A human behavior recognition method and system based on DualBiNet and a storage medium

Abstract

The application discloses a human behavior recognition method and system based on a DualBiNet and a storage medium, relates to the technical field of deep learning, and aims to solve the problems of the existing human behavior recognition method using channel state information, such as being easily affected by noise and irrelevant features, high computational complexity or large resource consumption, information leakage, and difficulty in accurately capturing subtle changes in human behavior. The application comprises the following steps: S1, collecting a human behavior CSI dataset and preprocessing CSI data; S2, constructing a DualBiNet model, wherein the DualBiNet model comprises a bidirectional BiTCAN module and a BiLSTM module; S3, inputting human behavior CSI data into the DualBiNet model, performing feature extraction through the BiTCAN module, capturing long-term dependence of the extracted features through the BiLSTM, weighting the output of the BiLSTM module through an attention module, and splicing the output of the BiLSTM module and the output of the attention module through residual connection; and finally, integrating features processed through multiple layers through a full connection layer, and classifying human behaviors.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a human behavior recognition method, system, and storage medium based on DualBiNet. Background Technology

[0002] When WiFi signals encounter a human body during propagation, phenomena such as reflection and scattering occur, altering the signal's propagation. These changes can reflect the dynamic behavior of objects and people in the environment. In recent years, utilizing WiFi signals for environmental perception and human behavior recognition has gradually become a hot research topic.

[0003] In the field of WiFi signal-based sensing technology, two commonly used parameters are Received Signal Strength Indicator (RSSI) and Channel State Information (CSI). RSSI was widely used in early research due to its ease of acquisition and simple processing; however, with the advancement of technology and the demand for high-precision sensing, CSI, which contains richer information, has received increasing attention.

[0004] CSI is extremely sensitive to environmental changes and can capture subtle physiological signals such as breathing and heartbeat. However, this characteristic also makes CSI data susceptible to noise, resulting in outliers and irrelevant features.

[0005] Traditional convolutional neural networks (CNNs) use channel state information for human behavior recognition. However, they are susceptible to noise and irrelevant features, resulting in high computational complexity and resource consumption. Furthermore, they often suffer from information leakage and are not suitable for processing time-series data. This is because the model may introduce future information during convolution operations, leading to a lack of causality and ultimately making it difficult to accurately capture subtle changes in human behavior, thus affecting the accuracy and robustness of the recognition. Summary of the Invention

[0006] The technical problem to be solved by this invention is:

[0007] Existing methods for human behavior recognition that utilize channel state information are susceptible to noise and irrelevant features, have high computational complexity or high resource consumption, and often suffer from information leakage, making it difficult to accurately capture subtle changes in human behavior.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] This invention provides a human behavior recognition method based on DualBiNet, comprising the following steps:

[0010] S1. Collect human behavior CSI dataset and preprocess the CSI data;

[0011] S2. Construct a DualBiNet model, which includes a bidirectional BiTCAN module and a BiLSTM module. The BiTCAN module uses 1D-CBAM, which adjusts two-dimensional convolution and two-dimensional pooling into one-dimensional convolution and one-dimensional pooling to obtain the 1D-CBAM. The 1D-CBAM is introduced into the TCN to replace the standard pooling layer, and the convolution operation of the TCN is replaced with bidirectional convolution to capture the forward and backward information of the input data simultaneously.

[0012] S3. Input human behavior CSI data into the DualBiNet model, extract features through the BiTCAN module, capture long-term dependencies of the extracted features through BiLSTM, weight the output of the BiLSTM module using an attention module, and concatenate the output of the BiLSTM module with the output of the attention module using a residual connection; finally, integrate the multi-layered processed features through a fully connected layer to classify human behavior.

[0013] Furthermore, the preprocessing of CSI data described in S1 includes filtering the data using Hampel filtering and Butterworth filtering.

[0014] Furthermore, the attention module described in S3 comprises two linear layers, a tanh activation function, and a softmax function.

[0015] Furthermore, the DualBiNet model described in S3 uses NLLoss as the loss function.

[0016] Furthermore, in S3, an early stopping method is used during the training of the DualBiNet model. The training is stopped when the performance metrics on the validation set are monitored in real time.

[0017] This invention provides a human behavior recognition system based on DualBiNet. The system has a program module corresponding to the steps of any of the above-described technical solutions, and executes the steps of the above-described human behavior recognition method based on DualBiNet when running.

[0018] The present invention provides a computer-readable storage medium storing a computer program configured to implement the steps of the human behavior recognition method based on DualBiNet as described in any of the above technical solutions when invoked by a processor.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] This invention proposes a composite classification model based on the bidirectional temporal convolutional attention network BiTCAN and the long short-term memory network Bi-LSTM, employing a one-dimensional convolutional attention module 1D-CBAM. By combining the attention mechanism with a residual network, it achieves a deep fusion of the attention mechanism and the convolutional neural network, significantly improving the model's performance in processing one-dimensional data with only a small increase in computational resource overhead. This allows the model to more accurately capture subtle changes in human behavior, thereby greatly improving the accuracy of human behavior recognition. The model was tested on two public datasets, UT-HAR and NTU-HAR, achieving recognition accuracies of 99.03% and 98.47%, respectively. This demonstrates the superior recognition performance of the proposed model on different datasets, providing new ideas and methods for research in related fields. Attached Figure Description

[0021] Figure 1 This is a flowchart of the human behavior recognition method based on DualBiNet in an embodiment of the present invention.

[0022] Figure 2 This is a diagram showing the data preprocessing effect in an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the DualBiNet model structure in an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram of the internal structure of BiTCAN in an embodiment of the present invention;

[0025] Figure 5 This is a schematic diagram of the 1D-CBAM module structure in an embodiment of the present invention;

[0026] Figure 6 This is a schematic diagram of bidirectional convolution in an embodiment of the present invention;

[0027] Figure 7 This is a schematic diagram of the attention module structure in an embodiment of the present invention;

[0028] Figure 8 This is a confusion matrix diagram in an embodiment of the present invention;

[0029] Figure 9 This is a comparison chart of the accuracy of DualBiNet in this embodiment of the invention and existing CNNs;

[0030] Figure 10 This is a comparison chart of the recognition effects of different algorithms in the embodiments of the present invention;

[0031] Figure 11This is a comparison chart of the performance of the DualBiNet model in this embodiment of the invention on two datasets. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] This invention provides a human behavior recognition method based on DualBiNet, such as... Figure 1 As shown, it includes the following steps:

[0035] S1. Collect human behavior CSI dataset and preprocess the CSI data;

[0036] S2, construct as follows Figure 3 The DualBiNet model shown includes a bidirectional BiTCAN module and a BiLSTM module. The BiTCAN module uses 1D-CBAM. The 1D-CBAM adjusts the convolution, pooling, concatenation, and weight calculation operations that are originally suitable for two-dimensional data to obtain the 1D-CBAM. The 1D-CBAM is introduced into the TCN to replace the standard pooling layer, and the convolution operation of the TCN is replaced with bidirectional convolution to capture the forward and backward information of the input data simultaneously.

[0037] S3. Input human behavior CSI data into the DualBiNet model, extract features through the BiTCAN module, capture long-term dependencies of the extracted features through BiLSTM, weight the output of the BiLSTM module using an attention module, and concatenate the output of the BiLSTM module with the output of the attention module using a residual connection; finally, integrate the multi-layered processed features through a fully connected layer to classify human behavior.

[0038] In this embodiment, the BiTCAN module is a Bidirectional Temporal Convolutional Attention Network (TCAN). First, the traditional Convolutional Block Attention Module (CBAM) is improved and transformed into a 1D-Convolutional Block Attention Module (1D-CBAM), enabling its proper application to one-dimensional signal processing. This 1D-CBAM module is then introduced into the TCN (Temporal Convolutional Network), replacing the standard pooling layer. This achieves a deep integration of the attention mechanism and the convolutional neural network, significantly improving the model's performance when processing one-dimensional data with only a small increase in computational resource overhead. The improved TCN model is called the Temporal Convolutional Attention Network (TCAN).

[0039] This embodiment, based on the TCAN model, further optimizes the convolution operation by replacing it with bidirectional convolution to simultaneously capture both forward and backward information from the input data. This significantly enhances the model's ability to perceive temporal features, resulting in the following: Figure 4 The example shown is the BiTCAN bidirectional convolutional attention network. This new framework not only improves the efficiency of feature extraction but also provides stronger support for processing complex one-dimensional signals.

[0040] TCN, or Temporal Convolutional Network, is a deep learning model used to process sequential data. TCN mainly captures temporal information in time series through a series of one-dimensional convolutional layers, and introduces causal convolution, dilated convolution, and residual connections on the basis of one-dimensional CNN. It is often used for tasks such as time series prediction and sequence classification.

[0041] TCN employs causal convolution, which appropriately pads the input to ensure that the convolution operation does not introduce future information, enhancing the model's ability to analyze time series data. The introduction of dilated convolution allows the model to capture longer dependencies in time series without increasing parameters, improving the ability to recognize complex sequence patterns. In addition, TCN also includes residual connections, which add the input before convolution and the output after convolution, enabling the model to effectively transfer information between different layers, alleviate gradient vanishing, and retain important features, allowing the model to learn complex sequence features more stably.

[0042] 1D-CBAM achieves this by adjusting operations such as convolution, pooling, concatenation, and weight calculation, which were originally suitable for two-dimensional data, to adapt them to one-dimensional structures, thus realizing... Figure 5 The 1D-CBAM module is shown. This module introduces channel attention and spatial attention mechanisms, enabling the model to adaptively focus on important parts of the input features, significantly improving the convolutional neural network's ability to recognize key features and important locations.

[0043] This attention mechanism not only enhances the overall interpretability of the model but also makes the features and locations it focuses on more transparent. By assigning higher weights to important features, 1D-CBAM can effectively mitigate the influence of noise and irrelevant features, improving the model's robustness in noisy and outlier data environments. Furthermore, 1D-CBAM primarily relies on global pooling and simple convolution operations, thus having low computational overhead and not significantly increasing the model's computational complexity or resource consumption, providing a good balance for practical applications.

[0044] In traditional TCNs, the main function of pooling layers is to extract important features and reduce feature size. This embodiment replaces traditional pooling layers with 1D-CBAM, introducing an attention mechanism to adaptively select important features, rather than relying solely on simple max pooling or average pooling, thus significantly improving feature extraction capabilities. Simultaneously, due to the smaller original size of the input data, 1D-CBAM not only maintains the stability of feature size but also reduces information loss, preventing feature maps from shrinking prematurely to a point where information cannot be effectively extracted. Therefore, the Convolutional Attention Network (TCAN) using 1D-CBAM to replace pooling layers demonstrates significant advantages over traditional TCNs in human behavior recognition tasks.

[0045] Traditional convolutions, processing only forward or backward information, can struggle to understand the global context, especially with long sequences where dependencies between preceding and following information are difficult to capture fully. Furthermore, unidirectional convolutions are highly sensitive to local noise or data loss, potentially leading to performance degradation under unstable input conditions. The core idea of ​​bidirectional convolution is to flip the input data during a regular convolution process, followed by a reverse convolution. This means that when processing the input sequence, the network considers not only its forward information but also its backward information. Finally, the results of the two convolutions are concatenated to form the output of the bidirectional convolution, as shown below. Figure 6 As shown.

[0046] The dual information processing mechanism in this embodiment enables the network to extract features from both forward and backward directions simultaneously, thereby achieving a more comprehensive and richer feature representation. By integrating information from different directions, the network can more effectively focus on key features, improving the model's generalization ability. Furthermore, since bidirectional convolution utilizes more information sources, the model is more resistant to the influence of local noise or missing data, thus enhancing the model's robustness.

[0047] BiLSTM (Bidirectional Long Short-Term Memory Network) is an improved recurrent neural network (RNN) for processing sequential data, such as in natural language processing and time series analysis. BiLSTM combines the advantages of traditional LSTM with enhanced contextual understanding through its bidirectional structure.

[0048] This embodiment combines BiTCAN and BiLSTM to effectively improve the performance of sequence data processing. BiTCAN can quickly identify key patterns and trends in input sequences, making it suitable for tasks with significant local properties, such as time series and text data. This structure not only accelerates feature extraction but also provides more refined feature representations for subsequent sequence modeling. BiLSTM, with its unique bidirectional structure, effectively integrates contextual information, enhancing the model's understanding of the context and thus improving the accuracy of predicting complex data patterns. This architecture fully leverages the advantages of both: first, BiTCAN is used for efficient local feature extraction, generating simplified and expressive feature representations; then, BiLSTM is used for in-depth analysis of the extracted features to capture long-term dependencies. The combination of BiTCAN and BiLSTM achieves efficient fusion of feature extraction and contextual modeling, improving overall efficiency and reducing the amount of data that BiLSTM needs to process, allowing it to focus more on capturing complex temporal dependencies. This strategy not only improves model performance but also provides a new perspective for the application of deep learning in sequence data analysis. The model in this embodiment exhibits low computational complexity when processing data, demonstrating its superior balance between speed and performance. This innovation not only provides a new solution for sequence data analysis, but also points the way for future related research and expands the boundaries of applications in the field of deep learning.

[0049] Since the attention module involves weight calculations, the matrix processed by this module may lose some useful information. To address this issue, this embodiment employs a residual connection strategy, concatenating the output of the Bidirectional Long Short-Term Memory (BiLSTM) network with the output of the attention module after refining the data. Residual connections introduce shortcut connections between layers, allowing gradients to propagate directly within the network, effectively mitigating the vanishing and exploding gradient problems. This not only enhances the model's training stability but also provides a direct gradient propagation path, enabling inputs to bypass the attention module and directly reach the output. This significantly reduces the potential negative impact of the attention module, further improving the model's expressive power and overall performance, allowing the network to better capture and utilize information, thus achieving superior performance in action recognition tasks.

[0050] Preferably, the preprocessing of CSI data in S1 includes filtering the data using Hampel filtering and Butterworth filtering.

[0051] This embodiment employs Hampel filtering to detect and remove outliers in the data. Hampel filtering compares each data point within a signal window to the median of that window. If the difference between a data point and the median exceeds a set threshold, that point is considered an outlier and is replaced with the median. Figure 2 The waveform changes before and after applying Hampel filtering are shown. Outliers in the signal are effectively removed, and large jumps no longer occur in the data, showing a smoother characteristic and significantly improving the reliability and usability of the data.

[0052] Typically, human activity occurs at low frequencies, while noise and other components in acquired CSI data occur at high frequencies. By applying a Butterworth low-pass filter, high-frequency components can be effectively suppressed, thus preserving low-frequency components and significantly improving data quality and reliability. The Butterworth filtering process is as follows:

[0053]

[0054] Where is the order of the filter. c This is the cutoff frequency.

[0055] like Figure 2 As shown, after Butterworth low-pass filtering, it is clear that the noise in the signal has been effectively removed. The two-step filtering process improves the smoothness and stability of the signal, laying a solid foundation for subsequent analysis.

[0056] Preferably, the attention module comprises two linear layers, a tanh activation function, and a softmax function; its structure is as follows: Figure 7 As shown. For the output of BiLSTM The output h at each time step t The query vector will be obtained after a linear transformation and a Tanh activation function. t :

[0057] Q t =tanh(W q h t )#(19)

[0058] in, It is the weight matrix of the query layer.

[0059] Next, the query is mapped to the score through another linear layer. t :

[0060] S t =W s Q t #(20)

[0061] in, It is the weight matrix of the scoring layer.

[0062] Then, the scores for all time steps. t Apply the softmax function to obtain attention weights. t :

[0063]

[0064] Finally, the context vector C is computed by multiplying the attention weights by the output of the BiLSTM.

[0065]

[0066] Attention mechanisms provide a rich information foundation for subsequent processing, not only optimizing the feature extraction process but also greatly improving the model's adaptability and robustness in dynamic scenarios, thus endowing complex behavior recognition tasks with stronger analytical capabilities.

[0067] Preferably, the DualBiNet model uses NLLoss as the loss function.

[0068] In the final stage of this embodiment, features processed through multiple layers are integrated via a fully connected layer and then input into a LogSoftmax layer for normalization. The LogSoftmax layer is typically used in conjunction with the Negative Log-Likelihood Loss (NLLLoss) function, offering greater stability. Furthermore, since the gradient of the logarithmic function is relatively smooth, the combination of LogSoftmax and NLLoss facilitates the efficient propagation of gradients within the network, significantly improving training efficiency. This not only optimizes the model's learning process but also fosters a closer connection between feature extraction and classification results, providing a more accurate and reliable output for the final behavior recognition.

[0069] Preferably, in S3, an early stopping method is used during the training of the DualBiNet model, and the timing of stopping training is determined by real-time monitoring of performance metrics on the validation set.

[0070] The human behavior recognition method (algorithm) based on DualBiNet proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0071] Based on the method proposed in this invention, a human behavior recognition system based on DualBiNet is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the above-mentioned human behavior recognition method based on DualBiNet when running.

[0072] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the DualBiNet-based human behavior recognition method described above when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0073] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0074] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0075] Example 1

[0076] This embodiment first uses the publicly available dataset UT-HAR for systematic testing, which includes seven actions: lying down, falling, picking up, running, sitting down, standing up, and walking. The CSI size is 1x250x90. This dataset is divided into training, validation, and test sets in an 8:1:1 ratio. During the DualBiNet model training process, an early stopping mechanism is introduced. The timing of training is determined by real-time monitoring of performance metrics on the validation set. This effectively avoids unnecessary training processes, saving computational resources and significantly improving the model's generalization ability, preventing overfitting. Some parameter settings are shown in Table 1.

[0077] Table 1

[0078]

[0079] After employing early stopping, the model automatically stopped training after approximately 100 epochs, at which point the classification accuracy reached 99.03%. Figure 8 The confusion matrix obtained after testing is shown, clearly demonstrating the performance of the DualBiNet model in various action recognition tasks. This result not only shows that the model has excellent classification capabilities, but also further verifies the effectiveness of the proposed method, providing strong support for action recognition in practical applications.

[0080] As can be seen, the DualBiNet model demonstrates strong recognition capabilities for seven categories of actions, with the recognition accuracy for five actions—falling, walking, picking up, running, and standing up—reaching 100%. This fully demonstrates the model's outstanding performance on the UT-HAR dataset.

[0081] To further verify the improvement brought by BiTCAN compared to conventional CNNs, this embodiment replaced the BiTCAN module in the model with a traditional CNN and conducted comparative experiments. The results are as follows: Figure 9 As shown, the performance differences between the two in behavior recognition tasks are illustrated.

[0082] Conventional CNNs no longer use the CBAM module, retain pooling layers for further feature extraction, and reduce the size of the convolutional kernels to avoid losing effective information. Results show that using BiTCAN significantly improves the model's performance.

[0083] To more intuitively demonstrate the superiority of this model, this embodiment also compares it with several commonly used classification algorithms, including Multi-Layer Perceptron (MLP), Recurrent Neural Network (RNN), Bidirectional Long Short-Term Memory (BiLSTM), and Convolutional Neural Network (CNN). The results are as follows: Figure 10 As shown.

[0084] Furthermore, this embodiment adjusted the model's parameters, including input / output dimensions and the number of convolutional layer channels, to adapt to the CSI information of the NTU-HAR dataset, and then trained and tested it again on this dataset. The test results are now compared with those on UT-HAR. Figure 11 As shown, the DualBiNet model exhibits an accuracy of over 98% on both datasets. Notably, on the NTU-HAR dataset, the relatively simple BiLSTM also achieves similar results. This is primarily due to the dataset's ease of recognition, containing six actions: running, walking, falling, boxing, crossing arms, and cleaning the floor. These actions are more distinguishable, resulting in more pronounced differences in the shape of the CSI signals. The NTU-HAR dataset utilizes different systems during acquisition, with each antenna having 114 subcarriers, providing more representative feature information. Furthermore, its significantly larger size compared to UT-HAR significantly increases the amount of data available for training.

[0085] Driven by numerous factors, BiLSTM alone is sufficient to achieve highly accurate recognition for the NTU-HAR dataset. Although DualBiNet can uncover deeper features, its recognition capabilities do not show a particularly significant advantage.

[0086] Nevertheless, considering that in real-world applications, it's impossible to guarantee that the collected data will always possess the same high quality and discriminative power as NTU-HAR, the DualBiNet model demonstrates a significant advantage when facing relatively difficult recognition tasks. Furthermore, the model achieves excellent recognition results on two significantly different datasets, proving its superior generalization ability and strong robustness.

[0087] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A human behavior recognition method based on DualBiNet, characterized in that, Includes the following steps: S1. Collect human behavior CSI dataset, and preprocess the CSI data. CSI stands for Channel State Information. S2. Construct a DualBiNet model, which includes a bidirectional BiTCAN module and a BiLSTM module. The BiTCAN module uses 1D-CBAM, which adjusts two-dimensional convolution and two-dimensional pooling into one-dimensional convolution and one-dimensional pooling to obtain the 1D-CBAM. The 1D-CBAM is introduced into the TCN to replace the standard pooling layer, and the convolution operation of the TCN is replaced with bidirectional convolution to capture the forward and backward information of the input data simultaneously. S3. Input human behavior CSI data into the DualBiNet model, extract features through the BiTCAN module, capture long-term dependencies of the extracted features through BiLSTM, weight the output of the BiLSTM module using the attention module, and concatenate the output of the BiLSTM module with the output of the attention module using residual connection. Finally, the features processed through multiple layers are integrated by a fully connected layer to classify human behavior.

2. The human behavior recognition method based on DualBiNet according to claim 1, characterized in that, The preprocessing of CSI data described in S1 includes filtering the data using Hampel filtering and Butterworth filtering.

3. The human behavior recognition method based on DualBiNet according to claim 2, characterized in that, The attention module described in S3 consists of two linear layers, a tanh activation function, and a softmax function.

4. The human behavior recognition method based on DualBiNet according to claim 3, characterized in that, The DualBiNet model described in S3 uses NLLoss as the loss function.

5. The human behavior recognition method based on DualBiNet according to claim 4, characterized in that, In S3, the DualBiNet model is trained using an early stopping method, which determines when to stop training by monitoring the performance metrics on the validation set in real time.

6. A human behavior recognition system based on DualBiNet, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 5, and executes the steps in the human behavior recognition method based on DualBiNet described above when it is run.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the human behavior recognition method based on DualBiNet as described in any one of claims 1 to 5.

Images