offset amplitude and velocity fluctuation rate.

[0029] Based on the first implementation plan, a second implementation plan is presented.

[0030] A device for estimating the trajectory of a person's center of gravity and detecting their balance ability includes:

[0031] The sensor module includes a first inertial measurement unit positioned at the head of the human body and a second inertial measurement unit positioned at the waist of the human body.

[0032] The data acquisition module, connected to the sensor module, is used to synchronously acquire multidimensional motion data output by the first inertial measurement unit and the second inertial measurement unit, and to obtain multidimensional individual characteristic data of the human body.

[0033] The processing module, connected to the data acquisition module, is configured for:

[0034] The multidimensional individual feature data is processed through an embedding layer to generate individual feature vectors;

[0035] The individual feature vectors are concatenated and fused with the multidimensional motion data to generate fused feature data;

[0036] The fused feature data is input into an end-to-end neural network model, which calculates and outputs the three-dimensional center of gravity trajectory parameters of the human body.

[0037] Furthermore, in the processing module, the embedding layer used to generate individual feature vectors is an adaptive attention fully connected embedding layer. This embedding layer is used to dynamically adjust the weights according to different individual features and reduce the dimensionality of the concatenated data through 1x1 convolution to generate the fused feature data. The end-to-end neural network model fixed in the processing module is a temporal convolutional network model. This model includes a multi-layer differential dilation convolutional structure with non-equidistant dilation factors and an adaptive activation function that can be dynamically selected during training.

[0038] Furthermore, the device also includes:

[0039] The evaluation and display module, connected to the processing module, is used to calculate the center of gravity offset amplitude and / or velocity fluctuation rate based on the output three-dimensional center of gravity trajectory parameters, evaluate the human body's balance ability, and visualize the center of gravity trajectory curve, balance ability level, or spatial distribution feature map of the center of gravity.

[0040] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0041] This invention abandons the physical model assumptions, fundamentally eliminates model errors, significantly improves computational efficiency and estimation accuracy in dynamic scenarios, and can adaptively learn individual differences, with strong generalization ability, which can meet the requirements of high-precision real-time balance capability detection.

[0042] Hardware-level time synchronization (error <0.05ms) ensures the consistency of the timing of head and waist data, avoiding errors in center of gravity trajectory estimation caused by time differences (error magnitude <0.1mm).

[0043] Dual IMUs in the head and waist capture high-frequency dynamic features of the vestibule-neck-trunk (such as rapid head turning angular velocity >40° / s) and low-frequency features of the trunk-lower limbs, making up for the shortcomings of a single sensor (such as only waist IMU) and reducing the center of gravity estimation error (such as the error in static standing scenes is reduced from 3.2cm to 1.5cm).

[0044] Introducing individual characteristic data provides a foundation for personalized modeling in the future, and improves the model's adaptability to different groups (such as the elderly and athletes).

[0045] Normalization and standardization improve the training efficiency of neural networks and reduce gradient instability caused by differences in data dimensions.

[0046] By adopting 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 flexible joint motion scenarios (such as squatting).

[0047] The attention mechanism embeds individual features, enhancing the model's adaptability to different populations (such as the middle-aged and elderly group), reducing the error by 28%. Through end-to-end learning, the error of the physical model assumption is eliminated, and the dynamic scene error is significantly reduced. The dynamic scene error using only the end-to-end neural network model is 5%–8%, while the dynamic scene error using the physical model is 15%–20%.

[0048] Visualized results (such as heat maps and graphs) help users and professionals quickly identify balance problems (such as excessive fluctuations in a certain axis), assisting in clinical diagnosis or exercise training. Combined with assessment results based on individual characteristics, they can be adapted to different populations (such as a 39% reduction in error in the middle-aged and elderly group), improving the relevance of the assessment.

[0049] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the human body center of gravity trajectory estimation and balance ability detection method of the present invention;

[0052] Figure 2This is a structural diagram of the human body center of gravity trajectory estimation and balance ability detection device of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0054] 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.

[0055] To address the shortcomings of existing technologies in terms of accuracy in estimating three-dimensional center of gravity trajectory, dynamic feature capture capability, and individual adaptability, the inventors propose an innovative architecture (Scheme 1) that integrates physical models and deep learning. This scheme synchronously acquires multi-dimensional motion data using multimodal inertial sensors (nine-axis IMUs for the head and waist), combining hierarchical physical modeling with adaptive neural network optimization. This not only compensates for the simplification errors of traditional physical models regarding joint flexibility but also enhances the learning ability for individual differences and high-frequency dynamic features through a data-driven approach, ultimately achieving high-precision estimation of the human body's center of gravity trajectory and quantitative detection of balance ability.

[0056] The specific implementation of Option 1 is as follows:

[0057] The system collects multidimensional motion data (18 dimensions) from a nine-axis sensor covering the head and waist, as well as measured or estimated values ​​of individual skeletal parameters. Multidimensional motion data is acquired via a high-precision AD converter built into the sensor, achieving a conversion accuracy of 16 bits to ensure data accuracy. Individual skeletal parameter measurements are obtained from medical imaging equipment (such as X-rays and CT scans) to obtain key bone lengths, joint angles, etc. If measurements are unavailable, estimations are made based on statistical models and individual physical characteristics (height, weight, etc.), or manual specification by experienced professionals is required.

[0058] The estimation error of the center of gravity trajectory between a single IMU (waist only) and two IMUs (head + waist) was calculated (based on the Vicon optical system accuracy of 0.1mm, the root mean square error RMSE was calculated). The comparison scheme selected 30 subjects (aged 20-50, covering the general population and sports enthusiasts) and conducted the study in three typical scenarios: static standing (eyes closed), dynamic walking (normal walking speed), and rapid head turning + trunk swaying. The results are shown in the table below.

[0059] Table 1: Comparison of center of gravity trajectory estimation errors between a single IMU (waist only) and two IMUs (head + waist).

[0060]

[0061] The reasons for this are as follows: the lumbar IMU only covers the "trunk-lower limb" kinetic chain, while the head IMU can capture high-frequency posture changes transmitted through the "vestibular-neck-trunk" chain (e.g., when the head turns rapidly at an angular velocity > 40° / s, the head IMU can sense the signal 15-30ms earlier), compensating for the response delay of the lumbar sensor to "head-triggered center of gravity disturbances"; at the same time, the fusion of the two sets of IMU data can distinguish the contribution of "active head movements (such as turning the head)" and "compensatory trunk movements (such as bending over)" to the center of gravity. In the compound action of "head rotation + trunk forward tilt", the single IMU scheme misjudges the "contribution of head movement" to 35%, while the two IMU scheme can control the decoupling error to within 8%.

[0062] Unlike traditional "single-site IMU" solutions (such as only waist / wrist), this solution's two-set IMU collaborative mechanism has unique advantages: it does not rely on "rigid body assumption" correction (directly covers the "head-chest-waist" key kinematic chain, reducing model assumption errors); it adapts to the universality of dynamic scenarios (in high-frequency scenarios such as running and jumping, the error of the single IMU solution increases linearly with the complexity of the action (R²=0.87), while the error growth trend of the two-set IMU solution is gradual (R²=0.32)), verifying the robustness of multi-site sensing to dynamic scenarios.

[0063] R² (coefficient of determination) is a measure of the goodness of fit of a regression model. Mathematically, it represents the proportion of the variation in the dependent variable that can be explained by the independent variables. 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.

[0064] The core improvements in Scheme 1 include physical model construction and neural network computation.

[0065] The physical model construction includes establishing kinematic mapping relationships based on a simplified head-thorax-lower back skeleton model (treating the head and chest, and the chest and lower back as rotatable connections) using the Newton-Euler equations. A layered modeling approach is employed, first modeling the head, chest, and lower back separately, then integrating them through connection points. Combined with the input individual skeletal parameters, a preliminary estimate of the center of gravity (3D coordinates G) is output. x G y G z The human body is simplified into three parts: head, chest, and lower chest, with the connections between these parts following the laws of rotation. To reflect real-world conditions, the range of head rotation relative to the chest is limited to ±60°, and the range of chest rotation relative to the lower chest is limited to ±30°. Furthermore, the model's range of motion and mechanical properties are adjusted based on individual skeletal parameters.

[0066] The neural network calculation includes: center of gravity estimate based on the output of the physical model + multidimensional motion data (18 dimensions) + multidimensional individual feature data (8 dimensions, including age, height, gender, sitting height, waist circumference, chest circumference, hip circumference, and length of bone from foot to knee joint) + individual skeletal parameters;

[0067] The unit for length or height data is cm.

[0068] Before inputting the data, the original sensor time series data is normalized by using the minimum-maximum normalization method to map the data to the [0,1] interval, thereby improving the training efficiency of the neural network.

[0069] The neural network employs a simple fully connected layer structure, consisting of two fully connected layers. The first layer contains 32 neurons, and the second layer contains 6 neurons. It directly processes the input data and outputs 6-dimensional centroid trajectory parameters (3D coordinates + 3D velocity). The loss function is optimized using the MSE loss function. To prevent overfitting, an L2 regularization term is added to the loss function, with a regularization coefficient set to 0.001.

[0070] The implementation process of Option 1 is as follows:

[0071] Data collection:

[0072] Simultaneously, 18-dimensional multi-dimensional motion data (100Hz) of the head / waist, individual skeletal parameter measurements (if applicable), and multi-dimensional individual feature data were acquired. Ground truth labels for the center of gravity were obtained using an optical capture system (error <2cm). Individual skeletal parameter estimates were calculated using a pre-trained regression model combined with multi-dimensional individual feature data. Multi-threading technology was employed during data acquisition to ensure synchronous acquisition of various data types and avoid data loss.

[0073] Physical model calculations:

[0074] Based on sensor data, the relative joint angles between the head and chest, and between the chest and subchest are calculated. Combined with measured or estimated values ​​of individual skeletal parameters, the preliminary center of gravity position is calculated using dynamic equations. When calculating the joint angles, a quaternion method is employed to avoid gimbal lock issues and improve calculation accuracy.

[0075] Neural network optimization:

[0076] The physical model output, multidimensional motion data, multidimensional individual feature data, and individual skeletal parameters are fused and input into a fully connected neural network to learn dynamically correct parameters, compensating for simplification assumption errors in the physical model. During model training, cross-validation is used, dividing the dataset into training, validation, and test sets in a 7:2:1 ratio.

[0077] After verification, Scheme 1 has significantly improved accuracy compared to traditional testing schemes, but the following problems still exist.

[0078] 1. Parameter sensitivity:

[0079] The physical model is highly dependent on the accuracy of multidimensional motion data and individual skeletal parameters. Both measurement errors and estimation biases can increase the error in center of gravity estimation. At the same time, simple fully connected neural networks are unable to fully extract complex temporal features, affecting the model's accuracy.

[0080] 2. Insufficient dynamic adaptability:

[0081] The physical model is based on simplified structural assumptions, which makes it difficult to adapt to joint flexibility in real time. It is inefficient in correcting errors in fast-moving scenarios, and the simple structure of the neural network makes it unable to effectively handle dynamically changing data.

[0082] 3. High computational complexity:

[0083] Solving the dynamic equations requires matrix iteration operations. In addition, the calculation or processing of individual skeleton parameters further increases the processing time of a single frame, far exceeding 80ms, making it even more difficult to meet the requirements of real-time detection (requirement <50ms).

[0084] After systematic research and analysis, a pure neural network adaptive scheme (Scheme Two) was proposed. This scheme, centered on end-to-end deep learning, abandons the structural assumptions of traditional physical models. Through deep fusion of multimodal sensor data and individual features, combined with the enhanced ability of dilated convolutional networks to capture temporal dynamic features, it achieves a direct mapping from raw data to the center of gravity trajectory. Compared to Scheme One, it significantly optimizes model generalization, computational efficiency, and accuracy in complex action scenarios, providing a more robust data-driven solution for balance ability detection.

[0085] Figure 1 The implementation flow of the detection method in Scheme 2 is shown.

[0086] A method for estimating the trajectory of the human body's center of gravity and detecting balance ability.

[0087] S1, Data Acquisition: Simultaneously acquire multidimensional motion data output from the first inertial measurement unit worn on the human head and the second inertial measurement unit worn on the human waist; and obtain multidimensional individual characteristic data of the human body;

[0088] The acquisition time of the first inertial measurement unit and the second inertial measurement unit is synchronized by comparing hardware trigger pulses and high-precision hardware timestamps.

[0089] S2, Feature Fusion: The multidimensional individual feature data is input into the embedding layer to generate an individual feature vector; and the individual feature vector is concatenated and fused with the multidimensional motion data to generate fused feature data;

[0090] The multidimensional individual feature data is input into an adaptive attention fully connected embedding layer, and the weights are dynamically adjusted according to different individual features to generate the individual feature vector.

[0091] The individual feature vector is concatenated with the multidimensional motion data along the feature dimension;

[0092] The fused feature data is generated by reducing the dimensionality of the concatenated data through 1x1 convolution.

[0093] S3, Model Inference: Input the fused feature data into the end-to-end neural network model, perform calculations, and output the three-dimensional center of gravity trajectory parameters of the human body;

[0094] The end-to-end neural network model is a temporal convolutional network model.

[0095] The temporal convolutional network model includes at least one dilated convolutional layer, which comprises a multi-layer differential dilated convolutional structure employing a non-equidistant dilation factor to expand the temporal receptive field and capture long-range dependencies.

[0096] The activation function in the temporal convolutional network model is an adaptive activation function, which dynamically switches among various preset activation functions during model training based on the feature complexity of the current time step.

[0097] 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 assess the human body's balance ability according to the index. The index includes the center of gravity offset amplitude and velocity fluctuation rate.

[0098] The core improvements in Scheme 2 include abandoning the reliance on physical models, adopting an individual feature embedding mechanism, and dynamic feature enhancement.

[0099] By abandoning reliance on physical models, the center of gravity trajectory mapping relationship is learned directly from "multidimensional motion data + individual feature data" through end-to-end deep learning, avoiding systematic errors caused by simplified structural assumptions.

[0100] Based on the methods for estimating the trajectory of the human body's center of gravity and testing balance ability, specific implementation examples are given.

[0101] 100 sets of dynamic action data (running, jumping, agile change of direction, etc.) were selected, and the error performance of Scheme 1 (physical model + neural network) and Scheme 2 (pure end-to-end learning) were compared.

[0102] Option 2 reduces the error by 35% compared to Option 1 (RMSE decreases from 5.2cm to 3.4cm). In scenarios involving flexible joint movements (such as squatting and standing up), Option 1 has an error rate of 18% due to the rigid body assumption, while Option 2 can control it to within 7%.

[0103] By using pre-trained model parameters from large-scale public datasets (such as Human3.6M and AMASS), this solution addresses the "head + waist multimodal + individual features" scenario. Through a feature adapter (a unique design distinct from general transfer learning), the "general human motion features" and "individual difference features" of the pre-trained model are decoupled, improving the model convergence speed by 40% (reducing the number of iterations from 200 rounds to 120 rounds), thus solving the training efficiency problem in small sample scenarios.

[0104] Individual feature embedding mechanism:

[0105] Eight-dimensional individual feature data (including age, height, gender, sitting height, waist circumference, chest circumference, hip circumference, and length of the bone from the sole of the foot to the knee joint, with the larger units of length and height being cm) are introduced. A 16-dimensional feature vector is generated through an adaptive attention fully connected embedding layer. This vector is then concatenated with the data obtained from the 18-dimensional sensor and compressed to 18 dimensions through a 1×1 convolution to guide the network in learning individual-specific movement patterns.

[0106] Furthermore, a specific design was added. Unlike the "general embedding + fixed weight" approach, the attention weights in this mechanism are dynamically adjusted according to individual characteristics: for example, the "height-center of gravity height" correlation weight is 32% higher for the elderly (age > 60 years) than for younger people; the "weight-horizontal center of gravity offset" weight is 27% higher for men than for women. By embedding an individual feature attention matrix (8×16 dimensions) into the fully connected layer, personalized binding of "feature-weight" is achieved, reducing the center of gravity trajectory estimation error by 28% in scenarios with significant individual differences (such as comparing the elderly with athletes).

[0107] We then conducted ablation experiments to verify that after removing the "individual feature embedding" or "attention mechanism", the average error of the model for subjects of different heights and weights increased by 25% and 19% respectively, proving that this mechanism plays an important role in the learning of individual-specific movement patterns.

[0108] Dynamic feature enhancement design:

[0109] A three-layer differential dilated convolutional structure (dilation factor 1 / 3 / 5, different from the general isometric dilation 1 / 2 / 4) is adopted to expand the receptive field to 45 time steps layer by layer. Combined with an 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 a BN layer, the ability to capture high-frequency dynamic features such as rapid head rotation (angular velocity > 50° / s) and trunk forward tilt is improved.

[0110] It should be noted that in general dilated convolutions, isometric dilation (e.g., factors 1 / 2 / 4) can easily lead to "feature hole overlap," resulting in a feature loss rate of up to 22% in scenarios with rapid head rotation (angular velocity of 60° / s).

[0111] 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.

[0112] Traditional fixed activation functions (such as a single ReLU or Leaky ReLU) have significant limitations during model training: when dealing with complex dynamic features such as rapid head rotation and forward torso tilt, the model is prone to getting stuck in local extrema due to gradient vanishing, causing training convergence to stall or slow down significantly. For example, in training scenarios involving high-frequency motion temporal features, fixed activation functions will significantly increase the number of iterations in local extrema regions, prolonging the overall training cycle.

[0113] The proposed adaptive activation function mechanism optimizes convergence characteristics through a dynamic learning strategy.

[0114] 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 temporal correlation). For example, when processing high-frequency acceleration signals generated by head rotation, the model adaptively selects Leaky ReLU to preserve gradient information of negative features, avoiding gradient vanishing due to the "neuron death" problem of ReLU.

[0115] Local extrema avoidance: By evaluating gradient change trends in real time, when a slowdown in model convergence is detected (e.g., the loss decrease is less than a preset threshold over multiple iterations), the activation function type is automatically switched, and the neuron activation mode is adjusted. This mechanism effectively reduces the time the model spends in local extrema regions, improving the stability of the training process.

[0116] Experimental results show that, compared with a fixed activation function, this adaptive mechanism can significantly reduce the convergence stagnation caused by local extrema during the training process of the model in complex dynamic scenarios, and optimize both the number of training rounds and the convergence time, which meets the requirements of dynamic equilibrium detection tasks for model real-time performance and robustness.

[0117] A specific implementation of Scheme 2 is described below.

[0118] 1. Multimodal data fusion:

[0119] A nine-axis IMU for the head (forehead) and a nine-axis IMU for the lumbar spine simultaneously acquire 18-dimensional motion data. High-precision time synchronization is achieved through a hardware-level synchronous sampling and control architecture. The specific solution is as follows:

[0120] Hardware collaboration:

[0121] An IMU sensor with a built-in synchronous sampling control module is selected. The main control SOC outputs a hardware trigger pulse (matching the sensor sampling rate) through the GPIO pin, thereby ensuring that the IMU starts synchronous sampling after receiving the trigger signal, and ensuring that the sampling time of the two IMUs is consistent.

[0122] Timestamp and Error Verification:

[0123] Each IMU embeds a 64-bit hardware timestamp (generated by a high-precision timer inside the sensor, with a timing accuracy better than 1μs) when outputting data. The data processing unit calculates the synchronization error by comparing the timestamps of the two IMUs. Validated using 100,000 sets of continuous sampling data. The maximum value is 0.8 μs, and the time fluctuation of 99.9% of the data is less than 0.05 ms, meeting the synchronization requirements in high-frequency motion scenarios. The hardware timestamp embedded when the head IMU collects data. Hardware timestamps embedded when acquiring data from the waist IMU.

[0124] Quantification of the impact of error:

[0125] The error in the center of gravity trajectory caused by time synchronization fluctuations can be derived through a kinematic model: Assuming the height of a human body's center of gravity is approximately 1.5m, the head and waist IMUs will have errors due to time differences. spatial displacement error for: (a represents the acceleration of human motion, taking a typical value of 5 m / s² in a running scenario) 2 Substitute ,have to The error is much smaller than the sensor measurement noise (approximately 0.1 mm), verifying that the error of this synchronization scheme is negligible and ensuring the accuracy of data fusion.

[0126] 2. TCN Model Architecture

[0127] Input layer:

[0128] T×34-dimensional data (T=100 time steps, 34-dimensional = 18-dimensional sensor data + 16-dimensional individual feature vector).

[0129] The input data is a time-series-feature two-dimensional matrix, defined as:

[0130] in:

[0131] T=100, time step, each step corresponds to a 10ms sampling interval, covering continuous motion data within 1 second;

[0132] 34-dimensional features = 18-dimensional motion data + 16-dimensional individual feature vectors;

[0133] 18-dimensional motion data: Head 9-axis IMU (3D acceleration + 3D angular velocity + 3D magnetic field) + Waist 9-axis IMU (3D acceleration + 3D angular velocity + 3D magnetic field), for ,Right now ;

[0134] Head acceleration [x,y,z] 3D row vector;

[0135] : Head angular velocity [x,y,z] 3D row vector;

[0136] : Head magnetic field [x,y,z] 3-dimensional row vector;

[0137] : Waist acceleration [x,y,z] 3D row vector;

[0138] : Waist angular velocity [x,y,z] 3D row vector;

[0139] The 3-dimensional row vector of the waist magnetic field [x,y,z].

[0140] 16-dimensional individual feature vector: generated from 8-dimensional original individual features (height, weight, sitting height, age, etc.) through an adaptive attention fully connected embedding layer. (The same individual feature vector is reused at each time step to ensure temporal consistency).

[0141] The input data is processed in blocks, with each block containing data at a certain time step, which facilitates parallel computation of the model and improves computational efficiency.

[0142] Feature fusion layer:

[0143] 1×1 convolutions compress 34 dimensions to 18 dimensions, enabling cross-modal feature interaction. During feature fusion, a channel attention mechanism is employed, assigning different weights based on the importance of features in different channels, thereby enhancing the model's ability to extract key features.

[0144] The specific explanation is as follows:

[0145] The feature fusion layer achieves cross-modal feature compression and fusion through 1×1 convolution operations. The specific process is as follows: Input matrix: Convolution kernel parameters: using A 1×1 convolution kernel (size 1×1, input channels 34, output channels 18).

[0146] Dimensionality reduction computation:

[0147]

[0148] in This represents a convolution operation, and the output matrix is... .

[0149] Dilated convolutional layer:

[0150] The system consists of 3 layers (dilation factors 1 / 3 / 5) with a kernel size of 5, capturing long-range dependencies in time series data. Residual connections are used within the dilated convolutional layers to avoid the vanishing gradient problem and improve the model's training stability.

[0151] The dilated convolutional layer consists of three differentially dilated convolutional structures. The matrix form of the input and output of each layer and the logic for expanding the receptive field are as follows:

[0152] Table 2 Differentiated dilated convolutional structures with 3 layers of dilated convolution.

[0153]

[0154] K1, K2, and K3 are the convolution kernel parameters corresponding to the three convolutional layers, respectively. Its characteristics are indicated as 3D matrix;

[0155] X1, X2, and X3 are the output matrices corresponding to the three convolutional layers. Its characteristics are indicated as 3D matrix.

[0156] Explanation: The convolution kernel size is fixed at 5×1 (5 steps in the time dimension, 1 step in the feature dimension) to ensure the continuity of temporal features; the non-equidistant dilation factor (1→3→5) causes the receptive field to expand non-linearly, avoiding the "feature holes" caused by the equidistant factor (e.g., the traditional factor 1→2→4 will result in a 12% feature loss at T=100, which this scheme reduces to less than 5%); the output matrix is ​​preserved. The dimension ensures temporal integrity (the features at each time step correspond one-to-one with the original samples), preserving temporal correlation for subsequent dynamic estimation of the centroid trajectory.

[0157] Output layer:

[0158] The model employs 6-dimensional centroid trajectory parameters and a loss function consisting of MSE plus an individual feature regularization term (λ=0.01), forcing the model to learn individual differences. During training, the regularization coefficient is dynamically adjusted based on the model's performance to balance its fitting and generalization abilities.

[0159] After extracting temporal features through 3 layers of dilated convolution, then... The convolutional + fully connected layer outputs 6-dimensional centroid trajectory parameters at each time step, in vector form: The 6-dimensional parameters are defined as follows: These correspond to the user's Cartesian coordinate system (unit: m) and three-dimensional velocity (unit: m / s) based on the world coordinates and orientation at the initial moment, respectively.

[0160] The coordinates of the centroid; The velocity vector relative to the previous point;

[0161] Y out This is a fully linked output vector with T rows and 6 columns.

[0162] The core basis for dynamic adjustment is to define two key indicators as adjustment trigger conditions:

[0163] Loss difference ratio:

[0164] in, Let MSE be the training set in round t. Let be the MSE loss of the validation set in round t. Reflecting the risk of model overfitting: The larger the value, the higher the risk of overfitting (the validation loss is much greater than the training loss).

[0165] Rate of change of loss:

[0166] The trend reflecting validation loss: At that time, the validation loss increases and the model's generalization ability decreases.

[0167] The update formula for the regularization coefficient λ is:

[0168] Adjustment conditions:

[0169] Condition 1: Increase λ (strengthen regularization): when (Loss difference ratio ≥ 30%) and When the validation loss increases, λ increases by 20% to force the model to reduce overfitting to individual features (e.g., to avoid overfitting to extreme skeletal parameters of a certain population group).

[0170] Condition 2: Decrease λ (weaken regularization): when (Loss difference ratio ≤ 10%) and When verifying loss decreases or stabilizes, λ is reduced by 20%, allowing the model to learn more fully about individual differences (such as differences in bone parameters between older adults and athletes).

[0171] In other cases where conditions 1 or 2 are not met, λ remains unchanged.

[0172] Below is an example of dynamic adjustment.

[0173] From the initial For example, let's simulate the adjustment process of 5 rounds of training:

[0174] Table 3. Adjustment process of simulation training

[0175]

[0176] Balance ability assessment indicators

[0177] Center of gravity shift range:

[0178]

[0179] D: It can usually represent a certain distance (such as the distance between two points in space, which is determined in combination with the specific application scenario, and the formula is similar to the spatial distance formula).

[0180] , , : Represents the coordinates of a point in the x, y, z dimensions (such as the three coordinate axes of a spatial rectangular coordinate system) or related physical quantities, characteristic quantities, etc.

[0181] , , :Usually , , The mean of the corresponding dimension is the mean of the static standing center of gravity. By calculating the magnitude of the center of gravity shift, the stability of the human body during movement can be assessed.

[0182] Velocity volatility:

[0183]

[0184] It reflects the stability of the center of gravity shift. The smaller the velocity fluctuation rate, the more stable the center of gravity shift, and the better the human body's balance ability.

[0185] in:

[0186] : This usually represents the standard deviation associated with the variable V, and is used to measure The degree of dispersion of the data.

[0187] N: Represents the total number of terms in the summation. It is a positive integer representing the number of data points.

[0188] : is the nth observation or data point, where n ranges from 1 to N.

[0189] :yes The mean is used to measure the central tendency of the data.

[0190] The following comparison between Scheme 1 and Scheme 2 further illustrates the advantages of Scheme 2 in terms of testing accuracy and generalization ability.

[0191] Table 4 Comparison of the advantages of each option

[0192]

[0193] 5. Through actual analysis and comparison, Scheme 1 has achieved a significant improvement in accuracy compared to the existing technology of solving the human body center of gravity physical model based on single gyroscope data, which uses plantar pressure plate (matrix) measurement. Scheme 2, through optimization of convolutional network, not only did not reduce accuracy but also further improved it, while also improving computational speed and generalization ability.

[0194] Specifically, it fundamentally eliminates model assumption errors, improves model convergence speed by 40%, provides more than 3 times the computational efficiency enhancement, reduces dynamic errors by more than 50%, and features a network structure specifically designed for temporal dynamic features with dynamic selection of activation functions for greater generalization capability.

[0195] Figure 2 The structure of the device for estimating the trajectory of the human body's center of gravity and detecting balance ability in implementing the second method is shown.

[0196] The following describes a device for estimating the center of gravity trajectory and detecting balance ability of a human body, which is used to implement the method of estimating the center of gravity trajectory and detecting balance ability.

[0197] The setup of the human body center of gravity trajectory estimation and balance ability detection device is explained, and the parts of the device that operate the human body center of gravity trajectory estimation and balance ability detection method will not be described in detail.

[0198] (a) Sensor Module:

[0199] Head-mounted nine-axis IMU (M1): integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, fixed to the forehead, with a coordinate system consistent with the human sagittal plane.

[0200] Lumbar 9-axis IMU (M2): Same specification model, installed in the middle of the lumbar spine, with the coordinate system aligned with the trunk axis.

[0201] (II) Data Acquisition Module:

[0202] Simultaneously 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-input individual features.

[0203] 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.

[0204] Data processing flow:

[0205] Synchronous calibration:

[0206] Dual-sensor time synchronization (error <1μs) is achieved through a hardware timer, and spatial alignment is completed based on the anatomical coordinate system.

[0207] Preprocessing:

[0208] Data filtering removes high-frequency noise such as muscle vibration; extended Kalman filtering fuses multi-sensor data to output calibrated acceleration, angular velocity, and attitude angles (roll, pitch, and yaw).

[0209] The data processing flow includes:

[0210] 1. Preprocessing stage

[0211] Time synchronization:

[0212] The hardware timer calibrates the dual-sensor clock, ensuring the time consistency of sensor data through precise timing and providing an accurate time reference for subsequent data processing.

[0213] Spatial alignment:

[0214] Based on the anatomical coordinate system, head and waist data are unified to the world coordinate system. A coordinate transformation matrix is ​​used to transform sensor data from its own coordinate system to the world coordinate system, facilitating data fusion and analysis.

[0215] Noise filtering:

[0216] A 50Hz low-pass filter combined with an extended Kalman filter outputs calibrated acceleration, angular velocity, and attitude angles. Appropriate filtering methods are selected based on different noise characteristics to effectively remove noise and improve data quality. Those skilled in the art can flexibly choose denoising methods according to testing needs, and these will not be elaborated upon here.

[0217] (III) Processing Module: Receives data from the data acquisition module, performs calculations through a neural network model, and outputs the three-dimensional center of gravity trajectory parameters of the human body.

[0218] Received data:

[0219] It receives and processes 18-dimensional sensor data and 8-dimensional individual feature data from the current time window in real time to ensure that the model can acquire the latest data in a timely manner.

[0220] Feature embedding:

[0221] Individual features are used to generate embedding vectors through fully connected layers (8→16 dimensions), which are then concatenated with sensor data to form a 34-dimensional vector. During feature embedding, a non-linear activation function is employed to enhance the model's non-linear expressive power.

[0222] Convolution processing:

[0223] A 1×1 convolution is used to compress the model to 18 dimensions, followed by three layers of dilated convolution to extract temporal features, ultimately outputting 6-dimensional centroid trajectory parameters. Parallel computing techniques are employed during the convolution process to improve the model's inference speed.

[0224] Processing delay:

[0225] The single-frame inference time is 25ms, meeting the requirements for real-time detection. By optimizing the model structure and algorithm, the computational complexity of the model is reduced, enabling fast inference.

[0226] (iv) Evaluation and Display Module

[0227] Based on the balance capability evaluation index in the scheme, the center of gravity trajectory parameters output by the receiving and processing module are used to calculate the center of gravity offset amplitude D and velocity fluctuation rate in real time. And compare with the preset threshold (good: D<5cm and <0.2m / s). Simultaneously, mathematical features are extracted from the model's output center of gravity estimates using manual or automatic methods (computer graphics algorithms) to generate spatiotemporal feature maps (time series plots, 3D polylines, 2D comparisons, etc.), allowing for a direct observation of the center of gravity's movement patterns and distribution. For example, if the center of gravity distribution is found to be relatively discrete in the spatial distribution heatmap, it indicates poor balance control during movement; in the coordinate distribution curve, if the coordinate fluctuation along a certain axis is large and irregular, there may be a balance obstacle in that direction. Based on the analysis results of the feature maps, the system further assists in determining the level of human balance ability, and visualizes the center of gravity trajectory curve, balance ability level, and related feature maps on terminal devices (mobile phones / PCs), facilitating users and professionals to intuitively understand their own or the subject's balance status.

[0228] Manual extraction:

[0229] Professionals analyze the trend of center of gravity changes at different stages of movement based on center of gravity coordinate data, mark abnormal fluctuation points, and combine clinical experience to judge the state of balance ability.

[0230] Automatic extraction:

[0231] Using computer graphics algorithms, the three-dimensional coordinate data of the center of gravity over a period of time are processed to create a spatial distribution heatmap and a coordinate distribution curve of the center of gravity. In the spatial distribution heatmap, the intensity of the color represents the frequency of the center of gravity appearing in a certain spatial region; the darker the color, the higher the probability of the center of gravity appearing in that region. The coordinate distribution curve shows the changes of the coordinates of the center of gravity on the X, Y, and Z axes over time.

[0232] The following is an explanation of model training and evaluation.

[0233] 1. Dataset Construction

[0234] Subject distribution:

[0235] 200 participants (aged 18-75, male-to-female ratio 1:1; general population / athletes / physical condition difference group = 6:2:2. The physical condition difference group includes fitness enthusiasts, sedentary individuals, and retired active individuals to ensure data diversity and improve model generalization ability).

[0236] Action Type:

[0237] Eight categories (standing upright, standing on one leg, walking, running, lateral jumping, head turning, trunk leaning forward, and disturbance actions) cover static maintenance, dynamic movement, and sudden disturbance scenarios, allowing the model to adapt to different sports / daily activities.

[0238] Data augmentation:

[0239] Gaussian noise (σ=0.05g) and random time offset (±50ms) were added, and the sample size was expanded to 100,000+. Sensor noise and action time sequence changes during motion were simulated to enhance the robustness of the model.

[0240] 2. Training Configuration

[0241] Optimizer:

[0242] Adam (lr=0.001, decay rate 0.9), combined with the advantages of Adagrad and RMSProp, accelerates convergence.

[0243] Training parameters:

[0244] Batch size 32, training epochs 50. Matching dataset size with computational resources allows the model to fully learn features.

[0245] Early shutdown mechanism:

[0246] If the validation set error does not decrease for five consecutive rounds, training is stopped to avoid overfitting and ensure generalization ability.

[0247] 3. Evaluation Results

[0248] Core error comparison:

[0249] Compared to the basic TCN, the optimized model reduced the overall coordinate error by 25%; the error in dynamic movements (such as walking and running) of the middle-aged and elderly group (≥51 years old) was reduced by 40%, validating the effectiveness of the improved accuracy.

[0250] Statistical significance:

[0251] Independent samples t-test showed that the error difference between the groups (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 and is suitable for diverse scenarios.

[0252] The device with a specific implementation method is used to implement a method for estimating the trajectory of the human body's center of gravity and detecting balance ability. Multi-dimensional error analysis is performed on the test data, and the results are as follows.

[0253] The technical solution of this invention not only performs excellently in terms of average performance, but also demonstrates strong generalization ability and high precision in a wide variety of people and action types.

[0254] Adaptability to different population groups: As shown in the table below, the optimized model of this invention achieved a significant reduction in error across all population groups compared to the basic TCN model. Particularly for groups with differences in physical condition (e.g., sedentary individuals, fitness enthusiasts) and the middle-aged and elderly groups, the error decreased significantly by 40% and 39%, respectively, fully demonstrating the effectiveness of the individualized feature embedding mechanism. The error differences between all groups were statistically significant (p<0.01) according to the independent samples t-test, indicating that the model can effectively distinguish and adapt to the balance characteristics of different population groups.

[0255] Table 5. Grouping Errors by Population (Center of Gravity Trajectory RMSE, Unit: cm)

[0256]

[0257] Adaptability to different movements: As shown in the table below, the model performed excellently in eight different types of movements. In particular, the error reduction was the largest in the two types of movements, "head rotation" and "torso forward tilt," which have extremely high requirements for dual-sensor collaboration and dynamic feature capture, reaching 33.3% and 34.5% respectively. This directly verifies the superiority of the dual IMU layout and differentiated expansion TCN architecture of this invention.

[0258] Table 6 Error by Action Type (RMSE of Center of Gravity Trajectory, Unit: cm)

[0259]

[0260] Note: The above data is a comparison of results calculated using the same set of test reference samples.

[0261] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the 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: Simultaneously acquire multidimensional motion data output from the first inertial measurement unit worn on the human head and the second inertial measurement unit worn on the human waist; And to obtain multidimensional individual characteristic data of the human body; S2, Feature Fusion: The multidimensional individual feature data is input into the embedding layer to generate an individual feature vector; and the individual feature vector is concatenated and fused with the multidimensional motion data to generate fused feature data; S3, Model Inference: Input the fused feature data into the end-to-end neural network model, perform calculations, and output the three-dimensional center of gravity trajectory parameters of the human body; The neural network model includes at least one dilated convolutional layer, which comprises a multi-layer differential dilated convolutional structure employing a non-equidistant dilation factor to expand the temporal receptive field and capture long-range dependencies.

2. The method for estimating the trajectory of the human body's center of gravity and detecting balance ability according to claim 1, characterized in that, The specific steps of S2 include: The multidimensional individual feature data is input into an adaptive attention fully connected embedding layer, and the weights are dynamically adjusted according to different individual features to generate the individual feature vector. The individual feature vector is concatenated with the multidimensional motion data along the feature dimension; The fused feature data is generated by reducing the dimensionality of the concatenated data through 1x1 convolution.

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

4. The method for estimating the trajectory of the human body's center of gravity and detecting balance ability according to claim 3, characterized in that, The activation function in the temporal convolutional network model is an adaptive activation function, which dynamically switches among various preset activation functions during model training based on the feature complexity of the current time step.

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

6. The method for estimating the trajectory of the human body's center of gravity and detecting balance ability according to claim 1, characterized in that, The method also includes, 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 assess the human body's balance ability according to the index. The index includes the center of gravity offset amplitude and velocity fluctuation rate.

7. A device for estimating the trajectory of a person's center of gravity and detecting their balance ability, characterized in that, include: The sensor module includes a first inertial measurement unit positioned at the head of the human body and a second inertial measurement unit positioned at the waist of the human body. The data acquisition module, connected to the sensor module, is used to synchronously acquire multidimensional motion data output by the first inertial measurement unit and the second inertial measurement unit, and to obtain multidimensional individual characteristic data of the human body. The processing module, connected to the data acquisition module, is configured for: The multidimensional individual feature data is processed through an embedding layer to generate individual feature vectors; The individual feature vectors are concatenated and fused with the multidimensional motion data to generate fused feature data; The fused feature data is input into an end-to-end neural network model, which calculates and outputs the three-dimensional center of gravity trajectory parameters of the human body.

8. The human body center of gravity trajectory estimation and balance ability detection device 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. This embedding layer is used to dynamically adjust the weights according to different individual features and reduce the dimensionality of the concatenated data through 1x1 convolution to generate the fused feature data. The end-to-end neural network model fixed in the processing module is a temporal convolutional network model. This model includes a multi-layer differential dilation convolutional structure with non-equidistant dilation factors and an adaptive activation function that can be dynamically selected during training.

9. The human body center of gravity trajectory estimation and balance ability detection device according to claim 7, characterized in that, The device further includes: The evaluation and display module, connected to the processing module, is used to calculate the center of gravity offset amplitude and / or velocity fluctuation rate based on the output three-dimensional center of gravity trajectory parameters, evaluate the human body's balance ability, and visualize the center of gravity trajectory curve, balance ability level, or spatial distribution feature map of the center of gravity.

Images