A cross-individual human behavior recognition method based on self-training and active inquiry

Abstract

This invention provides a cross-individual human behavior recognition method based on self-training and active querying. The method includes: obtaining a labeled source domain dataset and an unlabeled target domain dataset; processing acceleration data from both datasets; proportionally dividing the target domain data into unlabeled training and testing sets; constructing a feature extractor based on a dual-channel convolutional network to simultaneously extract temporal and spatial features from the input two-dimensional acceleration data; and employing a cross-individual adaptation algorithm based on confidence thresholding, sparse querying, and label propagation to train the model on the unlabeled training set in the target domain, thus mitigating the adaptation problem caused by differences in cross-individual data distribution. This invention can improve the accuracy of behavior recognition in cross-individual scenarios, reduce development costs and user burden, and provide a more intelligent and adaptive solution for wearable devices such as exoskeleton robots.

Description

Technical Field

[0001] This invention relates to the field of behavior recognition technology, specifically to a cross-individual human behavior recognition method based on self-training and active inquiry, applicable to human behavior recognition in scenarios such as exoskeleton robots and wearable devices. Background Technology

[0002] With the rapid development of robotics, edge computing, and wearable device technology, exoskeleton robots are increasingly being used in various fields such as medical rehabilitation, industrial assistance, and military training. Particularly in postoperative rehabilitation, exoskeleton robots can provide patients with personalized movement support and rehabilitation training, effectively promoting their recovery. Human behavior recognition technology, as one of the key technologies for achieving intelligent and adaptive control of exoskeleton robots, is of paramount importance. This technology can accurately identify and predict the wearer's movement intentions and behavioral patterns in real time, thereby providing precise control commands to the exoskeleton robot, supporting various types of movement modes, and meeting the diverse needs of users.

[0003] Currently, human behavior recognition technology is mainly implemented through various devices such as computer vision, Wi-Fi signals, and wearable sensors. Computer vision technology captures images of user movements using cameras and then uses image processing and machine learning algorithms for behavior recognition. However, this method has high requirements for the operating environment; once the user moves out of the camera's field of view, the recognition accuracy drops significantly, or even fails. Wi-Fi signal technology relies on wireless signal transmission. Although not limited by camera range, it is still affected by signal coverage and stability. When the user moves out of the Wi-Fi signal's transmission range, behavior recognition will also be affected.

[0004] In contrast, wearable sensor technology offers significant advantages such as being unrestricted by location and protecting user privacy. These sensors can be worn directly on the user's body to collect real-time motion data, such as acceleration and angular velocity, providing a rich data source for human behavior recognition. Among various wearable sensor signals, acceleration signals are frequently used for human behavior recognition tasks due to their ease of acquisition and processing. Inertial Measurement Units (IMUs), as a commonly used wearable sensor, can accurately collect acceleration information of different parts of the human body relative to the three axes of the coordinate system during movement, providing crucial data support for behavior recognition.

[0005] However, despite significant progress in human behavior recognition technology, wearable sensor technology still faces several challenges. Developing human behavior recognition models typically involves using one or more sensors to collect data, extracting features from the data, and then inputting these features, along with motion category labels, into a classification model for training. This approach performs well when training and testing data follow the same distribution, but due to differences in individual behavioral habits and movement patterns, acceleration data collected from different individuals also varies significantly. Even when performing the same type of activity, the data distribution from different individuals can differ. Therefore, models trained in this way often only apply to specific users and are difficult to generalize to new users.

[0006] When human behavior recognition tasks involve multiple users, data collection, processing, and model training must be performed separately for each user. This not only increases development costs and time but also consumes a significant amount of user time and effort, reducing the user experience. Therefore, enhancing the generalization ability of human behavior recognition models and improving their accuracy in recognizing new users has become a crucial issue that urgently needs to be addressed. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a cross-individual human behavior recognition method based on self-training and proactive inquiry. The aim is to improve the accuracy of behavior recognition in cross-individual scenarios by optimizing model training and adaptation mechanisms, reducing development costs and user burden, and providing a more intelligent and adaptive solution for the application of wearable devices such as exoskeleton robots.

[0008] The present invention achieves the above objectives through the following technical solutions: A cross-individual human behavior recognition method based on self-training and proactive inquiry includes: Wearable sensors are fitted to the first user group. When the first user group performs the required behavior activities, the wearable sensors are used to synchronously collect the acceleration data of the corresponding body parts during the movement, and a labeled dataset of the source domain is constructed. At the same time, the synchronous collection process is repeated for the second user group to obtain an unlabeled dataset of the target domain. Data processing is performed on acceleration data in the source domain labeled dataset and the target domain unlabeled dataset, including detecting and processing missing values ​​in the data, and performing sliding window sampling on the processed data; The processed target domain data is divided into unlabeled training and test sets according to a certain ratio; A feature extractor based on a dual-channel convolutional network is built to extract features in both time and space dimensions from the input two-dimensional acceleration data. A cross-individual adaptation algorithm based on confidence threshold, sparse query, and label propagation is adopted to train the model on the unlabeled training set in the target domain and alleviate the adaptation problem caused by the difference in cross-individual data distribution.

[0009] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry is provided. When performing missing value detection, the method detects abnormal missing points in the collected acceleration time series data by using the sliding difference method. That is, when the sampling time interval between consecutive data points exceeds a preset threshold, it is determined that there is a missing value at that position. At the same time, the method assists in identifying non-continuous missing values ​​caused by transmission interruption by detecting sudden changes in data amplitude. For the detected missing values, a time series-based bidirectional linear interpolation method is used, which uses the acceleration values ​​of N normal data points before and after the missing point to perform weighted average interpolation, with the weights decreasing exponentially with time distance; if the missing value is located at the beginning or end of the data sequence or the consecutive missing length exceeds M points, the nearest neighbor imputation method is used, which replaces the missing value with the nearest valid data point value. The interpolated data must pass a consistency check, meaning the sum of the absolute values ​​of the second-order differences between the interpolated segment and the original data must be ≤0.5 m / s. 4 Otherwise, it will be marked as abnormal and trigger manual review.

[0010] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry is provided. When performing sliding window sampling, the sliding window length is set to 300ms, the sliding step size is set to 30ms, and the ratio of window length to step size is 10:1. The interpolated acceleration time series data is truncated in units of sliding windows, and the data in each window are arranged in chronological order as a one-dimensional vector. If the data at the end of the window is less than 300ms, it is padded to the window length by zero padding or mirror padding, but the padded segment is marked and excluded in subsequent feature extraction. For the multi-sensor data within each sliding window, they are concatenated into a two-dimensional array according to the sensor dimension, where each row represents a sliding window sample and each column represents a data channel [ax, ay, az] of a sensor.

[0011] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry enhances data diversity in the dataset after sliding window sampling in the following ways: Random Gaussian noise was added to simulate sensor noise; Random time shifting to simulate sampling time jitter; The data-augmented samples need to pass validity verification, meaning the acceleration amplitude of the samples must be within the range of -10 m / s². 2 10m / s 2 Within the range, and the frequency domain energy distribution conforms to the characteristics of human movement; The generated source domain labeled dataset and target domain unlabeled dataset each contain the following information: Source domain samples containing sliding window data and motion category labels; Target domain sample containing sliding window data.

[0012] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry is provided. The feature extractor based on a dual-channel convolutional network achieves joint extraction of temporal and spatial dimensions of features through the following design: Upper channel: Contains N levels of convolutional modules, each module consisting of a one-dimensional convolutional layer, a batch normalization layer, and a non-linear activation function layer; the one-dimensional convolutional layer slides along the time dimension, extracting local dynamic features in the time series in parallel through multiple convolutional kernels; the number of output channels of each level of convolutional module increases exponentially to abstract higher-order features in the time dimension layer by layer; The lower-level channels contain M levels of convolutional modules. Each module consists of a two-dimensional convolutional layer, a batch normalization layer, and a non-linear activation function layer. The two-dimensional convolutional layer slides along the sensor dimension and extracts the spatial correlation features between multiple sensor data in parallel through multiple convolutional kernels. The number of output channels of each level of convolutional module increases synchronously with the upper-level channels to ensure the alignment of temporal and spatial features.

[0013] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry is provided. When extracting time features, the upper channel performs time-dimensional convolution on the input two-dimensional data, with a dimension of [L×C], where L is the time step and C is the number of sensors. The output feature map has a dimension of [L'×F_t], where L' is the length of the time dimension after compression and F_t is the number of time feature channels. When extracting spatial features, the lower-level channels perform spatial dimension convolution on the same input data, and the output feature map dimension is [L×F_s], where F_s is the number of spatial feature channels; the spatial features are aggregated along the time dimension through the global average pooling layer, and the output dimension is [1×F_s], so as to extract the global spatial correlation pattern between sensors; The temporal feature map ([L'×F_t]) is flattened into a one-dimensional vector (dimension [L'×F_t]), and the spatial feature vector ([1×F_s]) is copied and expanded to the same length and then concatenated. The final fused feature dimension is [L'×F_t + F_s]. The concatenated feature vector is then subjected to dimensionality reduction and nonlinear mapping through a fully connected layer to form a joint spatiotemporal feature representation. The output feature map of the upper channel is transposed, i.e., the dimension is transformed [L'×F_t]→[F_t×L'] and multiplied point by point with the intermediate feature map of the lower channel, i.e. the dimension is [L×F_s], to achieve cross-modal fusion of time-space features; the output joint spatiotemporal feature vector is input into the classifier for motion category prediction.

[0014] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry is provided, wherein the classifier includes: Linear layer: The input is the joint spatiotemporal feature vector output by the dual-channel convolutional network, with dimension [F_final]. The output dimension is the total number of motion categories K. The linear mapping from features to categories is achieved through the learnable weight matrix W∈R^{F_final×K} and the bias vector b∈R^K. The weights of the linear layer are initialized with a Xavier uniform distribution, ranging from [-√(6 / (F_final+K)), √(6 / (F_final+K))], to alleviate the gradient vanishing problem. Batch standardization layer: Immediately following the linear layer, the output features are normalized, and the mean μ and variance σ are calculated. 2 It achieves adaptive adjustment through learnable parameters γ and β; the batch normalization formula is: y = γ ×(x - μ) / sqrt(σ 2 + ε) + β, where ε=1e-5 is the numerical stability constant, where γ is the scaling factor and β is the translation factor.

[0015] According to the present invention, a cross-individual human behavior recognition method based on self-training and active inquiry is provided. Before the linear layer, i.e. between the output of the feature extractor and the input of the linear layer, the ReLU activation function is used to perform a nonlinear transformation on the joint spatiotemporal feature vector, with the formula f(x)=max(0, x); the ReLU function is used to sparse the feature representation. After the batch normalization layer, the Softmax function is applied to the output vector to convert the original scores into a probability distribution, with the formula: P(y=k|x) = exp(z_k) / Σ_{i=1}^K exp(z_i), where z_k is the k-th element of the batch normalization layer output; the probability vector output by the Softmax function has a dimension of [K], and each element represents the predicted probability that the input sample belongs to the corresponding motion category, and satisfies Σ_{k=1}^KP(y=k|x)=1; The classifier's final output is a probability vector [p_1, p_2, ..., p_K], where p_k is the predicted probability of the k-th class of motion; the predicted class label is determined by the argmax operation, i.e., y_pred = argmax_k(p_k); The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. The formula is: L = -Σ_{k=1}^K y_k log(p_k), where y_k is the one-hot encoding of the true label.

[0016] According to the present invention, a cross-individual human behavior recognition method based on self-training and active querying is provided. The cross-individual adaptation algorithm adopts a joint optimization strategy based on confidence threshold, sparse query, and label propagation, and specifically achieves adaptive adjustment of the cross-domain human behavior recognition model through the following steps: Train a human behavior recognition model on a source domain labeled dataset, and save the parameters of the feature extractor and classifier. The feature extractor is composed of a dual-channel convolutional network, and the classifier is composed of a linear layer and a batch normalization layer. Input the unlabeled training set samples of the target domain into the pre-trained model, and save the sample feature vector output by the feature extractor, the predicted label output by the classifier, and the confidence vector. For each sports category, samples are sorted in descending order of confidence and the top 10% are selected as high-confidence samples. For each sports category, the mean vector of the high-confidence sample feature vector is calculated as the feature center of that category. Calculate the Euclidean distance from all samples in the target domain training set to the center of each motion class; For each sample, determine its nearest class center c_1 and second nearest class center c_2, and calculate the distance ratio r = d(x,c_1) / d(x, c_2); If r is close to 1 (e.g., 0.9 ≤ r ≤ 1.1), then the sample is considered a boundary sample and there may be classification uncertainty.

[0017] According to the present invention, a cross-individual human behavior recognition method based on self-training and active querying is provided. When performing sparse label query, for each pair of motion categories, the top 10 samples with the largest distance ratio r are selected as boundary samples. The true labels of these boundary samples are queried from the label library of the target domain training set; for each boundary sample, the top 3% of samples with the closest Euclidean distance in the feature space are searched as nearest neighbors; if the boundary sample has a true label, its label is propagated to the nearest neighbors; if the boundary sample label is "to be propagated", propagation is skipped. The predicted labels of high-confidence samples are used as pseudo-labels; the propagated labels are merged with the high-confidence pseudo-labels to construct a labeled training set for the target domain, where the source of the sample labels is marked as "high confidence" or "propagation"; With the feature extractor parameters fixed, only the linear and batch normalization layers of the classifier are fine-tuned; Using the merged target domain labeled training set, the classifier parameters are updated through the cross-entropy loss function (formula: L = -Σ_{i=1}^N y_i log(p_i)) and the Adam optimizer (learning rate = 1e-4, β1 = 0.9, β2 = 0.999); The fine-tuned model is evaluated on the target domain test set to calculate the accuracy of human behavior recognition.

[0018] Therefore, compared with the existing technology, the cross-individual human behavior recognition method based on self-training and active inquiry proposed in this invention has the following beneficial effects: This invention effectively enhances the model's adaptability to differences in individual behavior by introducing self-training and proactive query mechanisms. During model training, pre-training is performed using a labeled dataset from the source domain, and fine-tuning is achieved through a cross-individual adaptation algorithm using an unlabeled dataset from the target domain. This allows the model to better capture and adapt to behavioral feature differences among individuals. This invention significantly improves the model's generalization ability in cross-individual scenarios, enabling it to achieve high behavior recognition accuracy without requiring extensive retraining for each new user.

[0019] This invention significantly reduces the need for manual annotation through a sparse query and label propagation strategy. During cross-individual adaptation, sparse queries are performed only on a small number of boundary samples to obtain true labels, and these labels are then extended to more similar samples through label propagation. This method not only reduces the workload of manual annotation but also mitigates model performance degradation caused by annotation errors or inconsistencies. Simultaneously, since users are not required to participate in a large amount of data annotation, it also reduces their burden and improves the user experience.

[0020] The method of this invention achieves high behavior recognition accuracy in cross-individual scenarios. Through self-training and proactive querying mechanisms, the model can better learn and adapt to the behavioral characteristics of different individuals, thus achieving accurate recognition even when facing new users. Furthermore, sparse query and label propagation strategies further optimize the model's label utilization efficiency, improving its learning performance under limited label conditions. These advantages work together to significantly improve the behavior recognition accuracy of the method in cross-individual scenarios.

[0021] In summary, the cross-individual human behavior recognition method based on self-training and active inquiry proposed in this invention has achieved significant beneficial effects in improving model generalization ability, reducing annotation costs, and improving recognition accuracy. It not only provides new ideas and methods for the development of human behavior recognition technology, but also provides strong technical support for the application and promotion in related fields.

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0023] Figure 1 This is a flowchart of an embodiment of a cross-individual human behavior recognition method based on self-training and active inquiry according to the present invention.

[0024] Figure 2 This is a schematic diagram of the dual-channel convolutional network in an embodiment of a cross-individual human behavior recognition method based on self-training and active querying according to the present invention.

[0025] Figure 3 This is a schematic diagram of the cross-individual adaptation algorithm in an embodiment of a cross-individual human behavior recognition method based on self-training and active querying according to the present invention. Detailed Implementation

[0026] 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.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] See Figure 1 This embodiment provides a cross-individual human behavior recognition method based on self-training and active inquiry, including: Step S1: Equip the first user group with wearable sensors. When the first user group performs the required behavioral activities, use the wearable sensors to synchronously collect the acceleration data of the corresponding body parts during the movement, and construct the source domain labeled dataset. Step S2 involves repeating the synchronous data collection process for the second user group to obtain the unlabeled dataset for the target domain. Step S3 involves processing the acceleration data in the source domain labeled dataset and the target domain unlabeled dataset, including detecting and processing missing values ​​in the data, and performing sliding window sampling on the processed data. Step S4: Divide the processed target domain data into unlabeled training and test sets according to the proportions. Step S5: Build a feature extractor based on a dual-channel convolutional network to extract features in both time and space dimensions from the input two-dimensional acceleration data. Step S6: A cross-individual adaptation algorithm based on confidence threshold, sparse query and label propagation is adopted to train the model on the unlabeled training set in the target domain and alleviate the adaptation problem caused by the difference in cross-individual data distribution.

[0029] In step S1 above, the first user group consists of existing subjects, who are fitted with wearable sensors. The placement of the sensors can be selected based on the characteristics of the target action, such as placing them on the chest, wrist, thigh, or ankle.

[0030] In step S2 above, the second user group is the new subjects. The collection process in step S1 is repeated for the new subjects to obtain the target domain unlabeled dataset. This dataset only contains the collected acceleration data and lacks the real labels representing the motion category.

[0031] Specifically, this embodiment uses a low-power, high-precision triaxial accelerometer array integrated into a flexible wearable substrate material. It achieves real-time sensing of acceleration signals through piezoelectric or capacitive working principles. The sensor array establishes a low-latency communication link with the mobile smart terminal via Bluetooth 5.0 protocol and adopts a multi-threaded concurrent transmission mechanism to ensure that each sensor node independently collects data and transmits it to the terminal in real time.

[0032] The hardware system includes a dynamic power management module that adaptively adjusts the sensor sampling frequency based on motion intensity, reducing power consumption to below 10mW when stationary and increasing the sampling rate to 200Hz when in motion. When the first user group performs preset activities, the sensor array data is synchronously triggered via a mobile terminal APP, and the behavior type label is recorded. The collected raw acceleration data undergoes hardware-level filtering, using dual third-order Butterworth filters (0.3Hz high-pass filter + 25Hz low-pass filter) to remove gravity components and high-frequency noise. After the data is transmitted to the terminal, a sliding window segmentation algorithm (window length 1.28 seconds, overlap rate 50%) is used to extract time-domain features (such as mean, variance, and energy) and frequency-domain features (such as FFT spectrum energy distribution), which are then combined with the behavior labels to construct a source domain dataset.

[0033] In this embodiment, when the second user group performs free activities, the sensor array collects acceleration data with the same hardware configuration, but the terminal APP does not record behavior labels; after the collected data is processed by the same filtering and feature extraction process as the source domain, only the original feature vector is stored to form an unlabeled target domain dataset; the mobile terminal encrypts and uploads the source domain and target domain data to the cloud server through a distributed storage protocol, and uses homomorphic encryption technology to ensure privacy and security during data transmission and storage.

[0034] Specifically, the cloud server performs spatiotemporal alignment of the source and target domain datasets, matching user activity data within the same time period based on GPS positioning information and timestamps; the Z-score normalization method is used to normalize the feature vectors to eliminate the dimensional differences between different user groups; the constructed dataset meets the following conditions: the source domain data volume is ≥5000 samples, the target domain data volume is ≥10000 samples, and the KL divergence of the two datasets in the feature space distribution is ≤0.2.

[0035] Specifically, in this embodiment, the wearable sensor array equipped for the second user group only includes an acceleration acquisition module, disabling or removing any auxiliary sensors related to behavior classification (such as gyroscopes and electromyography sensors), and does not integrate behavior recognition algorithms in the hardware circuitry; the sensors communicate with the mobile terminal through a one-way data transmission mode, only uploading raw triaxial acceleration data (unit: m / s²). 2 The terminal APP does not receive or store any user-inputted behavioral tag information; the hardware system adopts a physical isolation design, and blocks other data interaction channels between the sensor and the terminal except for acceleration data through a hardware firewall.

[0036] During the free activity period of the second user group, the mobile terminal APP runs in silent mode, only triggering the data acquisition function of the sensor array, without displaying any interactive interface related to behavior labeling; the acquired acceleration data is marked with the acquisition time by timestamp, but is not associated with any motion category label, and is directly stored as a raw binary file after being transmitted to the terminal; the terminal APP has a built-in logic verification module that monitors in real time whether there are label fields in the data stream, and if label data is detected, it is automatically discarded and a hardware reset is triggered to clear residual information.

[0037] In this embodiment, the collected raw acceleration data is preprocessed, and a sliding window is used to detect and remove silent states (acceleration amplitude < 0.1 m / s²). 2 Redundant data segments (with a duration > 30 seconds) were excluded; random noise (noise amplitude ≤ 0.05 m / s²) was added to the acceleration data using differential privacy technology. 2 This approach preserves motion characteristics while preventing user identification; the final unlabeled dataset for the target domain contains only time-series triaxial acceleration data (format: [timestamp, ax, ay, az]), and no sample in the dataset contains any metadata related to the motion category.

[0038] Furthermore, the unlabeled dataset is encrypted and stored in a multi-node cloud using distributed hash table (DHT) technology, and data access logs are recorded using blockchain technology to ensure that the data is immutable and used only for authorized model training. The constructed target domain unlabeled dataset meets the following conditions: the data volume is ≥15,000 samples, the duration of each sample is ≥5 seconds, and the probability of overlap of motion categories between any two samples in the dataset is <5% (based on the cross-validation results of subsequent model predictions).

[0039] In step S3 above, during missing value detection, the collected acceleration time series data is analyzed using the sliding difference method to detect abnormal missing points. Specifically, if the sampling time interval between consecutive data points exceeds a preset threshold (e.g., 50ms), a missing value is determined to exist at that location. Simultaneously, abrupt changes in data amplitude are detected (e.g., acceleration difference between adjacent points > 5m / s²). 2 ) Assist in identifying discontinuities caused by transmission interruptions; For the detected missing values, a time-series-based bidirectional linear interpolation method is used, which uses the acceleration values ​​of N normal data points (N=3) before and after the missing point for weighted average interpolation. The weights decay exponentially with time distance (decay coefficient α=0.8). If the missing value is located at the beginning or end of the data sequence or the consecutive missing length exceeds M points (M=10), the nearest neighbor imputation method is used to replace the missing value with the nearest valid data point value. The interpolated data must pass a consistency check, meaning the sum of the absolute values ​​of the second-order differences between the interpolated segment and the original data must be ≤0.5 m / s. 4 Otherwise, it will be marked as abnormal and trigger manual review.

[0040] When performing sliding window sampling, the sliding window length is set to 300ms, the sliding step size is set to 30ms, and the ratio of window length to step size is 10:1 to ensure that there is appropriate overlap between samples to enhance the generalization ability of the model. The interpolated acceleration time series data is truncated in units of sliding windows, and the data in each window are arranged in chronological order as a one-dimensional vector. If the data at the end of the window is less than 300ms, it is padded to the window length by zero padding or mirror padding, but the padded segment is marked and excluded in subsequent feature extraction. For the multi-sensor data (such as a triaxial accelerometer) within each sliding window, it is concatenated into a two-dimensional array according to the sensor dimension, where each row represents a sliding window sample and each column represents a data channel [ax, ay, az] of a sensor; if the number of sensors is K, then the dimension of each sample is (window length / sampling interval) × K. For example, when the sampling rate is 100Hz, the dimension of a single sample is 30 × K.

[0041] For the dataset sampled by the sliding window, data diversity is enhanced in the following ways: Randomly add Gaussian noise (noise amplitude ≤ 0.02 m / s) 2 To simulate sensor noise; Random time shift (shift range ±10ms) to simulate sampling time jitter; The data-augmented samples need to pass validity verification, meaning the acceleration amplitude of the samples must be within the range of -10 m / s². 2 10m / s 2 Within ], and the frequency domain energy distribution conforms to the characteristics of human movement (main frequency ≤ 20Hz). The generated source domain labeled dataset and target domain unlabeled dataset each contain the following information: Source domain samples containing sliding window data and motion category labels; Target domain sample containing sliding window data.

[0042] The dataset is stored in HDF5 format, which supports efficient random access, and each data block contains metadata (such as sampling time, sensor ID, interpolation label) to facilitate subsequent model training and debugging.

[0043] Through the above steps, this embodiment achieves efficient repair of missing values ​​in acceleration data and precise control of sliding window sampling. While ensuring data integrity, it significantly improves the quantity and diversity of data samples, providing high-quality, large-scale, and structured data input for neural network training, and effectively enhancing the model's generalization ability and training efficiency.

[0044] In step S4 above, all data in the target domain are proportionally divided into an unlabeled training set (70%) and a test set (30%). All data in the source domain are labeled and used for pre-training the model.

[0045] In step S5 above, such as Figure 2 As shown, the feature extractor based on a dual-channel convolutional network achieves joint extraction of temporal and spatial features through the following design: Upper-level channel (temporal feature extraction channel): Contains N (N≥2) levels of convolutional modules. Each module consists of a one-dimensional convolutional layer (with a kernel size of [T×1], where T is the time dimension window size), a batch normalization layer, and a non-linear activation function layer. The one-dimensional convolutional layer slides along the time dimension (the row direction of the two-dimensional data) and extracts local dynamic features (such as acceleration change trends and periodic patterns) in the time series in parallel through multiple convolutional kernels (number ≥16). The number of output channels of each level of convolutional module increases exponentially (e.g., 16→32→64) to abstract higher-order features of the time dimension layer by layer. The lower-level channel contains M (M≥2) level convolutional modules. Each module consists of a two-dimensional convolutional layer (with a kernel size of [1×C], where C is the number of sensor channels), a batch normalization layer, and a non-linear activation function layer (LeakyReLU, slope=0.1). The two-dimensional convolutional layer slides along the sensor dimension (column direction of the two-dimensional data) and extracts spatial correlation features between multiple sensor data (such as the coordination of limb movements and the spatial distribution of motion patterns) in parallel through multiple convolutional kernels (number ≥16). The number of output channels of each level of convolutional module increases synchronously with the upper-level channel to ensure the alignment of the temporal and spatial features.

[0046] When extracting temporal features, the upper channel performs temporal convolution on the input two-dimensional data, with a dimension of [L×C], where L is the time step and C is the number of sensors. The output feature map has a dimension of [L'×F_t], where L' is the length of the time dimension after compression and F_t is the number of temporal feature channels. The temporal dimension is further compressed through a max pooling layer (pooling window [2×1], stride 2) to enhance feature robustness.

[0047] When extracting spatial features, the lower-level channels perform spatial dimension convolution on the same input data, and the output feature map dimension is [L×F_s], where F_s is the number of spatial feature channels; the spatial features are aggregated along the time dimension through the global average pooling layer, and the output dimension is [1×F_s], so as to extract the global spatial correlation pattern between sensors; The temporal feature map ([L'×F_t]) is flattened into a one-dimensional vector (dimension [L'×F_t]), and the spatial feature vector ([1×F_s]) is copied and expanded to the same length and then concatenated. The final fused feature dimension is [L'×F_t + F_s]. The concatenated feature vector is then passed through a fully connected layer (dimension compressed to [F_final], F_final≤256) for dimensionality reduction and nonlinear mapping to form a joint spatiotemporal feature representation. A cross-channel attention mechanism (such as the Squeeze-and-Excitation module) is introduced between adjacent convolutional modules in a dual-channel architecture. Channel weight vectors are generated through global pooling and fully connected layers to dynamically adjust the extraction intensity of temporal and spatial features. The output feature map of the upper channel is transposed, i.e., the dimension is transformed [L'×F_t]→[F_t×L'] and multiplied point by point with the intermediate feature map of the lower channel, i.e. the dimension is [L×F_s], to achieve cross-modal fusion of time-space features; the output joint spatiotemporal feature vector is input into the classifier for motion category prediction.

[0048] Through the above steps, the dual-channel convolutional network efficiently extracts the dynamic features and structural correlations of acceleration data from the temporal and spatial dimensions, and achieves complementary enhancement of spatiotemporal information through cross-channel interaction and feature fusion mechanisms, which significantly improves the motion recognition model's ability to represent complex human movements and its classification accuracy.

[0049] Specifically, the classifier in this embodiment includes: Linear layer: The input is the joint spatiotemporal feature vector output by the dual-channel convolutional network, with dimension [F_final]. The output dimension is the total number of motion categories K. The linear mapping from features to categories is achieved through the learnable weight matrix W∈R^{F_final×K} and the bias vector b∈R^K. The weights of the linear layer are initialized with a Xavier uniform distribution, ranging from [-√(6 / (F_final+K)), √(6 / (F_final+K))], to alleviate the gradient vanishing problem. Batch standardization layer: Immediately following the linear layer, the output features are normalized, and the mean μ and variance σ are calculated. 2 It achieves adaptive adjustment through learnable parameters γ and β; the batch normalization formula is: y = γ ×(x - μ) / sqrt(σ 2 + ε) + β, where ε=1e-5 is the numerical stability constant, where γ is the scaling factor and β is the translation factor.

[0050] Before the linear layer, that is, between the output of the feature extractor and the input of the linear layer, the ReLU activation function is used to perform a nonlinear transformation on the joint spatiotemporal feature vector, with the formula f(x)=max(0, x). The ReLU function enhances the model's ability to fit nonlinearly to complex motion patterns by sparsening the feature representation.

[0051] After the batch normalization layer, the Softmax function is applied to the output vector to convert the original scores into a probability distribution, with the formula: P(y=k|x) = exp(z_k) / Σ_{i=1}^K exp(z_i), where z_k is the k-th element of the batch normalization layer output; the probability vector output by the Softmax function has a dimension of [K], and each element represents the predicted probability that the input sample belongs to the corresponding motion category, and satisfies Σ_{k=1}^KP(y=k|x)=1; The classifier's final output is a probability vector [p_1, p_2, ..., p_K], where p_k is the predicted probability of the k-th class of motion; the predicted class label is determined by the argmax operation, i.e., y_pred = argmax_k(p_k); The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. The formula is: L = -Σ_{k=1}^K y_k log(p_k), where y_k is the one-hot encoding of the true label.

[0052] During training, the label smoothing technique is used to adjust the one-hot encoding of the real label to y_k' = (1-ε)×y_k + ε / K (ε=0.1) to alleviate overfitting and enhance the model's generalization ability.

[0053] Through the above steps, the classifier, with linear layers and batch normalization layers as its core architecture, combined with ReLU and Softmax activation functions, achieves efficient and stable motion category probability prediction. The ReLU function enhances the non-linear expressive power of features, the Softmax function provides interpretable probability output, and the batch normalization layer and label smoothing technique further improve the model's convergence speed and generalization performance, providing a high-performance classification solution for human motion recognition tasks.

[0054] In step S6 above, during the model training process in the target domain, the following method is used: Figure 3 The algorithm shown alleviates the cross-individual adaptation problem. This algorithm is a cross-individual adaptation algorithm based on confidence thresholding, sparse querying, and label propagation, used to improve the accuracy of human behavior recognition in cross-individual scenarios. The algorithm first uses a confidence threshold to select samples with relatively high confidence, then obtains the true labels of a small number of samples by sparsely querying the target individual. Finally, it maximizes the query benefit through label propagation. The algorithm includes the following steps: 1) Train the human behavior recognition model on the source domain dataset and save the pre-trained model parameters.

[0055] 2) Input all samples in the target domain training set into the pre-trained model above, and save the sample feature vectors output by the feature extractor, the classification labels output by the model, and the corresponding confidence scores.

[0056] 3) Based on the classification results, select the top 10% of samples with confidence for each sports category, calculate the average feature vector of these samples for each sports category, and obtain the center of the sports category in the demand space.

[0057] 4) Calculate the Euclidean distance from all samples in the target domain training set to the center of each class, and then calculate the ratio of the sample's distance to the nearest and second nearest center. The closer this ratio is to 1, the more likely the sample is to be located at the boundary between the two classes, meaning that the model's classification of this sample has greater uncertainty.

[0058] 5) For each pair of motion categories, select the top ten samples with the largest distance ratios as their boundary samples. Query the true labels of the selected boundary samples from the label library of the target domain training set. Then, propagate the queried labels to their nearest neighbors in the feature space (the top 3% closest based on Euclidean distance).

[0059] 6) Merge the propagated sample labels with the pseudo-labels of the high-confidence center samples to construct a labeled training set in the target domain.

[0060] 7) After constructing the labeled training set, fine-tune the classifier part of the pre-trained model. During the fine-tuning process, use the cross-entropy loss function and the Adam optimizer to update the network.

[0061] 8) After the model parameters are updated, test the current model in the target domain test set and calculate the accuracy of human behavior recognition in cross-individual scenarios.

[0062] In this embodiment, the cross-individual adaptation algorithm adopts a joint optimization strategy based on confidence threshold, sparse query, and label propagation, specifically achieving adaptive adjustment of the cross-domain human behavior recognition model in the following ways: Train a human behavior recognition model on a source domain labeled dataset, and save the parameters of the feature extractor and classifier. The feature extractor is composed of a dual-channel convolutional network, and the classifier is composed of a linear layer and a batch normalization layer. Input the unlabeled training set samples of the target domain into the pre-trained model, and save the sample feature vector output by the feature extractor, the predicted label output by the classifier, and the confidence vector. For each sports category, samples are sorted in descending order of confidence and the top 10% are selected as high-confidence samples. For each sports category, the mean vector of the high-confidence sample feature vector is calculated as the feature center of that category. Calculate the Euclidean distance from all samples in the target domain training set to the center of each motion class; For each sample, determine its nearest class center c_1 and second nearest class center c_2, and calculate the distance ratio r = d(x,c_1) / d(x, c_2); If r is close to 1 (e.g., 0.9 ≤ r ≤ 1.1), then the sample is considered a boundary sample and there may be classification uncertainty.

[0063] When performing sparse label queries, for every two groups of motion categories, the top 10 samples with the largest distance ratio r are selected as boundary samples; The true labels of these boundary samples are queried from the label library of the target domain training set; for each boundary sample, the top 3% of samples with the closest Euclidean distance in the feature space are searched as nearest neighbors; if the boundary sample has a true label, its label is propagated to the nearest neighbors; if the boundary sample label is "to be propagated", propagation is skipped. The predicted labels of high-confidence samples are used as pseudo-labels; the propagated labels are merged with the high-confidence pseudo-labels to construct a labeled training set for the target domain, where the source of the sample labels is marked as "high confidence" or "propagation"; With the feature extractor parameters fixed, only the linear and batch normalization layers of the classifier are fine-tuned; Using the merged target domain labeled training set, the classifier parameters are updated through the cross-entropy loss function (formula: L = -Σ_{i=1}^N y_i log(p_i)) and the Adam optimizer (learning rate = 1e-4, β1 = 0.9, β2 = 0.999); The fine-tuned model is evaluated on the target domain test set to calculate the accuracy of human behavior recognition.

[0064] In practical applications, the method proposed in this invention is tested using two publicly available datasets commonly used in the field.

[0065] 1) The first dataset used is the Daily and Sports Activities Dataset (DSAD), and the second dataset is the Physical Activity Monitoring for Aging People Dataset (PAMAP2).

[0066] 2) Methods used for comparison include baseline methods and fine-tuning methods. Baseline methods involve obtaining a pre-trained model on the source domain training set and then directly applying it to the test set of new individuals in the target domain. The pre-trained model in baseline methods has never learned from data on new individuals; therefore, this method is used to directly test the cross-individual generalization ability of the pre-trained model.

[0067] Fine-tuning: Fine-tuning a pre-trained model is a common transfer learning method and is considered an upper bound on the model's performance on new individuals in the target domain. This method consists of a pre-training phase and a fine-tuning phase. The purpose of pre-training is to allow the model to learn the knowledge required for the task from existing data, so that it can achieve better results more quickly when performing similar tasks again. In the fine-tuning phase, some parameters of the existing model are retrained with a new dataset to adapt it to the new dataset. Fine-tuning after pre-training can improve the performance of the pre-trained model on new tasks.

[0068] The specific steps of the fine-tuning method are as follows: First, construct the network structure and randomly initialize the parameters.

[0069] Second, the model is trained on the source domain training set during the pre-training phase. When training on the DSAD dataset, the hyperparameters are set as follows: training epochs of 30, batch size of 512, and learning rate of 0.00001. When training on the PAMAP2 dataset, the hyperparameters are set as follows: training epochs of 30, batch size of 256, and learning rate of 0.00001.

[0070] Third, the fine-tuning stage: After obtaining the pre-trained source model, the parameters of its feature extraction part are fixed and no longer updated. The classifier part is randomly initialized and updated on the target domain training set. When fine-tuning on the DSAD dataset, the hyperparameter settings are as follows: training epochs of 50, batch size of 512, and learning rate of 0.00000001. When fine-tuning on the PAMAP2 dataset, the hyperparameter settings are as follows: training epochs of 50, batch size of 512, and learning rate of 0.00000001.

[0071] Fourth, the testing phase: After fine-tuning using the target domain training set, the recognition accuracy is tested using the test set of the target individuals.

[0072] 3) The test results are shown in Table 1 and Table 2.

[0073] Table 1: Test results of the DSAD dataset

[0074] Based on the results in the table above, when performing leave-one-out trials on the DSAD dataset, the present invention outperforms the baseline and approaches the upper limit fine-tuning method on all target individuals.

[0075] Table 2: Test results of the PAMAP2 dataset

[0076] Based on the results in the table above, when the leave-one-out-of-subjects test was performed on the PAMAP2 dataset, the present invention outperformed the baseline and approached the effect of the fine-tuning method on all target individuals.

[0077] In summary, this invention proposes a cross-individual human behavior recognition method utilizing confidence thresholding, sparse queries, and label propagation. The neural network structure design fully considers the spatiotemporal characteristics of wearable data, thereby enabling better automatic feature extraction from sensor time-series data. The cross-individual adaptation algorithm proposed in this invention directly filters high-confidence samples in the target domain through confidence thresholding and assigns labels to uncertain samples at class boundaries using sparse queries and label propagation. Compared to re-collecting sufficient data and retraining the model for individuals in the target domain, the algorithm in this invention, through optimizing the neural network structure and cross-individual adaptation algorithm, significantly reduces the annotation cost required for adapting to new individuals, alleviates the burden on users, improves the model's behavior recognition accuracy in cross-individual scenarios, and reduces annotation costs and development burden, demonstrating broad application prospects and market value.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A cross-individual human behavior recognition method based on self-training and active inquiry, characterized in that, include: Wearable sensors are fitted to the first user group. When the first user group performs the required behavior activities, the wearable sensors are used to synchronously collect the acceleration data of the corresponding body parts during the movement, and a labeled dataset of the source domain is constructed. At the same time, the synchronous collection process is repeated for the second user group to obtain an unlabeled dataset of the target domain. Data processing is performed on acceleration data in the source domain labeled dataset and the target domain unlabeled dataset, including detecting and processing missing values ​​in the data, and performing sliding window sampling on the processed data; The processed target domain data is divided into unlabeled training and test sets according to a certain ratio; A feature extractor based on a dual-channel convolutional network is built to extract features in both time and space dimensions from the input two-dimensional acceleration data. A cross-individual adaptation algorithm based on confidence threshold, sparse query and label propagation is adopted to train the model on the unlabeled training set in the target domain and alleviate the adaptation problem caused by the difference in cross-individual data distribution. The cross-individual adaptation algorithm employs a joint optimization strategy based on confidence threshold, sparse query, and label propagation. Specifically, it achieves adaptive adjustment of the cross-domain human behavior recognition model through the following steps: Train a human behavior recognition model on a source domain labeled dataset, and save the parameters of the feature extractor and classifier. The feature extractor is composed of a dual-channel convolutional network, and the classifier is composed of a linear layer and a batch normalization layer. Input the unlabeled training set samples of the target domain into the pre-trained model, and save the sample feature vector output by the feature extractor, the predicted label output by the classifier, and the confidence vector. For each sports category, samples are sorted in descending order of confidence and the top 10% are selected as high-confidence samples. For each sports category, the mean vector of the high-confidence sample feature vector is calculated as the feature center of that category. Calculate the Euclidean distance from all samples in the target domain training set to the center of each motion class; For each sample, determine its nearest class center c_1 and second nearest class center c_2, and calculate the distance ratio r = d(x, c_1) / d(x, c_2); If r is close to 1, the sample is considered a boundary sample, which may have classification uncertainty.

2. The method according to claim 1, characterized in that: When performing missing value detection, the moving difference method is used to detect abnormal missing points in the collected acceleration time series data. That is, when the sampling time interval between consecutive data points exceeds a preset threshold, it is determined that there is a missing value at that position. At the same time, the detection of sudden changes in data amplitude is used to help identify discontinuous missing values ​​caused by transmission interruption. For the detected missing values, a time series-based bidirectional linear interpolation method is used, which uses the acceleration values ​​of N normal data points before and after the missing point to perform weighted average interpolation, with the weights decreasing exponentially with time distance; if the missing value is located at the beginning or end of the data sequence or the consecutive missing length exceeds M points, the nearest neighbor imputation method is used, which replaces the missing value with the nearest valid data point value. The interpolated data needs to pass a consistency check, i.e. the sum of the absolute values of the second-order differences between the interpolated section and the original data is ≤ 0.5 m / s 4 , otherwise it is marked as abnormal and triggers manual review.

3. The method according to claim 2, characterized in that: When performing sliding window sampling, the sliding window length is set to 300ms, the sliding step size is set to 30ms, and the ratio of window length to step size is 10:

1. The interpolated acceleration time series data is truncated in units of sliding windows, and the data in each window are arranged in chronological order as a one-dimensional vector. If the data at the end of the window is less than 300ms, it is padded to the window length by zero padding or mirror padding, but the padded segment is marked and excluded in subsequent feature extraction. For the multi-sensor data within each sliding window, they are concatenated into a two-dimensional array according to the sensor dimension, where each row represents a sliding window sample and each column represents a data channel [ax, ay, az] of a sensor.

4. The method according to claim 3, characterized in that: For the dataset sampled by the sliding window, data diversity is enhanced in the following ways: Random Gaussian noise was added to simulate sensor noise; Random time shifting to simulate sampling time jitter; The data-augmented samples need to pass validity verification, meaning the acceleration amplitude of the samples must be within the range of -10 m / s². 2 10m / s 2 Within the range, and the frequency domain energy distribution conforms to the characteristics of human movement; The generated source domain labeled dataset and target domain unlabeled dataset each contain the following information: Source domain samples containing sliding window data and motion category labels; Target domain sample containing sliding window data.

5. The method according to claim 1, characterized in that: The feature extractor based on a dual-channel convolutional network achieves joint extraction of temporal and spatial features through the following design: Upper channel: Contains N levels of convolutional modules, each module consisting of a one-dimensional convolutional layer, a batch normalization layer, and a non-linear activation function layer; the one-dimensional convolutional layer slides along the time dimension, extracting local dynamic features in the time series in parallel through multiple convolutional kernels; the number of output channels of each level of convolutional module increases exponentially to abstract higher-order features in the time dimension layer by layer; The lower-level channels contain M levels of convolutional modules. Each module consists of a two-dimensional convolutional layer, a batch normalization layer, and a non-linear activation function layer. The two-dimensional convolutional layer slides along the sensor dimension and extracts the spatial correlation features between multiple sensor data in parallel through multiple convolutional kernels. The number of output channels of each level of convolutional module increases synchronously with the upper-level channels to ensure the alignment of temporal and spatial features.

6. The method according to claim 5, characterized in that: When extracting temporal features, the upper channel performs temporal convolution on the input two-dimensional data, with a dimension of [L×C], where L is the time step and C is the number of sensors. The output feature map has a dimension of [L'×F_t], where L' is the length of the time dimension after compression and F_t is the number of temporal feature channels. When extracting spatial features, the lower-level channels perform spatial dimension convolution on the same input data, and the output feature map dimension is [L×F_s], where F_s is the number of spatial feature channels; the spatial features are aggregated along the time dimension through the global average pooling layer, and the output dimension is [1×F_s], so as to extract the global spatial correlation pattern between sensors; The temporal feature map [L'×F_t] is flattened into a one-dimensional vector dimension [L'×F_t]. The spatial feature vector [1×F_s] is expanded to the same length by copying and then concatenated. The final fused feature dimension is [L'×F_t + F_s]. The concatenated feature vector is then subjected to dimensionality reduction and nonlinear mapping through a fully connected layer to form a joint spatiotemporal feature representation. The output feature map of the upper channel is transposed, i.e., the dimension is transformed [L'×F_t]→[F_t×L'] and multiplied point by point with the intermediate feature map of the lower channel, i.e. the dimension is [L×F_s], to achieve cross-modal fusion of time-space features; the output joint spatiotemporal feature vector is input into the classifier for motion category prediction.

7. The method according to claim 6, characterized in that, The classifier includes: Linear layer: The input is the joint spatiotemporal feature vector output by the dual-channel convolutional network, with a dimension of [F_final]. The output dimension is the total number of motion categories K. The linear mapping from features to categories is achieved through the learnable weight matrix W∈R^{F_final×K} and the bias vector b∈R^K. The weights of the linear layer are initialized with a Xavier uniform distribution, ranging from [-√(6 / (F_final+K)),√(6 / (F_final+K))], to alleviate the gradient vanishing problem. Batch standardization layer: Immediately following the linear layer, the output features are normalized, and the mean μ and variance σ are calculated. 2 It achieves adaptive adjustment through learnable parameters γ and β; the batch normalization formula is: y = γ ×(x - μ) / sqrt(σ 2 +ε) + β, where ε=1e-5 is the numerical stability constant, where γ is the scaling factor and β is the translation factor.

8. The method according to claim 7, characterized in that: Before the linear layer, i.e. between the output of the feature extractor and the input of the linear layer, the ReLU activation function is used to perform a nonlinear transformation on the joint spatiotemporal feature vector, with the formula f(x)=max(0, x); the ReLU function sparsifies the feature representation. After the batch normalization layer, the Softmax function is applied to the output vector to convert the original scores into a probability distribution, with the formula: P(y=k|x) = exp(z_k) / Σ_{i=1}^K exp(z_i), where z_k is the k-th element of the batch normalization layer output; the probability vector output by the Softmax function has a dimension of [K], and each element represents the predicted probability that the input sample belongs to the corresponding motion category, and satisfies Σ_{k=1}^KP(y=k|x)=1; The classifier's final output is a probability vector [p_1, p_2, ..., p_K], where p_k is the predicted probability of the k-th class of motion; the predicted class label is determined by the argmax operation, i.e., y_pred = argmax_k(p_k); The cross-entropy loss function is used to measure the difference between the predicted probability and the true label. The formula is: L = -Σ_{k=1}^K y_k log(p_k), where y_k is the one-hot encoding of the true label.

9. The method according to any one of claims 1 to 8, characterized in that: When performing sparse label queries, for every two groups of motion categories, the top 10 samples with the largest distance ratio r are selected as boundary samples; Query the true labels of these boundary samples from the label library of the target domain training set; For each boundary sample, search the feature space for the top 3% of samples with the closest Euclidean distance as its nearest neighbors; if the boundary sample has a true label, propagate its label to its nearest neighbors; if the boundary sample's label is "to be propagated", skip the propagation. The predicted labels of high-confidence samples are used as pseudo-labels; the propagated labels are merged with the high-confidence pseudo-labels to construct a labeled training set for the target domain, where the source of the sample labels is marked as high-confidence or propagation. With the feature extractor parameters fixed, only the linear and batch normalization layers of the classifier are fine-tuned; Using the merged target domain labeled training set, update the classifier parameters with the Adam optimizer through the cross-entropy loss function; The fine-tuned model is evaluated on the target domain test set to calculate the accuracy of human behavior recognition.

Images