A motion environment perception method based on biomechanical characteristics and gait adaptation

Abstract

The application provides a motion environment perception method based on biomechanical characteristics and gait adaptation, and belongs to the field of human-computer interaction perception, and comprises the following steps: acquiring original motion characteristic sequences and biomechanical characteristic vectors of a subject, performing physical dimension scaling alignment through a biomechanical scaling matrix to obtain a motion characteristic tensor; inputting the motion characteristic tensor and the biomechanical characteristic vector into a multi-task time sequence convolution network to perform characteristic-level linear modulation, obtaining modulated deep-level characteristics, processing the modulated deep-level characteristics through a backbone network, outputting a terrain category by a classification branch, and outputting initial values of environmental geometric parameters by a regression branch; freezing the parameters of the backbone network, fine-tuning the regression branch based on an adaptive loss function driven by gait phase and arch characteristics, obtaining a motion environment perception model, inputting real-time data of the subject, and outputting the terrain category and the environmental geometric parameters. The application solves the technical problem of low accuracy of environmental geometric parameter estimation caused by biomechanical differences between subjects in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of human-computer interaction perception, and in particular relates to a motion environment perception method based on biomechanical characteristics and gait adaptation. Background Technology

[0002] With the rapid development of multimodal sensor integration technology and high-performance embedded computing platforms, the requirements for the perception accuracy of human motion intentions and environmental features in human-machine coupling systems are increasing. By acquiring high-frequency temporal motion feature signals such as joint angular displacement and dynamic interaction forces, human-machine coupling systems can estimate complex environmental geometric parameters, such as slope and step height, in real time, thus providing a physical basis for downstream control decisions. When processing such high-dimensional and dynamically changing motion feature data, how to maintain the stability of parameter estimation across sample applications and ensure the real-time performance and accuracy of perception during drastic environmental changes has become a primary problem that must be solved in the field of high-dynamic human-machine interaction perception.

[0003] Traditional methods for estimating motion environment parameters often employ fixed thresholding or intra-subject modeling techniques. While these methods perform well when the target and training samples are highly consistent, they are highly susceptible to individual biomechanical differences, such as variations in limb length and muscle strength, when applied across different samples. When faced with nonlinear noise generated by high-frequency motion impacts, traditional methods often require a compromise between smoothness and real-time performance, leading to significant amplitude deviations or phase lags in the perception curve, rendering them unreliable in real-world high-dynamic perception tasks. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a motion environment perception method based on biomechanical characteristics and gait adaptation. This method solves the technical problems of low accuracy in estimating environmental geometric parameters, phase lag, and poor cross-sample generalization ability caused by biomechanical differences between individual subjects and foot impact noise in the gait cycle in existing motion environment perception technologies.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a motion environment perception method based on biomechanical characteristics and gait adaptation, comprising the following steps: The original motion feature sequence and biomechanical feature vector of the subject are obtained, and the original motion features are scaled and aligned by physical dimensions using a biomechanical scaling matrix to obtain the motion feature tensor. The motion feature tensor and biomechanical feature vector are input into a multi-task temporal convolutional network. The motion feature tensor is linearly modulated at the feature level by the biomechanical feature vector to obtain the modulated deep features. Based on the deep modulation features, the data are processed through a backbone network. The classification branch outputs the terrain category, and the regression branch outputs the initial values ​​of the environmental geometric parameters. The backbone network parameters are frozen, only the regression branch is retained, and fine-tuning is performed based on an adaptive loss function jointly driven by the subject's gait phase and foot arch features to obtain the motion environment perception model. Based on the motion environment perception model, the real-time raw motion feature sequence and biomechanical feature vector of the subject are input. The real-time terrain category is output through the classification branch, and the real-time environmental geometric parameters are output through the regression branch. The filtering window is dynamically adjusted based on the real-time terrain category to smooth the real-time environmental geometric parameters, and the final motion environment perception result is obtained.

[0006] Existing motion environment perception methods mostly employ fixed thresholding or intra-subject modeling techniques. They collect temporal motion features such as joint angles, angular velocities, and interaction torques to regress and estimate environmental geometric parameters like slope and step height. However, when applied across different samples, biomechanical differences in limb length, weight, and arch shape among subjects can cause amplitude drift in the original signal. Traditional methods struggle to eliminate these physical dimension differences, leading to decreased accuracy in perceiving environmental parameters for new individuals. Furthermore, the rigid impact of the heel striking the ground during walking introduces high-frequency nonlinear noise, causing amplitude deviations or phase lag in the perception curve, making it difficult to meet the real-time requirements of highly dynamic human-computer interaction scenarios. This invention uses a biomechanical scaling matrix to physically scale and align the original motion features, reducing amplitude drift caused by differences in body shape among individual subjects. The biomechanical feature vector is input into a multi-task temporal convolutional network, allowing the output to match the subject's biological characteristics, improving perception accuracy and suppressing noise. Finally, it achieves high-precision environmental perception without phase lag during terrain step transitions, enhancing the adaptability and robustness of the human-computer coupling system in highly dynamic interaction scenarios.

[0007] Further: the obtained motion feature tensor specifically includes: Multimodal data is collected by sensors worn by the subjects to form raw motion feature sequences, and the subjects' static biological constants are collected simultaneously to construct biomechanical feature vectors; Based on static biological constants and combined with a dynamic model based on the movement characteristics of the human lower limbs, the biomechanical scaling matrix is ​​derived. Based on the biomechanical scaling matrix, the original motion feature sequence is scaled and aligned according to physical dimensions, and then time-series sliced ​​according to a set time window to obtain the motion feature tensor.

[0008] Furthermore, the expression for the motion feature tensor is as follows:

[0009]

[0010]

[0011]

[0012] in, For the motion feature tensor, For biomechanical scaling matrix, This is the original motion feature sequence. The mean of the training set of the baseline model. The standard deviation of the training set of the benchmark model. Let be the error constant. For diagonal matrix operators, Composed of all 1 elements The dimension row vector corresponds to the scaling factor of the four joint features. This is a scaling vector of the angular velocity characteristics. This is a scaling vector representing the torque characteristics. Composed of all 1 elements Dimensional row vectors correspond to geometric vector features and do not require scaling. Angular velocity scaling factor This is the torque scaling factor. The leg length of the baseline subject, The current subject's leg length, The baseline subject's weight, This refers to the current weight of the subject.

[0013] The further beneficial effects mentioned above are as follows: Existing technologies are difficult to eliminate the perceptual result deviation caused by individual differences among different subjects. This invention introduces a biomechanical scaling matrix derived from the inverted pendulum dynamics model to differentially scale features of different physical dimensions such as angle, angular velocity, and torque, so that the motion feature tensor is aligned at the physical level, resulting in motion equivalent to that of the benchmark subject, thereby reducing the impact of differences in body size among different individuals on the perceptual accuracy.

[0014] Further: the obtained deep modulation features specifically include: The motion feature tensor and biomechanical feature vector are input into a multi-task temporal convolutional network; The adaptive expansion factor is calculated based on the leg length parameter in the biomechanical feature vector; Based on the adaptive expansion factor, the biomechanical receptive field of the multi-task temporal convolutional network is adaptively adjusted to obtain the multi-task temporal convolutional network with the optimal time window. Based on the multi-task temporal convolutional network with the optimal time window, temporal convolutional feature extraction is performed on the motion feature tensor to obtain the feature matrix. Then, feature-level linear modulation is performed using biomechanical feature vectors to obtain the modulated deep features.

[0015] Furthermore, the expression for the modulation deep feature is as follows:

[0016]

[0017] in, To modulate deep features, This is the scaling factor. The bias coefficient, This is a biomechanical feature vector. For element-wise multiplication, For the first task in a multi-task temporal convolutional network The feature matrix output by each convolutional layer The current subject's leg length, The arch index of the current subjects, This represents the gait habit coefficient of the current test subject; The expression for the biomechanical receptive field is as follows:

[0018]

[0019] in, For biomechanical receptive fields The kernel size is [size]. As an adaptive expansion factor, The leg length of the baseline subject, For the floor operator, This is the index for the convolutional layer.

[0020] The further beneficial effects mentioned above are as follows: Most temporal convolutional networks in existing environmental perception technologies use fixed receptive fields, which cannot match the gait cycle differences of subjects with different leg lengths, resulting in phase misalignment or information loss during feature extraction; This invention calculates an adaptive expansion factor through leg length parameters, which can adjust the receptive field to cover the single-step gait cycle of different individuals, and combines biomechanical feature vectors to linearly modulate the feature matrix, thus adapting to individual biomechanical characteristics at the physical level, reducing the phase misalignment problem of perception results caused by different subjects, and improving the accuracy of feature expression.

[0021] Furthermore, the expression for the adaptive loss function jointly driven by the subject's gait phase and foot arch features is as follows:

[0022]

[0023] in, For adaptive loss function, For the current time step, This represents the total number of time steps within the time window. For gait phase adaptive weighting function, For real-time gait phase, The arch index of the current subjects, For smoothing loss function For time step Real terrain geometry feature labels These are the terrain feature values ​​predicted by the model. For frozen backbone network parameters, The impact penalty coefficient modulated by the arch index. It is a natural exponential function. The impact phase of the heel strike. The standard deviation parameter is used to control the width of the impact penalty phase window.

[0024] The further beneficial effects mentioned above are: the present invention utilizes a gait phase adaptive weighting function. In the impact phase The system automatically reduces loss weights, suppresses the impact of ground contact impulse noise on parameter updates, and incorporates the arch index. Modulation impact penalty coefficient By applying stronger deweighting to flat-footed subjects with poor buffering capacity, the robustness of cross-individual fine-tuning and the accuracy of environmental parameter estimation are effectively improved.

[0025] Furthermore, the expression for the final motion environment perception result is as follows:

[0026] in, For the final motion environment perception result, The normalization coefficient is... The dynamic window length is controlled by terrain semantics. For the current time step, This is the relative time offset index within the window. For the pre-calculated convolution coefficients, For the model in The initial environmental perception prediction value output at each moment.

[0027] The beneficial effects of this invention are: This invention uses a biomechanical scaling matrix built based on an inverted pendulum dynamics model to scale and align the original motion features by physical dimensions, reducing signal amplitude drift caused by differences in body shape such as leg length and weight among different subjects. At the same time, it uses the leg length parameter to adaptively adjust the expansion factor of the temporal convolutional network, so that the receptive field accurately matches the individual gait cycle. Furthermore, it combines biomechanical feature vectors to linearly modulate the convolutional features, solving the phase misalignment problem when applied across samples, and improving the model's generalization ability and the accuracy of environmental parameter estimation among different subjects.

[0028] In the personalized fine-tuning stage, this invention adopts an adaptive loss function jointly driven by gait phase and arch index. It automatically reduces the loss weight at the heel strike impact phase, suppresses the interference of ground impact impulse noise on parameter updates, and applies a stronger penalty coefficient for special arch shapes such as flat feet. This avoids the model misjudging impact noise as terrain changes and improves the robustness of environmental perception during high-dynamic walking.

[0029] This invention drives two tasks, terrain classification and environmental parameter regression, simultaneously through a shared backbone network. When faced with a new individual, the backbone network is frozen, and only the regression branch is fine-tuned, reducing the computational overhead and memory usage of the embedded platform. At the same time, the filter window width is dynamically adjusted according to the semantic category probability of the terrain, and automatically adjusted at the terrain step edge, outputting lag-free and high-precision environmental geometric parameters. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a motion environment perception method based on biomechanical characteristics and gait adaptation. Detailed Implementation

[0031] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0032] Example 1 like Figure 1 The diagram shows a flowchart of a motion environment perception method based on biomechanical features and gait adaptation, which includes the following steps: The original motion feature sequence and biomechanical feature vector of the subject are obtained, and the original motion features are scaled and aligned by physical dimensions using a biomechanical scaling matrix to obtain the motion feature tensor. The motion feature tensor and biomechanical feature vector are input into a multi-task temporal convolutional network. The motion feature tensor is linearly modulated at the feature level by the biomechanical feature vector to obtain the modulated deep features. Based on the deep modulation features, the data are processed through a backbone network. The classification branch outputs the terrain category, and the regression branch outputs the initial values ​​of the environmental geometric parameters. The backbone network parameters are frozen, only the regression branch is retained, and fine-tuning is performed based on an adaptive loss function jointly driven by the subject's gait phase and foot arch features to obtain the motion environment perception model. Based on the motion environment perception model, the real-time raw motion feature sequence and biomechanical feature vector of the subject are input. The real-time terrain category is output through the classification branch, and the real-time environmental geometric parameters are output through the regression branch. The filtering window is dynamically adjusted based on the real-time terrain category to smooth the real-time environmental geometric parameters, and the final motion environment perception result is obtained.

[0033] In one embodiment of the present invention, the subject's original motion feature sequence is: ,in, For the current time step, For angle, Angular velocity, For torque, Traditional environmental sensing methods for ground-sensing pulses often suffer from the technical problem of failing to eliminate differences in subject body size. Current common approaches involve linear normalization or scaling, which involves a simple linear transformation of sensor signals based on global parameters such as the subject's height and weight. This ignores the independent dynamic dimensions of different physical quantities such as angle, angular velocity, and torque. For example, angular velocity is inversely proportional to leg length, and torque is directly proportional to weight and leg length. This results in physical biases in the processed data, failing to truly eliminate amplitude drift caused by differences in leg length and weight. This invention introduces a biomechanical scaling matrix derived from an inverted pendulum dynamics model to differentially scale feature channels with different physical dimensions such as angle, angular velocity, and torque. This allows the motion feature tensor to be physically equivalent to the motion characteristics of a benchmark subject, reducing amplitude drift caused by differences in the body size of different subjects. The motion feature tensor obtained in this invention specifically includes: Multimodal data is collected using sensors worn by the subjects to form raw motion feature sequences. Simultaneously, the static biological constants of the subjects were collected to construct biomechanical feature vectors; among them, multimodal data included the subjects' joint angles, angular velocities, and interaction torques, while static biological constants included leg length, weight, arch index, and gait. Based on static biological constants and a dynamic model based on the movement characteristics of the human lower limbs, the biomechanical scaling matrix is ​​derived. ; Based on the biomechanical scaling matrix, the original motion feature sequence is scaled and aligned according to physical dimensions, and then time-series sliced ​​according to a set time window to obtain the motion feature tensor. The expression of the motion feature tensor is as follows:

[0034]

[0035]

[0036]

[0037] in, For the motion feature tensor, For biomechanical scaling matrix, It is a diagonal matrix. This is the original motion feature sequence. The mean of the training set of the baseline model. The standard deviation of the training set of the benchmark model. Let be the error constant. For diagonal matrix operators, Composed of all 1 elements The dimension row vector corresponds to the scaling factor of the four joint features. This is a scaling vector of the angular velocity characteristics. This is a scaling vector representing the torque characteristics. Composed of all 1 elements Dimensional row vectors correspond to geometric vector features and do not require scaling. Angular velocity scaling factor This is the torque scaling factor. The leg length of the baseline subject, The current subject's leg length, The baseline subject's weight, The current subject's weight is used; in the angular velocity scaling factor, the angular velocity is inversely proportional to the square root of the leg length; in the torque scaling factor, the joint torque is directly proportional to the product of the weight and the leg length; and the angle and impulse characteristics are treated as dimensionless or geometric absolute quantities, with a scaling factor of 1.

[0038] In one embodiment of the present invention, traditional human-computer coupling systems often use temporal convolutional networks to process the motion features of subjects. The present invention not only scales the input motion features using a biomechanical scaling matrix to reduce errors caused by differences in body shape among different subjects, but also considers that different subjects have different gait cycles, requiring different receptive fields for the temporal convolutional network. Conventional temporal convolutional networks use a fixed receptive field. For subjects with shorter leg lengths, the fixed receptive field may cover more than one gait cycle, introducing redundant noise between adjacent cycles; for subjects with longer leg lengths, the receptive field may be insufficient to encompass the complete single-step gait information, leading to phase misalignment during temporal feature extraction. The present invention designs an adaptive expansion factor that can dynamically adjust the receptive field according to the subject's arch shape, gait habits, and other biomechanical characteristics, improving the generalization ability across samples. The present invention inputs the motion feature tensor and biomechanical feature vector into a multi-task temporal convolutional network, and performs feature-level linear modulation on the motion feature tensor through the biomechanical feature vector to obtain modulated deep features, specifically including: The motion feature tensor and biomechanical feature vector are input into a multi-task temporal convolutional network, where the expression for the biomechanical features is as follows:

[0039] in, Biomechanical characteristics The current subject's leg length, The arch index of the current subjects, This represents the gait habit coefficient of the current test subjects. It is the transpose symbol; Based on the leg length parameter in the biomechanical feature vector, the adaptive expansion factor is calculated, and its expression is as follows:

[0040] in, As an adaptive expansion factor, The leg length of the baseline subject, For the floor operator, This is the index for the convolutional layer; the formula is based on the physical law in human gait kinematics that the natural period of an inverted pendulum is proportional to the square root of the leg length, so that the adaptive expansion factor scales linearly with the leg length. Based on the adaptive dilation factor, the biomechanical receptive field of the multi-task temporal convolutional network is adaptively adjusted to obtain the multi-task temporal convolutional network with the optimal time window. The expression for the adaptively adjusted receptive field is as follows:

[0041] in, For biomechanical receptive fields The kernel size is determined by the adaptive expansion factor. When the subject's leg length is relatively long, the adaptive expansion factor automatically increases, and the receptive field automatically widens to cover the entire gait cycle. When the leg length is relatively short, the adaptive expansion factor automatically decreases, and the receptive field automatically shrinks to avoid interference between adjacent cycles, thereby achieving different individual matching within the time window at the physical level. Based on a multi-task temporal convolutional network with an optimal time window, temporal convolutional feature extraction is performed on the motion feature tensor to obtain a feature matrix. Then, feature-level linear modulation is performed using biomechanical feature vectors to obtain deep modulation features. Specifically, this invention inputs the biomechanical feature vectors into a multilayer perceptron to generate affine transformation coefficients. and For multi-task temporal convolutional networks, the first Feature matrix output by convolution layer Performing linear modulation yields the following expression for the deep modulation features:

[0042] in, To modulate deep features, This is the scaling factor. The bias coefficient, For element-wise multiplication, For the first task in a multi-task temporal convolutional network The feature matrix output by the convolutional layer; the modulation operation of this invention enables the multi-task temporal convolutional network to adaptively enhance or suppress the expression of high-frequency features, such as the impact frequency component of the plantar surface of flat feet, based on the arch index of the subject, and adjust the bias of joint angle features according to the gait habit coefficient; to further embed physical priors into the bottom layer of the convolutional operator, so that the deep semantics of the network can perceive individual body shape differences, this invention also defines the first Layer convolution at time The modulation response is:

[0043] in, For the first Layer convolution in Modulation response at time, For a moment, For the size of the core, The expansion factor is used to force the pure mathematical convolution features to be stretched or compressed into a physical space that conforms to the current subject dynamics dimensions from the bottom layer of the operator.

[0044] This invention obtains deep modulation features through adaptive expansion factor adjustment and biomechanical feature-level linear modulation. These features are highly aligned with the individual gait characteristics of the subjects at the physical level, providing accurate dynamic representations for subsequent multi-task classification and regression, and improving the cross-individual generalization ability of environmental parameter estimation.

[0045] In one embodiment of the present invention, based on the modulated deep features, the data are processed through a backbone network. The classification branch outputs the terrain category, and the regression branch outputs the initial values ​​of the environmental geometric parameters. The backbone network parameters are frozen, only the regression branch is retained, and fine-tuning is performed based on an adaptive loss function jointly driven by the subject's gait phase and foot arch features to obtain a motion environment perception model.

[0046] Traditional multi-task environmental perception methods in human-machine coupled systems typically train terrain classification and parametric regression as two independent models. This not only incurs high computational costs but also fails to utilize shared feature representations between the two. During the cross-sample adaptation phase, traditional methods often employ global fine-tuning or static weight transfer. However, differences in gait habits and foot arch morphology among subjects can lead to amplitude biases in the regression curves. Particularly at the moment of heel strike, the impulse noise generated by rigid impact can be misinterpreted by the model as a step change in terrain, resulting in incorrect parameter update directions. Existing technologies lack fine-grained modeling of the differences in phase noise within the gait cycle, causing estimation jitter to persist even in highly dynamic walking scenarios after personalized calibration.

[0047] This invention employs a shared backbone network to simultaneously drive two tasks: terrain classification and parameter regression. The classification branch outputs the terrain category, while the regression branch outputs the initial values ​​of environmental geometric parameters. When performing personalized fine-tuning for different subjects, the backbone network parameters are frozen, and only the regression branch is updated. An adaptive loss function jointly driven by gait phase and arch index is introduced. The specific implementation is as follows: After the modulated deep features are input into the backbone network, they are sent to two branches. The classification branch outputs the probability of the terrain category at the current time, such as flat land, up stairs, down slope, etc.; the regression branch outputs the initial estimate of environmental geometric parameters, such as slope angle or step height. The two branches share the backbone network parameters, which reduces redundant computation. When dealing with new subjects, initiate a personalized fine-tuning process; freeze all parameters of the backbone network. Only the regression branch is retained for training. Combined with target terrain samples, terrain typically within 10-20 gait cycles can be selected. A gait phase adaptive weighted loss function is used for fine-tuning. The expression of the adaptive loss function jointly driven by the subject's gait phase and foot arch features is as follows:

[0048] The gait phase adaptive weighting function is defined as follows:

[0049] in, For adaptive loss function, For the current time step, This represents the total number of time steps within the time window. For gait phase adaptive weighting function, Real-time gait phase can be estimated using plantar pressure or an inertial measurement unit. The arch index of the current subjects, For smoothing loss function For time step Real terrain geometry feature labels These are the terrain feature values ​​predicted by the model. For frozen backbone network parameters, The impact penalty coefficient modulated by the arch index. It is a natural exponential function. The impact phase of the heel strike. The standard deviation parameter is used to control the width of the impact penalty phase window.

[0050] With the adaptive loss function of this invention, when the gait phase approaches the impact phase, the exponential term approaches 1, and the gait phase adaptive weight... To minimize the loss contribution at that moment, avoid impulse noise dominating gradient updates, and rely on historical smooth features for inference; The penalty coefficient modulated by the arch index is used for flat-footed subjects. Larger impacts result in poor foot cushioning, leading to longer duration and higher amplitude of impact noise. Automatically increase and strengthen the weighting of the impact phase to prevent the model from misinterpreting noise as a terrain step; for normal foot arches, Taking a smaller value only moderately weakens the impact phase.

[0051] The parameter update rules during the personalized fine-tuning process are as follows:

[0052] in, This is the weight matrix for the regression branch. For learning rate, For gradient, loss function Weights The gradient indicates the direction of parameter updates, i.e. the direction of the fastest gradient descent. Through gait phase adaptive weighted loss, the model can learn to rely on historical smooth features for inference when an impact occurs, rather than forcibly fitting noise spikes. This maintains the sensitivity of terrain step response while suppressing periodic estimation jitter caused by individual gait habit differences, such as pigeon toes and landing force preferences.

[0053] In one embodiment of the present invention, based on the motion environment perception model, the real-time raw motion feature sequence and biomechanical feature vector of the subject are input, the real-time terrain category is output through the classification branch, the real-time environmental geometric parameters are output through the regression branch, and the real-time environmental geometric parameters are smoothed by dynamically adjusting the filtering window based on the real-time terrain category to obtain the final motion environment perception result.

[0054] Traditional environmental perception methods typically employ fixed-window low-pass filtering or moving averages after regression output to eliminate high-frequency oscillation noise in the gait cycle. However, an excessively wide fixed window can lead to significant phase lag during terrain step transitions, causing exoskeletons or robots to miss optimal assist opportunities. Conversely, an excessively narrow window cannot effectively suppress impact spikes within the gait cycle, resulting in control command jitter. Furthermore, traditional filters rely solely on signal amplitude changes, lacking prior knowledge of the current terrain semantics. This makes it difficult to employ different smoothing strategies for different terrains such as slopes and stairs, resulting in filtering accuracy and speed failing to meet practical requirements. To address this, this invention uses the terrain category probability output from the classification branch as the basis for adjusting the filtering parameters, yielding the following expression for the final motion environment perception result:

[0055] in, For the final motion environment perception result, The normalization coefficient is... The dynamic window length is controlled by terrain semantics. For the current time step, This is the relative time offset index within the window. For the pre-calculated convolution coefficients, For the model in The initial environmental perception prediction value output at each moment; The dynamic window length is controlled by terrain semantics. When the system predicts that it is currently at the edge of a terrain step transition, Rapid contraction to eliminate phase lag caused by the filtering algorithm, when in steady-state terrain. The expansion is used to suppress oscillations in the motion cycle, ultimately outputting hysteresis-free, high-precision environmental geometry parameters.

[0056] This invention uses a biomechanical scaling matrix built based on an inverted pendulum dynamics model to scale and align the original motion features by physical dimensions, reducing signal amplitude drift caused by differences in body shape such as leg length and weight among different subjects. At the same time, it uses the leg length parameter to adaptively adjust the expansion factor of the temporal convolutional network, so that the receptive field accurately matches the individual gait cycle. Furthermore, it combines biomechanical feature vectors to linearly modulate the convolutional features, solving the phase misalignment problem when applied across samples, and improving the model's generalization ability and the accuracy of environmental parameter estimation among different subjects.

[0057] In the personalized fine-tuning stage, this invention adopts an adaptive loss function jointly driven by gait phase and arch index. It automatically reduces the loss weight at the heel strike impact phase, suppresses the interference of ground impact impulse noise on parameter updates, and applies a stronger penalty coefficient for special arch shapes such as flat feet. This avoids the model misjudging impact noise as terrain changes and improves the robustness of environmental perception during high-dynamic walking.

[0058] This invention drives two tasks, terrain classification and environmental parameter regression, simultaneously through a shared backbone network. When faced with a new individual, the backbone network is frozen, and only the regression branch is fine-tuned, reducing the computational overhead and memory usage of the embedded platform. At the same time, the filter window width is dynamically adjusted according to the semantic category probability of the terrain, and automatically adjusted at the terrain step edge, outputting lag-free and high-precision environmental geometric parameters.

Claims

1. A motion environment perception method based on biomechanical characteristics and gait adaptation, characterized in that, Includes the following steps: The original motion feature sequence and biomechanical feature vector of the subject are obtained, and the original motion features are scaled and aligned by physical dimensions using a biomechanical scaling matrix to obtain the motion feature tensor. The motion feature tensor and biomechanical feature vector are input into a multi-task temporal convolutional network. The motion feature tensor is linearly modulated at the feature level by the biomechanical feature vector to obtain the modulated deep features. Based on the deep modulation features, the data are processed through a backbone network. The classification branch outputs the terrain category, and the regression branch outputs the initial values ​​of the environmental geometric parameters. The backbone network parameters are frozen, only the regression branch is retained, and fine-tuning is performed based on an adaptive loss function jointly driven by the subject's gait phase and foot arch features to obtain the motion environment perception model. Based on the motion environment perception model, the real-time raw motion feature sequence and biomechanical feature vector of the subject are input. The real-time terrain category is output through the classification branch, and the real-time environmental geometric parameters are output through the regression branch. The filtering window is dynamically adjusted based on the real-time terrain category to smooth the real-time environmental geometric parameters and obtain the final motion environment perception result. The obtained motion feature tensor specifically includes: Multimodal data is collected by sensors worn by the subjects to form raw motion feature sequences, and the subjects' static biological constants are collected simultaneously to construct biomechanical feature vectors; Based on static biological constants and combined with a dynamic model based on the movement characteristics of the human lower limbs, the biomechanical scaling matrix is ​​derived. Based on the biomechanical scaling matrix, the original motion feature sequence is scaled and aligned according to physical dimensions, and then time-series sliced ​​according to a set time window to obtain the motion feature tensor. The expression for the motion feature tensor is as follows: in, For the motion feature tensor, For the current time step, For biomechanical scaling matrix, This is the original motion feature sequence. The mean of the training set of the baseline model. The standard deviation of the training set of the benchmark model. Let be the error constant. For diagonal matrix operators, Composed of all 1 elements The dimension row vector corresponds to the scaling factor of the four joint features. This is a scaling vector of the angular velocity characteristics. This is a scaling vector representing the torque characteristics. Composed of all 1 elements Dimensional row vector Angular velocity scaling factor This is the torque scaling factor. The leg length of the baseline subject, The current subject's leg length, The baseline subject's weight, This refers to the current weight of the subject.

2. The motion environment perception method based on biomechanical characteristics and gait adaptation according to claim 1, characterized in that, The obtained deep modulation features specifically include: The motion feature tensor and biomechanical feature vector are input into a multi-task temporal convolutional network; The adaptive expansion factor is calculated based on the leg length parameter in the biomechanical feature vector; Based on the adaptive expansion factor, the biomechanical receptive field of the multi-task temporal convolutional network is adaptively adjusted to obtain the multi-task temporal convolutional network with the optimal time window. Based on the multi-task temporal convolutional network with the optimal time window, temporal convolutional feature extraction is performed on the motion feature tensor to obtain the feature matrix. Then, feature-level linear modulation is performed using biomechanical feature vectors to obtain the modulated deep features.

3. The motion environment perception method based on biomechanical characteristics and gait adaptation according to claim 2, characterized in that, The expression for the modulation deep feature is as follows: in, To modulate deep features, This is the scaling factor. The bias coefficient, This is a biomechanical feature vector. For element-wise multiplication, For the first task in a multi-task temporal convolutional network The feature matrix output by each convolutional layer The current subject's leg length, The arch index of the current subjects, This represents the gait habit coefficient of the current test subject; The expression for the biomechanical receptive field is as follows: in, For biomechanical receptive fields The kernel size is [size]. As an adaptive expansion factor, The leg length of the baseline subject, For the floor operator, For convolutional layer index, This represents the total number of convolutional layers.

4. The motion environment perception method based on biomechanical characteristics and gait adaptation according to claim 1, characterized in that, The expression for the adaptive loss function jointly driven by the subject's gait phase and foot arch features is as follows: in, For adaptive loss function, For the current time step, This represents the total number of time steps within the time window. For gait phase adaptive weighting function, For real-time gait phase, The arch index of the current subjects, For smoothing loss function For time step Real terrain geometry feature labels These are the terrain feature values ​​predicted by the model. For frozen backbone network parameters, The impact penalty coefficient modulated by the arch index. It is a natural exponential function. The impact phase of the heel strike. The standard deviation parameter is used to control the width of the impact penalty phase window.

5. The motion environment perception method based on biomechanical characteristics and gait adaptation according to claim 1, characterized in that, The expression for the final motion environment perception result is as follows: in, For the final motion environment perception result, The normalization coefficient is... The dynamic window length is controlled by terrain semantics. For the current time step, This is the relative time offset index within the window. For the pre-calculated convolution coefficients, For the model in The initial environmental perception prediction value output at each moment.

Images