A human motion tracking method based on flexible sensor

Abstract

The present application relates to a human motion tracking method based on flexible sensors, comprising: simultaneously collecting data of four flexible sensors worn by a test person at four joints of the limbs of the human body and human motion data of the test person within a period of time, and generating data pairs corresponding one-to-one between the flexible sensor data and the human motion data based on the collection time, and constructing a training set based on all data pairs of all test persons; constructing a human motion tracking model based on a Bi-LSTM network, taking the flexible sensor data as the input of the model, and the output of the model being predicted human motion data, and training the model through the training set; and tracking human motion based on the trained human motion tracking model. The present application uses a bidirectional LSTM neural network model which fully utilizes short sequence information, thereby reducing the possibility of the model collapsing to an average posture.

Description

Technical Field

[0001] This invention relates to the field of human motion tracking, and more particularly to a human motion tracking method based on a flexible sensor. Background Technology

[0002] Human motion tracking is widely used in animation, video games, human-computer interaction, and medical rehabilitation applications. Currently, optical solutions and inertial measurement units (IMUs) are the most common methods for tracking human motion and have mature applications. Between the two, optical solutions are affected by harsh environmental conditions (such as occlusion or poor lighting) and are also limited by clothing used to place optical markers; inertial measurement units (IMUs) do not have these limitations and are a more robust choice. While this method can effectively track human motion, its reliance on densely packed IMUs can interfere with the user experience. Commercial systems even use more than 10 IMUs for motion tracking.

[0003] Using multiple flexible sensors for measurement is a time-consuming and labor-intensive task, due to the long wear and tear on the devices, high cost, and the inconvenience of users wearing multiple flexible sensors. Researchers have also attempted to track human motion using a small set of sensors, employing only six IMUs, but this method requires significant computational costs and offline optimization for non-convex problems across the entire sequence. Although researchers have achieved good tracking results with six IMUs, exploring how to track human motion with fewer sensors to improve device usability remains meaningful.

[0004] Current methods for predicting human motion still have some shortcomings: 1) widely adopted data-driven methods are prone to collapse to average posture; 2) predicted motion contains unnatural jitter. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a human motion tracking method based on flexible sensors.

[0006] The specific plan is as follows:

[0007] A human motion tracking method based on flexible sensors includes the following steps:

[0008] S1: Simultaneously collect data from four flexible sensors worn by the tester at four joints of the limbs and human motion data of the tester within a certain period of time, and generate data pairs that correspond one-to-one between flexible sensor data and human motion data based on the collection time. Construct a training set based on all data pairs collected from all testers.

[0009] S2: Construct a human motion tracking model based on a Bi-LSTM network, using flexible sensor data as the input to the model and the model output as predicted human motion data. Train the model using a training set.

[0010] S3: Track human motion based on the trained human motion tracking model.

[0011] Furthermore, human motion data includes data on arm joints, forearm joints, thigh joints, leg joints, and spinal joints.

[0012] Furthermore, the flexible sensors are configured to connect to the four joints of the corresponding limbs of the bodysuit worn on the human body.

[0013] Furthermore, the flexible sensor is connected to the bodysuit by placing the flexible sensor in a pocket at the joint position of the corresponding limb of the bodysuit.

[0014] Furthermore, before training the model, the process includes: dynamically calibrating the flexible sensor data in the training set, and then using the dynamically calibrated data to train the model; the specific formula for dynamic calibration is:

[0015] Y out =Y in -(LN(ave(min(Y in ))))

[0016] Among them, Y in Y represents normalized flexible sensor data for a time series. out This represents the calibrated flexible sensor data, where LN represents the fully connected layer, min represents the minimum pooling layer, and ave represents the average pooling layer.

[0017] Furthermore, the model's network structure embeds a median pooling layer at the end of the Bi-LSTM network.

[0018] Furthermore, the connection between the flexible sensor and the wire is pluggable.

[0019] The present invention adopts the above technical solution, which reduces the possibility of the model collapsing to the average pose by using a bidirectional LSTM neural network model that makes full use of short time series information; and reduces jitter by adding a median pooling layer to the network, making the prediction results smoother. Attached Figure Description

[0020] Figure 1 The diagram shows a flowchart of a method according to an embodiment of the present invention.

[0021] Figure 2 The diagram shows the arrangement of the flexible sensor in an embodiment of the present invention.

[0022] Figure 3 The diagram shown is a schematic of the flexible sensor used in an embodiment of the present invention.

[0023] Figure 4 The figure shown is a graph illustrating the change in resistance of the flexible sensor as a function of bending angle in an embodiment of the present invention.

[0024] Figure 5 The image shown is an external view of the PCB board in an embodiment of the present invention.

[0025] Figure 6 The diagram shown is a circuit diagram of the PCB board in an embodiment of the present invention.

[0026] Figure 7 The diagram shown is a schematic representation of the network structure of the model in an embodiment of the present invention.

[0027] Figure 8 The image shown is a comparison diagram of jitter distribution in an embodiment of the present invention.

[0028] Figure 9 The image shown is a comparison of the bending angle of the left leg with and without a median pooling layer in an embodiment of the present invention.

[0029] Figure 10 The image shows a visual evaluation comparison of methods for different types of motion in an embodiment of the present invention.

[0030] Figure 11 The figure shows a qualitative comparison between the method of this embodiment and other methods in this invention.

[0031] Figure 12 The figure shows a qualitative comparison between the method of this embodiment and the IMU in this embodiment of the invention. Detailed Implementation

[0032] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention.

[0033] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0034] Multi-sensor approaches aim to maximize the advantages of different sensor types while reducing the number of sensors placed on the human body, thereby minimizing disruption to the user experience. Most multi-sensor approaches utilize optical sensors for higher accuracy and a better user experience, and work in conjunction with an IMU to obtain precise limb orientation, which is challenging for purely optical systems in fast-moving or occluded scenarios. Although researchers have strived to reduce the number and complexity of hardware, the facility requirements (such as equipment maintenance costs and user familiarity with costs) for this integrated approach are higher than for single-sensor systems. Furthermore, this method, combined with an optical system, is susceptible to occlusion and lighting conditions. Therefore, this embodiment employs a single-sensor approach, which is easier to deploy in practical applications.

[0035] To extend the application of human motion tracking beyond laboratory environments, we have made significant efforts to reduce the number of sensors used, which not only lowers costs but also improves wearing comfort. Research has shown that researchers primarily use IMUs (Insulated Measurement Units) to track human motion because they are small, inexpensive, and easily configured outside of laboratory settings. However, IMUs still have some inherent problems: 1) While the number of sensors used is reduced, they are not flexible and therefore can still interfere with human motion; 2) They are susceptible to electromagnetic interference; 3) The data captured by the sensors can be very noisy: motion amplitude can occasionally be over-amplified due to errors in measuring acceleration. To address these limitations, this embodiment preferably uses flexible sensors instead of IMUs for human motion tracking, resulting in a better user experience.

[0036] In human motion tracking based on flexible sensors, existing solutions either employ dense, user-unfriendly flexible sensors or focus on tracking localized motion. Based on the above research results, this invention provides a human motion tracking method based on flexible sensors, such as... Figure 1 As shown, the method includes the following steps:

[0037] S1: Simultaneously collect data from four flexible sensors worn by the tester at four joints of the limbs and human motion data of the tester over a period of time, and generate data pairs that correspond one-to-one between the flexible sensor data and human motion data based on the collection time. Construct a training set based on all data pairs collected from all testers.

[0038] In this embodiment, the flexible sensor is connected to the four joints of the corresponding limbs of the bodysuit worn on the human body. The data acquisition system includes the bodysuit, the flexible sensor, and the PCB board. These will be described separately below.

[0039] (1) Bodysuit: Before designing the bodysuit, we conducted a preliminary study to determine the placement of the flexible sensors. The study found that, with sparse flexible sensors, placing them on the pelvis and limbs was the optimal choice. Given that we only placed four sensors, we used an optical system to collect human motion data and employed a bidirectional LSTM neural network to verify the optimal placement one by one. The results showed that placing the sensors on the limbs (Y-axis of the forearm and X-axis of the leg) yielded the best tracking results.

[0040] Therefore, in this embodiment, four flexible sensors are placed at four joint positions of the bodysuit (the X-axis of the two knees and the Y-axis of the two elbows), which can closely fit the bodies of different users. Figure 2 Specifically, flexible sensors can be placed in four pockets (such as long pockets) sewn into the bodysuit to reduce the probability of large sensor displacements. These flexible sensors are secured to the pockets with Velcro and rubber balls attached to wires.

[0041] Furthermore, in this embodiment, the connection between the flexible sensor and the wire is designed to be pluggable. The pluggable connection has the following advantages: 1) easy cleaning; 2) easy sensor replacement when damaged; 3) avoids unnecessary damage caused by excessive bending and collision when the flexible sensor is worn.

[0042] In this embodiment, a single tin-plated copper wire with an insulating layer is soldered to the flexible sensor at 350 degrees Celsius. The PCB board is placed in front of the user to facilitate interaction during data acquisition. The wire can be inserted into the circuit board, with its ends encased in plastic. Without sacrificing performance, most of the wire can be hidden within the seam of the integrated design or in a double-layered fabric bag connected to the seam for improved aesthetics.

[0043] (2) Flexible sensor

[0044] The flexible sensor used in this embodiment is a resistive type, such as... Figure 3 As shown. Although the resistance of the flexible sensor monotonically increases as the side containing the grid (inner side) extends outward (as shown). Figure 4 (As shown), but the relationship between resistance and bending angle is non-linear, making it difficult to analytically derive their relationship. Furthermore, the resistance-bending relationship differs during stretching and bending, making learning regression models challenging. Since only the inner side of the flexible sensor is designed to change significantly with bending, the inner side of the sensor must face the outer side of the body during installation.

[0045] (3) PCB board:

[0046] In this embodiment, a PCB board is used (its appearance and design are as follows). Figure 5 and Figure 6 (As shown) is used for flexible sensor data acquisition. Specifically, the PCB board is configured to save the collected flexible sensor data to the SD card every 0.05 seconds.

[0047] During data acquisition, this embodiment uses a multi-channel voltage divider to select channels sequentially, and the reference voltage is obtained through... Figure 6 The image shows a pair of uniform resistive units in the lower left corner. A Wheatstone bridge structure is used to calculate the difference (V) between the voltage of each sensor and the reference voltage. ref =V CC / 2), where V CC This represents the input voltage of the circuit. Ignoring the effect of the low-pass filter, the input voltage of the digital-to-analog converter can be defined as follows:

[0048]

[0049] Among them, R i Indicates a voltage divider resistor, Rsensor i Let represent the resistance of the i-th flexible sensor, and Gain represent the amplification factor of the amplifier unit. The amplified voltage measurement is processed by a low-pass filter with a bandwidth of 300Hz. Finally, the output signal is converted into digital form within the range [0, 4096].

[0050] In this embodiment, 10 testers were invited for data acquisition, including 4 men and 6 women, aged 19 to 41. Each tester was guided to wear the bodysuit provided by the acquisition system, the sensors and PCB board were installed in their respective positions, and the PCB board was switched on. Data was acquired through the PCB board from four flexible sensors, while human motion data was simultaneously acquired through an optical measurement system (Qualisys motion capture system).

[0051] For human motion data, this embodiment only collects data on the arm joints, forearm joints, thigh joints, leg joints, and spinal joints (including all three spinal joints). It is not necessary to collect motion data on the rotation of the head, thumb joints, finger joints, foot joints, hand joints, neck joints, and hip joints, because the movements of these joints are independent of the movements of other joints. For example, a person can independently turn their head left or right while walking.

[0052] Furthermore, to compare the solution used in this embodiment with traditional solutions using IMUs, we further required two subjects to wear clothing with IMUs on our bodysuits. Data acquisition from the IMUs was synchronized with flexible sensors and an optical measurement system.

[0053] S2: Construct a human motion tracking model based on a Bi-LSTM network, using flexible sensor data as the input to the model and predicting human motion data as the output. Train the model using a training set.

[0054] Human motion tracking models are used to learn the relationship between flexible sensor data and human motion data.

[0055] like Figure 7 As shown, the network structure of the model in this embodiment adopts a Bi-LSTM network with one hidden layer and is implemented using PyTorch. Specifically, it uses linear input and output layers and a bidirectional LSTM layer with 512 hidden units. Training stops when the average tracking error drops below 0.1 degrees within 10 consecutive iterations. This embodiment uses an Adam optimizer with an initial learning rate of 0.001, which decays to 10% of the current rate every 10 iterations. To avoid gradient explosion, gradient pruning is preferably used in this embodiment. Furthermore, the batch size of input data to the network is set to 64 data points.

[0056] Since the flexible sensor is attached to clothing, it inevitably deforms with body movement, which accelerates its aging. During data collection, we observed irreversible bending deformation (i.e., aging) in the flexible sensor, causing its resistance to increase with data acquisition, thus increasing the flexible sensor baseline. This phenomenon can obfuscate the neural network during training. To mitigate the effects of this flexible sensor aging, this embodiment includes dynamic calibration of the flexible sensor data in the training set before model training, and proposes a novel dynamic calibration method:

[0057] Y out =Y in -(LN(ave(min(Y in ))))

[0058] Among them, Y in This represents a normalized time series of flexible sensor data (the flexible sensor data is normalized using min-max normalization). Specifically, a sliding window is used to process the raw flexible sensor data. The series length is 100. The sliding window moves two steps at a time; Y out This represents the calibrated flexible sensor data; LN represents the fully connected layer; min represents the min-pooling layer; ave represents the average-pooling layer; in this embodiment, the kernel sizes of the min and average-pooling layers are set to 25 and 10, respectively; the padding sizes of the min and average-pooling layers are set to 0 and 8, respectively; LN will restore the number of features reduced due to the min and average-pooling layers, and further learn the sensor baseline and ave(min(Y)). inA linear mapping between )); in addition, the data passes through two more fully connected layers before being input into the Bi-LSTM layer. ave(min(Y in The ')' represents the sensor baseline for each sample with a sequence length of 100. It's important to note that using only minimum pooling or average pooling layers will negatively impact tracking results. Therefore, we hypothesize that the sensor baseline calculated solely by minimum pooling layers is insufficiently accurate because people may still move the joint to a small extent. Thus, local minimum sensor data cannot adequately represent the sensor baseline. However, averaging the output of the minimum pooling layer can reduce this effect. Sensor data obtained through average pooling layers also cannot represent the baseline because the average value is susceptible to variations in the range of motion caused by human error. Therefore, combining both methods aids in sensor calibration.

[0059] In this embodiment, sensor data is normalized using min-max normalization before model training. Then, only data from (denoted as X = (x0, x1, ..., x...)) is processed. t ,…,x T Extract the lengths of five sequences from X and combine X with its corresponding ground motion data Y = (y0, y1, ..., y2). t ,…,y T The input sequence length is then fed into the Bi-LSTM to distinguish it from the sequence length of the sensor data before calibration. (To differentiate the input sequence length from the sequence length of the sensor data before calibration, we use the input sequence length and calibration sequence length as shown below.) Finally, we optimize the Bi-LSTM parameters and save them. The loss function used during training is:

[0060]

[0061] Among them, L S Let X(t) represent the loss value, X(t) represent the flexible sensor data, and Y(t) represent the human motion data (such as measured whole-body joint angles collected by the optical system). Although using future frames (where the last data point in the X sequence precedes the corresponding Y) would help with tracking results, given that our data sequence length is 5, there is little difference whether or not future frames are learned, so future frames are not used to avoid the degradation that can occur with future frame learning.

[0062] We observed jitter in the predicted motion, indicating a lack of smoothness and naturalness in the prediction results. To address this issue, this embodiment embeds a median pooling layer at the end of the Bi-LSTM network to smooth the predicted values ​​and obtain natural results. Specifically, the kernel size of the median pooling layer is set to 3, and the padding is set to 1.

[0063] S3: Track human motion based on the trained human motion tracking model.

[0064] By inputting the collected data from the flexible sensor into the trained human motion tracking model, the tracking results can be obtained.

[0065] The technical effects of the method in this embodiment will be explained through experiments below.

[0066] 1. Experimental equipment

[0067] We ran our experiment on a server with a triple-core CPU and a GPU (NVIDIA GTX Titan Xp, 12GB). The operating system was 64-bit Ubuntu 16.04.

[0068] 2. Quantitative assessment

[0069] (1) Overall and Detailed Performance: Overall, our model achieves a low average tracking error of 4.51°. To further observe, we further decompose the model evaluation into sub-evaluation tasks based on motion type (Table 1) and joint position (Table 2):

[0070] As shown in Table 1, we can conclude that: i) our model has a relatively high average tracking error for whole-body movements with a larger range of motion (such as climbing) and lower limb movements (such as bending over); ii) our model has a relatively low average tracking error for movements involving only limb joints (such as throwing a baseball) or whole-body movements with a smaller range of motion (such as swinging). iii) It can be observed that, except for bending over, the tracking error for all other common human movements is very low, less than 7°. This demonstrates the superiority of the method in this embodiment for tracking everyday human movements.

[0071] Table 1

[0072] action Error (degrees) action Error (degrees) Walk forward 3.67 swing 2.40 bend over 5.92 Rotation movement 2.26 Dribbling with both hands 5.07 walk with big strides 4.09 Climb up 6.96 Tai Chi 5.76 Walk back 3.89 Turn left 2.43 Throwing a baseball 4.15 Punch 5.34 Normal kick 3.80 sweep the floor 4.97

[0073] As shown in Table 2, the average tracking error of the arm and leg joints is greater than that of the shoulder and spine joints. This is likely because the arm and leg joints have a greater range of motion.

[0074] Table 2

[0075] joint Error (degrees) joint Error (degrees) right shoulder 1.83 Spine 1 1.07 left shoulder 1.65 Spine 2 1.07 Right upper limb 6.68 right arm 8.88 Right leg 5.28 Right forearm 6.58 left upper limb 5.69 left arm 8.22 left leg 4.78 left forearm 5.91 spine 0.79

[0076] (2) The comparison with the IMU solution shown in Table 3 also demonstrates the superiority of the method of this embodiment over its IMU alternative. The tracking error and jitter were evaluated using the same motion data collected from two subjects in the experiment.

[0077] Table 3

[0078] IMUs Method of this embodiment Error (degrees) 6.04 4.51 <![CDATA[Jitter (10 4 degrees / second 3 )]]> 4.451 2.450

[0079] As can be observed from Table 3, the average tracking error of the method in this embodiment is 1.53 degrees lower than that of the IMU alternative, and the jitter is 2.090 × m / s lower. Therefore, it can be concluded that the solution adopted in this embodiment is more accurate and more stable than its IMU alternative.

[0080] (3) Ablation Research

[0081] To demonstrate the effectiveness of the different components in the method of this embodiment, we conducted an ablation study, as shown in Table 4:

[0082] Table 4

[0083]

[0084]

[0085] Rows 1-4 of Table 4 demonstrate the effectiveness of the Bi-LSTM network used in this embodiment compared to other neural network architectures used in previous work. Rows 5-7 demonstrate the effectiveness of the median pooling layer (denoted as M) and dynamic calibration (denoted as C) used in this embodiment. It can be observed that adding the median pooling layer and dynamic calibration improves accuracy and reduces jitter. To further investigate the impact of the median pooling layer, we... Figure 8 The jitter distribution is plotted. It can be observed that the jitter of the prediction results using the median pooling layer is mainly distributed between 0 and 0.5. Furthermore, the jitter value using the median pooling layer is low, indicating that the median pooling layer used in this embodiment effectively reduces jitter. Figure 9 The changes in the left forearm over 5 seconds were depicted. It can be observed that after using the median pooling layer, the prediction curve becomes smoother and closer to the tracking result of the optical system, which proves the necessity of the median pooling layer.

[0086] (4) Justification for the selection of calibration sequence length

[0087] As shown in Table 5, we demonstrate the hyperparameters set in the method of this embodiment by conducting human motion prediction experiments with different calibration sequence lengths and comparing the results.

[0088] Table 5

[0089] Calibration sequence length 120 100 80 60 40 <![CDATA[D TE Error (degrees) on 4.58 4.51 4.58 4.56 5.6

[0090] We fixed the kernel sizes of the min-pooling and average-pooling layers to 15 and 10, respectively; then, we varied the input length of the data one by one. The results are shown in Table 5. It can be observed that the training set is smallest when the calibration sequence length is 100. Furthermore, the error is slightly larger when the calibration sequence length is 120 or 80. The average tracking error increases significantly when the calibration sequence length is 60 or 40. In summary, the optimal calibration sequence length is 100. Therefore, we set the input sequence length to 100. After the flexible sensor data passes through two pooling layers, we only extract the last 5 samples as input because the Bi-LSTM network performs best when the sequence length is 5. We also found that the model performs best when the kernel sizes of the min-pooling and average-pooling layers are 15 and 10, respectively. Although there is no significant difference in the overall results between different kernel sizes, the error is slightly better when the kernel size of the average-pooling layer is 10. Therefore, in this embodiment, kernel sizes of 15 and 10 were used for the min-pooling and average-pooling layers, respectively.

[0091] 3. Qualitative assessment

[0092] We also qualitatively evaluate the method of this embodiment by visualizing its output motion on a virtual human body. Figure 10 The figure shows the overall performance of the method of this embodiment for different types of motion.

[0093] (1) Ablation study:

[0094] Visual results of ablation studies, such as Figure 11 As shown, it can be observed that the Bi-LSTM method and the method of this embodiment have significant effects in most cases, especially when the object moves its arm, the tracking is closer to the real-world situation on the ground. In contrast, even though the TCN, FCN, and Transformer models track large and rapid arm movements, they tend to jitter slightly or move at incorrect angles. This may be due to their limited ability to fit small-range movements, which causes them to collapse slightly to the average pose. The performance of the method of this embodiment is closer to that of Bi-LSTM. However, when there are some large arm movements, the method of this embodiment performs better than Bi-LSTM.

[0095] (2) Subjective results compared to the IMU alternative

[0096] from Figure 12It can be concluded that the method of this embodiment outperforms the IMU, especially in the arm and leg joints. It is worth noting that the gait of the IMU sometimes differs from that of the optical system, possibly due to the magnetic environment underground and noisy data captured by the sensor. Therefore, the amplitude of motion is occasionally over-amplified due to errors in the acceleration term of the inertial measurement unit. Although we employed dynamic calibration and a median pooling layer to mitigate sensor aging and potential jitter, our tracking is closer to the results of the optical system. Furthermore, the method of this embodiment is unaffected by magnetic fields.

[0097] This invention utilizes only four flexible sensors to track 48 types of full-body human movements, covering most situations in our daily lives. Because this method is based on sparse flexible sensors, it achieves a better user experience without sacrificing motion tracking accuracy. To address unnatural jitter in the prediction results, this embodiment employs a median pooling layer to smooth the movement. This method improves the wearing experience of the motion tracking flexible sensor suit while maintaining stable tracking of everyday human activities using only four sensors.

[0098] Furthermore, while the flexible sensor used in this embodiment is biocompatible and ideal for improving the wearing experience, it is prone to aging, which can increase prediction errors. Since aging of flexible sensors typically leads to changes in their reading range, this embodiment further proposes a novel dynamic calibration technique to readjust the data range of the flexible sensor, which further improves prediction accuracy. Experimental results show that the method in this embodiment achieves a stable tracking error as low as 4.51 degrees across different motion types using only four sensors.

[0099] The method in this embodiment has the following technical effects:

[0100] (1) A data-driven method based on flexible sensors for tracking daily human activities. Through several pooling layers, this embodiment improves robustness in the face of sensor aging and replacement.

[0101] (2) A method to mitigate prediction output jitter and handle sparse input. The Bi-LSTM model and median pooling layer not only help reduce the jitter of the prediction results, but also have good generality on the test set.

[0102] (3) With only four sensors in the prototype, the tracking error of the system prototype is 4.51 degrees, which can be used to track users’ daily activities.

[0103] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A human motion tracking method based on a flexible sensor, characterized in that, Includes the following steps: S1: Simultaneously collect data from four flexible sensors worn by the tester at four joints of the limbs and human motion data of the tester within a certain period of time, and generate data pairs that correspond one-to-one between flexible sensor data and human motion data based on the collection time. Construct a training set based on all data pairs collected from all testers. S2: Construct a human motion tracking model based on a Bi-LSTM network, using flexible sensor data as the input to the model and the model output as predicted human motion data. Train the model using a training set. S3: Track human motion based on the trained human motion tracking model; The flexible sensors are configured to connect to the four joints of the corresponding limbs of the bodysuit worn on the human body. The flexible sensor is connected to the bodysuit in the following way: the flexible sensor is placed in the pocket at the joint position of the corresponding limb of the bodysuit. Before training the model, the following steps are included: dynamically calibrating the flexible sensor data in the training set, and then using the dynamically calibrated data to train the model; the specific formula for dynamic calibration is: Among them, Y in Y represents normalized flexible sensor data for a time series. out This represents the calibrated flexible sensor data, where LN represents the fully connected layer, min represents the minimum pooling layer, and ave represents the average pooling layer.

2. The human motion tracking method based on a flexible sensor according to claim 1, characterized in that: Human motion data includes data on the joints of the arms, forearms, thighs, legs, and spine.

3. The human motion tracking method based on a flexible sensor according to claim 1, characterized in that: The model's network structure embeds a median pooling layer at the end of the Bi-LSTM network.

4. The human motion tracking method based on a flexible sensor according to claim 1, characterized in that: The connection between the flexible sensor and the wire is pluggable.

Images