© Wuhan University 2026

0 Introduction

Ubiquitous sensing technology is a promising research frontier in the IoT sensing domain and has broad application prospects in areas such as security, sports monitoring, health monitoring, and entertainment ^[1].Within this landscape, human activity recognition (HAR) is one of the key research topic^[2-3]. Traditional HAR relies on vision-based systems or wearable sensors, which face limitations: vision requires line-of-sight and raises privacy concerns, while wearables impose user compliance challenges. In contrast, ubiquitous WiFi infrastructure enables device-free, non-intrusive activity sensing through wireless signal analysis. In particular, Channel State Information (CSI) in WiFi signals has shown great potential in a variety of device-free human perception tasks, such as activity recognition^[4-5], fall detection^[6], gesture recognition ^[7], human identity recognition ^[8-9], and human counting ^[10]. WiFi CSI-based HAR offers distinct advantages: device-free operation, non-line-of-sight capability, privacy preservation, illumination independence, and cost-effectiveness.

Currently, there are two main types of WiFi-based sensing methods: model-based and learning-based. Model-based methods, representing early phases of development, primarily rely on physical signal propagation models. These evolved from early approaches using Received Signal Strength Indicator (RSSI)^[11-12] to more precise methodologies leveraging CSI ^[13-16]. However,the coarse granularity of RSSI and the complexity of physical modeling for intricate human movements limit their scalability.

Conversely, learning-based methods leverage deep neural networks to extract features directly from data, achieving superior performance in complex sensing tasks. Despite their success, current deep learning approaches for HAR still face significant limitations: 1) LSTM-based models^[17] that effectively capture temporal dependencies but neglect spatial relationships and joint spatial-temporal features, limiting discriminative capability; 2) CNN (Convolutional Neural Network)-based methods^[18-19] that extract spatial features from CSI-transformed images yet suffer from shallow architectures lacking adaptive multi-scale attention; 3) CNN-LSTM hybrids ^[20-22] that jointly model spatial-temporal features but lack explicit multi-scale attention and adaptive fusion; 4) Domain-specific attention mechanisms ^[23-24] that improve robustness through temporal or separable dimensional attention but miss multi-scale spatial receptive fields; 5) Transformer-based approaches ^[25-27] that capture long-range dependencies at the cost of high computational complexity and environmental sensitivity. To address these limitations,we draw inspiration from the transformative success of attention mechanisms in NLP (Natural Language Processing)^[28-30],computer vision^[31-34], and time-series analysis^[35-39].While attention has shown emerging potential in WiFi sensing^[40-42],standard mechanisms often fail to address the specific domain challenges of CSI data, such as efficiency, adaptability to signal variability, and the need for multi-scale spatial receptive fields^[43-44]. Therefore, beyond merely applying existing attention models, there is a critical need for a domain-specific innovation. This motivates our development of an Enhanced CBAM framework, explicitly tailored to overcome both the inherent limitations of prior HAR methods and the specific shortcomings of conventional attention mechanisms in robust and efficient CSI-based recognition.

In this paper, we propose a novel CNN-ECBAM approach, a learning-based framework specifically designed to overcome the aforementioned challenges. By integrating an Enhanced Convolutional Block Attention Module (ECBAM) into a convolutional neural network, our method effectively captures discriminative spatiotemporal patterns from CSI data. The proposed approach consists of three main components: CSI-to-pseudo-color image conversion, Enhanced CBAM architecture, and integrated CNN framework.The main contributions of this paper are as follows:

1) Novel CSI-to-image Conversion Methodology: A systematic approach is introduced for converting raw CSI signals into pseudo-color images that preserve essential signal characteristics while enabling effective CNN processing. The method incorporates optimized normalization, color mapping, and resolution adaptation techniques specifically designed for WiFi CSI data.

2) ECBAM Architecture for CSI Data: An Enhanced Convolutional Block Attention Module is developed for pseudo-color images derived from WiFi CSI signals. The design integrates standard deviation pooling in channel attention and dilated convolutions in spatial attention, facilitating the capture of both signal variability and multiscale spatial patterns.

3) Comprehensive CNN-ECBAM Framework: A unified CNN architecture is constructed by integrating the ECBAM modules, forming the CNN-ECBAM framework. This design leverages the spatial feature extraction capability of CNNs together with the refined attention mechanisms of ECBAM to achieve superior activity recognition performance.

The remainder of this paper is organized as follows. Section 1 briefly introduces CSI, CBAM, and CNN. Section 2 presents the detailed methodology of our CNN-ECBAM approach, including the CSI-to-pseudo-color image conversion process, ECBAM architecture design, and integrated CNN framework. Section 3 describes the experimental setup, datasets, evaluation metrics, training strategy used in our study. Section 4 presents the experimental results and comprehensive analysis. Section 5 concludes with a summary of contributions and practical implications.

1 Preliminaries and Background

1.1 Channel State Information (CSI)

Wireless signal propagation is sensitive to environmental factors, particularly human motion, which induces reflection, refraction, scattering, and diffraction. In a fixed environment, distinct movements produce characteristic temporal fluctuations in the received channel response. By learning the statistical mapping between these fluctuations and the underlying motion states, human activities can be inferred from variations in WiFi Channel State Information (CSI). Figure 1 presents the overall framework of WiFi CSI-based activity recognition, comprising signal acquisition, processing, and classification stages.

Fig.1 The overview of human activity recognition

The wireless channel can be expressed in the frequency domain as follows:

$\mathit{Y}=\mathit{H}\cdot \mathit{X}+\mathit{N}$(1)

where H is the channel state information matrix, Y is the received signal vector, X is the transmitted signal vector, $\cdot$ denotes matrix multiplication, and N is the additive Gaussian white noise vector. H can be further expressed as:

$\mathit{H}(j)=|H(j)|{\mathrm{e}}^{\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{\angle }H(j)}$(2)

Here $H(j)$ denotes the channel state information of the j-th subcarrier, which contains two components, CSI magnitude ($|H(j)|$) and phase ($\mathrm{\angle }H(j)$). According to multiple-input multiple-output (MIMO) technology, the combined CSI of the data streams obtained by combining multiple antennas at the transceiver end with each other can be expressed with matrix form as:

$\mathit{H}=\left[\begin{array}{ccc}{H}_{\mathrm{11}}& \cdots & H_{\mathrm{1}m}\\ ⋮& H_{ij}& ⋮\\ H_{n\mathrm{1}}& \cdots & H_{nm}\end{array}\right]$(3)

where $H_{ij}$ denotes the CSI between the i-th receiving antenna and the j-th transmitting antenna, and $m$ and $n$ denote the number of antennas at the receiving and transmitting ends, respectively.

1.2 Convolutional Block Attention Module(CBAM)

CBAM^[34] enhances CNN feature representations through cascaded channel and spatial attention mechanisms, as shown in Fig. 2. Channel attention weights feature maps based on inter-channel relationships, while spatial attention emphasizes informative regions within each feature map, jointly improving network discriminability.

Fig. 2 The overview of CBAM

Given an original feature map $F\in {\mathbb{R}}^{C\times H\times W}$ as input, CBAM sequentially derives a one-dimensional channel attention map $M_c\in {\mathbb{R}}^{C\times \mathrm{1}\times \mathrm{1}}$ and a two-dimensional spatial attention map $M_s\in {\mathbb{R}}^{\mathrm{1}\times H\times W}$. The one-dimensional channel attention map $M_c$ will be multiplied channel by channel with the original feature map to obtain the channel-weighted feature map $F^{\text{'}}$, $F^{\text{'}}$ will be multiplied position by position with the two-dimensional spatial attention map $M_s$ to obtain the final output. The overall attention process can be summarized as

$\begin{array}{l}{F}^{\text{'}}=M_c(F)\otimes F,\\ F^{″}=M_s(F^{\text{'}})\otimes F^{\text{'}},\end{array}$(4)

where $\otimes$ denotes element-by-element multiplication.

For the Channel Attention Module (CAM) in Fig.2, its goal is to learn the importance weights of each channel to enhance useful channels and suppress redundant channels. Its realization steps include three phases: feature compression, weight generation and feature re-calibration. Feature compression performs Global Average Pooling (GAP) and Global Maximum Pooling (GMP) on the input feature map $F\in {\mathbb{R}}^{C\times H\times W}$ to obtain two $C\times \mathrm{1}\times \mathrm{1}$ channel description vectors, respectively. Weight generation takes the two vector inputs into the shared multilayer perceptron (MLP) and generates the channel attention weights $M_c\in {\mathbb{R}}^{C\times \mathrm{1}\times \mathrm{1}}$. The formula is described as follows:

$M_c(F)=\sigma (\mathrm{M}\mathrm{L}\mathrm{P}(\mathrm{G}\mathrm{A}\mathrm{P}(F))+\mathrm{M}\mathrm{L}\mathrm{P}(\mathrm{G}\mathrm{M}\mathrm{P}(F)))$(5)

where $\sigma$ is the Sigmoid function. The feature re-calibration multiplies $M_{\mathrm{c}}$ with the original feature map channel by channel to get the channel-weighted feature map. The spatial attention module learns the importance of each spatial location within feature maps, thereby focusing on key regions. Its implementation comprises four sequential stages: channel compression, feature concatenation, weight generation, and feature re-calibration. Channel compression performs average pooling and maximum pooling on the weighted feature map along the channel dimensions to obtain two $\mathrm{1}\times \mathit{H}\times \mathit{W}$ feature maps. Feature splicing splices the two feature maps into $\mathrm{2}\times \mathit{H}\times \mathit{W}$. Weight generation compresses the channels to 1 by a $\mathrm{7}\times \mathrm{7}$ convolutional layer to generate spatial attention weights $M_s\in {\mathbb{R}}^{\mathrm{1}\times H\times W}$. The formula is as follows:

$M_s(F)=\sigma (f^{\mathrm{7}\times \mathrm{7}}([\mathrm{A}\mathrm{v}\mathrm{g}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F);\mathrm{M}\mathrm{a}\mathrm{x}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F)]))$(6)

Feature re-calibration: Multiply $M_s$ with the feature map position by position to get the final output.

1.3 Convolutional Neural Networks

The convolutional neural network was first introduced for handwritten ZIP code recognition ^[45] and was later formalized in their comprehensive work on document recognition^[46]. These networks overcome the limitations of multilayer perceptrons (MLPs) via two key mechanisms: parameter-sharing convolutional filters and hierarchical spatial subsampling. The core computational framework of CNNs involves layered feature extraction through learnable convolution operators followed by non-linear downsampling. Mathematically, the discrete convolution operation applies a filter kernel to an input tensor through sliding window operations, generating feature maps via element-wise multiplication and summation.Spatial pooling layers subsequently reduce feature dimensionality by aggregating local regions through maximization (max-pooling) or averaging operations. A standard CNN architecture comprises alternating convolutional blocks, pooling layers, and fully connected classifiers, as illustrated in Fig. 3.

Fig.3 Main CNN architecture

2 Proposed Method

2.1 CSI-to-Pseudo-Color Image Preprocessing

The conversion of raw CSI signals into pseudo-color images involves several critical preprocessing steps designed to preserve essential signal characteristics while enabling effective CNN processing.This section provides detailed specifications of each preprocessing component.

2.1.1 Data normalization and scaling

The raw CSI amplitude data shows considerable variability across environments, devices, and time periods. To enhance training consistency and stability, a comprehensive normalization strategy is employed.

1) Global Min-Max Normalization: For the entire dataset $\mathcal{D}=\{X_{\mathrm{1}},X_{\mathrm{2}},\cdots ,X_N\}$, where each $X_i$ represents a CSI sample, we compute:

$X_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}}=\frac{X-\mathrm{m}\mathrm{i}\mathrm{n}(\mathcal{D})}{\mathrm{m}\mathrm{a}\mathrm{x}(\mathcal{D})-\mathrm{m}\mathrm{i}\mathrm{n}(\mathcal{D})}$(7)

where $\mathrm{m}\mathrm{i}\mathrm{n}(\mathcal{D})$ and $\mathrm{m}\mathrm{a}\mathrm{x}(\mathcal{D})$ are the global minimum and maximum values across all samples in the dataset.

2) Pixel Value Mapping: The normalized values are then mapped to the standard 8-bit pixel intensity range:

$P=\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}(X_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}}\times \mathrm{255})$(8)

where $P\in [\mathrm{0,255}]$ represents the pixel intensity values, and the data type is converted to uint8 for memory efficiency and compatibility with image processing frameworks.

2.1.2 Color mapping and visualization

To convert single-channel normalized CSI data into informative pseudo-color representations, the viridis colormap is utilized. This colormap is widely adopted owing to its desirable properties: (i) perceptual uniformity, ensuring that equal data variations correspond to equal perceptual differences; (ii) monotonic luminance, providing a consistent brightness progression with increasing values; and (iii) accessibility for individuals with color vision deficiencies, while maintaining high contrast across the entire dynamic range. Formally, the mapping function translates normalized intensity values into RGB triplets, thereby enhancing the interpretability of CSI visualizations.

2.1.3 Image standardization and signal quality validation

To ensure consistent input dimensions for the CNN architecture, pseudo-color images are generated under standardized settings. Each image is initially rendered at 8×8 inches with 100 DPI and subsequently rescaled to 64×64 pixels for computational efficiency. The images are represented in three RGB channels and stored as float32 tensors. To preserve the intrinsic structure of CSI data, the antenna and subcarrier dimensions are mapped to the vertical and horizontal axes, respectively, while temporal variations are encoded through color intensity.

Signal integrity during preprocessing is ensured through validation of dynamic range preservation, noise floor distinguishability, and retention of activity-specific signatures. This pipeline effectively transforms raw CSI measurements into discriminative pseudo-color images, which are then used as input to the CNN-ECBAM model. Representative samples are presented in Fig. 4 and Fig. 5.

Fig. 4 UT_HAR generated RGB images

Fig.5 NTU-Fi_HAR generated RGB images

2.2 Enhanced CBAM

The core contribution of our work lies in the Enhanced Convolutional Block Attention Module (ECBAM), which significantly improves upon the original CBAM by incorporating several innovative mechanisms specifically designed for pseudo-color images derived from WiFi CSI data. Our Enhanced CBAM addresses the unique characteristics of CSI signals by integrating four key innovations: (1) an Efficient Channel Attention (ECA) mechanism, (2) Multi-scale Spatial Attention with parallel convolutions, (3) an Adaptive Weight Learning mechanism for dynamic feature fusion, and (4) explicit residual connections for gradient flow optimization and feature preservation.Its detailed structure is shown in Fig. 6.

Fig.6 Enhanced CBAM architecture

2.2.1 Efficient Channel Attention (ECA) module

Traditional CBAM applies global average and max pooling for channel attention, but this design is limited in capturing the complex statistics of CSI-based pseudo-color images. In the proposed framework, the conventional module is replaced with Efficient Channel Attention (ECA), which achieves more effective channel-wise recalibration while preserving computational efficiency. ECA adaptively determines kernel sizes according to channel dimensions, thereby enhancing the modeling of cross-channel interactions. The adaptive kernel size is defined as:

$t=|\mathrm{l}\mathrm{o}{\mathrm{g}}_{\mathrm{2}}(c)+b|/\gamma$(9)

$k=\{\begin{array}{ll}t,& \mathrm{i}\mathrm{f}t\mathrm{i}\mathrm{s}\mathrm{o}\mathrm{d}\mathrm{d},\\ t+\mathrm{1},& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{s},\end{array}$(10)

where $c$ is the number of input channels, $\gamma =\mathrm{2}$ and $b=\mathrm{1}$ are hyperparameters, and $k$ is the final kernel size for the 1D convolution. Hyperparameters $\gamma =\mathrm{2}$ and $b=\mathrm{1}$ are adopted following the original ECA work^[47], which determined these values through extensive experiments on large-scale datasets. The choice of $\gamma =\mathrm{2}$ ensures logarithmic growth of the kernel sizes with respect to the channel dimensions, while $b=\mathrm{1}$ provides an appropriate base offset for effective modeling of cross-channel interactions. The ECA mechanism then applies the 1D convolution along the channel dimension after global average pooling:

$\mathrm{E}\mathrm{C}\mathrm{A}(F)=\sigma (\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\mathrm{1}{\mathrm{D}}_k(\mathrm{G}\mathrm{A}\mathrm{P}(\mathrm{F})))$(11)

where $\mathrm{G}\mathrm{A}\mathrm{P}$ denotes global average pooling, $\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\mathrm{1}{\mathrm{D}}_k$ represents 1D convolution with kernel size $k$, and $\sigma$ is the sigmoid activation function.

The mathematical formulation for channel attention refinement is:

${F^{\text{'}}}_c=F_c\otimes \mathrm{E}\mathrm{C}\mathrm{A}{(F)}_c$(12)

where $F_c$ represents the c-th channel of the input feature map $F$, and $\otimes$ denotes element-wise multiplication. $\mathrm{E}\mathrm{C}\mathrm{A}{(F)}_c$ is an importance weight computed for the c-th channel via global average pooling and 1D convolution. ${F^{\text{'}}}_c$ is the characterization of the channel after enhancement by the attentional mechanism, where important channels are reinforced and unimportant channels are suppressed.

2.2.2 Multi-scale spatial attention module

The spatial attention component of our ECBAM incorporates parallel convolutions with multiple kernel sizes to capture spatial patterns at different receptive fields. This design is particularly crucial for CSI-derived pseudo-color images, where spatial relationships manifest at various scales. The input to this module is generated by concatenating the channel-wise average-pooled and max-pooled features from the input tensor.This combined feature descriptor is then processed through three parallel convolution branches with 3×3, 5×5, and 7×7 kernels:

$\begin{array}{l}\mathrm{S}{\mathrm{A}}_{\mathrm{3}\times \mathrm{3}}(F)=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{3}\times \mathrm{3}}([\mathrm{A}\mathrm{v}\mathrm{g}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F);\mathrm{M}\mathrm{a}\mathrm{x}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F)]),(\mathrm{13})\\ \mathrm{S}{\mathrm{A}}_{\mathrm{5}\times \mathrm{5}}(F)=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{5}\times \mathrm{5}}([\mathrm{A}\mathrm{v}\mathrm{g}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F);\mathrm{M}\mathrm{a}\mathrm{x}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F)]),(\mathrm{14})\\ \mathrm{S}{\mathrm{A}}_{\mathrm{7}\times \mathrm{7}}(F)=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{7}\times \mathrm{7}}([\mathrm{A}\mathrm{v}\mathrm{g}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F);\mathrm{M}\mathrm{a}\mathrm{x}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}(F)]),(\mathrm{15})\end{array}$

where $[\bullet ;\bullet ]$ denotes concatenation along the channel dimension, and the input to each branch is formed by concatenating channel-wise average and max pooled features.

The multi-scale features are then fused through a $\mathrm{1}\times \mathrm{1}$ convolution layer:

$\mathrm{S}{\mathrm{A}}_{\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}}(F)=\sigma (\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{1}\times \mathrm{1}}([\mathrm{S}{\mathrm{A}}_{\mathrm{3}\times \mathrm{3}}(F);\mathrm{S}{\mathrm{A}}_{\mathrm{5}\times \mathrm{5}}(F);\mathrm{S}{\mathrm{A}}_{\mathrm{7}\times \mathrm{7}}(F)]))$(16)

The final spatial attention map is generated using a sigmoid activation and applied to the feature map:

$F_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}^{″}=F^{\text{'}}\otimes \mathrm{S}{\mathrm{A}}_{\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}}(F^{\text{'}})$(17)

$F_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}^{″}$ is a feature map enhanced by a spatial attention mechanism, where important spatial regions are reinforced and unimportant regions are suppressed. $F^{\text{'}}$ is the channel-enhanced feature map with c channels, serving as input to the spatial attention module. $\mathrm{S}{\mathrm{A}}_{\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}}(F^{\text{'}})$ is the attention weight map generated by the multi-scale spatial attention module, indicating the importance of spatial locations.

2.2.3 Adaptive weight learning mechanism

A critical innovation in our Enhanced CBAM is the introduction of learnable adaptive weights that dynamically balance the outputs of the attention pathways. This mechanism combines the channel-attended features with the final spatially-attended features. The adaptive weight learning mechanism is formulated as:

$W_{\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{l}},W_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}\in \mathbb{R}>\mathrm{0}, W_{\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{l}}(\mathrm{0})=W_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}(\mathrm{0})=\mathrm{1.0}$(18)

where the initialization values are set to 1.0. The normalized weights are computed via softmax normalization:

$\begin{array}{l}{w}_c=\frac{\mathrm{e}\mathrm{x}\mathrm{p}(W_{\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{l}})}{\mathrm{e}\mathrm{x}\mathrm{p}(W_{\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{l}})+\mathrm{e}\mathrm{x}\mathrm{p}(W_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}})}, \\ w_s=\frac{\mathrm{e}\mathrm{x}\mathrm{p}(W_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}})}{\mathrm{e}\mathrm{x}\mathrm{p}(W_{\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{l}})+\mathrm{e}\mathrm{x}\mathrm{p}(W_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}})}\end{array}$(19)

and the final enhanced attention output is

$F_{\mathrm{o}\mathrm{u}\mathrm{t}}=w_c{F}^{\text{'}}+w_s{F}_{\mathrm{s}\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}^{″}$(20)

2.2.4 Residual connection integration

To maintain feature preservation and facilitate gradient flow, our Enhanced CBAM is integrated within residual blocks. The residual connection ensures that the original feature information is preserved while allowing the attention mechanism to refine the features:

$F_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}}=F+\mathrm{E}\mathrm{n}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{d}\_\mathrm{C}\mathrm{B}\mathrm{A}\mathrm{M}(F).$(21)

$F_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}}$ denotes the final output feature map obtained after enhanced CBAM processing with residual connections. In contrast, $F$ represents the original input feature map of the Enhanced CBAM residual block, and $\mathrm{E}\mathrm{n}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{d}\_\mathrm{C}\mathrm{B}\mathrm{A}\mathrm{M}(F)$ indicates the feature map generated by the Enhanced CBAM module, which integrates both channel and spatial attention.

The residual connection design in our Enhanced CBAM serves two critical functions: (1) gradient flow preservation during backpropagation, which prevents vanishing gradient problems in deep networks, and (2) feature identity preservation, ensuring that the original input features are maintained while allowing the attention mechanism to provide refinements. As depicted in Fig. 6, the skip connection pathway directly adds the input feature map to the attention-processed output, creating a robust learning pathway that combines both original and attention-enhanced features.

2.3 CNN-ECBAM Architecture

The proposed CNN-ECBAM architecture is systematically constructed to optimize pseudo-color image processing through strategically integrated attention mechanisms. The comprehensive framework comprises hierarchically organized components that facilitate multi-scale feature learning and adaptive attention allocation. Figure 7 illustrates the architecture.

Fig.7 CNN-ECBAM architecture

1) Input Processing Module: Raw CSI amplitude data is converted into 64×64×3 RGB pseudo-color images, encoding spatial-temporal patterns in a format compatible with convolutional processing.

2) Hierarchical Convolutional Framework: The network employs alternating 3×3 and 5×5 convolutions to capture local and contextual features, with batch normalization for training stability,ReLU activation, strategically positioned ECBAM modules for adaptive attention, and dropout (p=0.3) for regularization.

3) Multi-Resolution Feature Learning: Parallel convolutional branches with varying receptive fields simultaneously extract fine-grained local features and broader contextual patterns from CSI data.

4) Adaptive Feature Aggregation: Multi-scale features are integrated through concatenation (preserving feature diversity) and element-wise summation (dimension-preserving aggregation).

5) Decision Layer Architecture: Global average pooling reduces spatial dimensions while preserving channel information, followed by fully connected layers with ReLU activation and dropout, culminating in a softmax layer for activity classification.

3 Experimental Setup

3.1 Datasets

We evaluate our method on two public CSI-based HAR datasets: UT-HAR ^[17] and NTU-Fi_HAR ^[48]. UT-HAR comprises seven activity classes: bend, fall, lie down, run, sit down, stand up, and walk, collected using an Intel 5300 NIC with three antenna pairs, each capturing 30 subcarriers. Data collection was conducted in a controlled indoor environment.

NTU-Fi_HAR was acquired via Atheros CSI Tool and encompasses six activity classes: box, circle, clean, fall, run, and walk, representing diverse movement intensities and complexities.

3.2 Evaluation Metrics

To comprehensively assess the efficacy of our proposed methodology, we evaluate model performance using standard classification metrics: accuracy and F₁-score. Accuracy measures overall classification correctness, while F₁-score provides a balanced measure particularly valuable for imbalanced datasets. Finally, we present a comprehensive confusion matrix analysis to reveal inter-class confusion patterns.

3.3 Baseline Methods and Comparison

We compare CNN-ECBAM against the following baseline methods:

1) Classical Deep Learning Architectures: We evaluate against a modified LeNet-5 ^[46] where the original activation functions have been replaced with ReLU units to enhance training efficiency, and ResNet-18^[49], which represents a standard convolutional neural network architecture widely adopted in computer vision tasks. We also incorporate the LSTM-based approach proposed in Ref. [17], which has shown effectiveness in capturing temporal dependencies for sequential CSI data in human activity recognition tasks.

2) WiFi-based HAR Specialized Methods: Our comparison includes ABLSTM ^[41], a prominent attention-based bidirectional LSTM approach specifically designed for WiFi sensing applications, and the CSI-to-image transformation method proposed in Ref.[19], which converts CSI data into Gramian Angular Field (GAF) representations followed by a four-layer CNN classifier for activity recognition.

To ensure the reliability of performance comparisons, paired t-tests are employed. Each model is trained and evaluated five times with different random seeds to derive performance distributions. Statistical significance is determined by comparing CNN-ECBAM against the strongest baseline on each dataset, where * indicates p < 0.05 and ** indicates p < 0.01.

3.4 Training Strategy

For all experiments, we train models for 200 epochs with adaptive batch size selection: LSTM models use batch size 64 for improved training stability, GAF-CNN follows the paper specification with batch size 32, while other models dynamically adjust batch sizes (32-200) based on GPU memory availability. The optimization strategy varies by model type: LSTM models on UT_HAR use Adam optimizer ($\mathrm{l}\mathrm{r}=\mathrm{0.001}$) with cosine annealing scheduling and gradient clipping ($\mathrm{m}\mathrm{a}\mathrm{x}\_\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}=\mathrm{1.0}$), while LSTM models on NTU-Fi_HAR follow paper specifications using SGD optimizer ($\mathrm{l}\mathrm{r}=\mathrm{0.001}$, $\mathrm{m}\mathrm{o}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{u}\mathrm{m}=\mathrm{0.9}$). GAF-CNN strictly adheres to paper requirements with Adam optimizer ($\mathrm{l}\mathrm{r}=\mathrm{0.001}$, $\mathrm{w}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\_\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{y}=\mathrm{0.000}\mathrm{1}$), and other models use Adam with Reduce LR On Plateau scheduling. We implemented mixed-precision training for memory-intensive models (CNN-ECBAM, ResNet-18, GAF-CNN) to optimize GPU utilization, employ gradient accumulation for large datasets, and apply Xavier normal initialization for convolutional layers with forgetting gate bias set to 1.0 for LSTM models. Additionally, we incorporated early stopping mechanism with $\mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}=\mathrm{10}$, dropout regularization (0.3-0.5), and systematic memory management including periodic cleanup to ensure stable training across all experimental configurations.

4 Results and Analysis

4.1 Overall Performance Comparison

Table 1 summarizes the classification accuracy and F₁-scores for all models across both datasets. The results demonstrate the distinct robustness of our proposed CNN-ECBAM architecture. On the UT-HAR dataset, CNN-ECBAM achieved the highest accuracy of 99.20%, outperforming the strong ResNet-18 baseline (98.60%) and significantly surpassing ABLSTM (94.40%) and LeNet-5 (94.20%). Notably, GAF-CNN failed to capture discriminative features, yielding only 68.00% accuracy.

Conversely, on the NTU-Fi_HAR dataset, CNN-ECBAM attained 100% accuracy, confirming its superior feature extraction capability. A critical finding is the performance volatility of baseline models: while GAF-CNN improved dramatically to 99.62% (indicating dataset-specific specialization), ResNet-18 degraded significantly to 87.50%, revealing severe overfitting. In contrast, CNN-ECBAM maintained state-of-the-art performance across both scenarios. Statistical significance testing further confirms these improvements: the gain over the strongest baselines is significant at p < 0.05 for UT-HAR and p < 0.01 for NTU-Fi_HAR, validating the effectiveness of the proposed attention-enhanced framework in overcoming the generalization limitations observed in conventional architectures.

4.2 Convergence Analysis for Different Methods

This section provides a comparative analysis of convergence behavior, final accuracy, and generalization capability across all models on the UT-HAR dataset and NTU-Fi_HAR dataset , as illustrated in Fig. 8 and Fig. 9.

Fig.8 Accuracy convergence curves for all evaluated models on the UT_HAR dataset Note:   Each subplot shows the training accuracy (blue) and test accuracy (orange) over 200 epochs.

Fig.9 Accuracy convergence curves for all evaluated models on the NTU-Fi_HAR dataset Note:   Each subplot shows the training accuracy (blue) and test accuracy (orange) over 200 epochs.

The proposed CNN-ECBAM demonstrated consistent stability and high accuracy across both datasets. On UT-HAR dataset, it converged to approximately 98% accuracy, matching the stability of ResNet-18 but with superior generalization. Notably, on the NTU-Fi_HAR dataset, where ResNet-18 exhibited severe overfitting (plateauing at 87% test accuracy despite 100% training accuracy), CNN-ECBAM maintained robust performance with minimal generalization gap. This contrast indicates that while standard deep architectures like ResNet-18 risk overfitting on specific CSI datasets, the domain-specific attention mechanisms in CNN-ECBAM effectively regularize the learning process. The Modified LeNet-5 achieved respectable stability but generally trailed the deeper architectures in peak accuracy.

A comparison between LSTM and ABLSTM highlights the critical role of attention mechanisms in temporal modeling. The standard LSTM struggled with volatility and lower accuracy (plateauing around 85% on UT-HAR), whereas ABLSTM achieved significantly more stable convergence (reaching 94% on UT-HAR and 97% on NTU-Fi_HAR). This confirms that attention mechanisms successfully mitigate the limitations of standard RNNs in capturing long-range dependencies within noisy CSI data.

The GAF-CNN model displayed significant performance variability. It failed to converge effectively on UT-HAR (fluctuating between 60%-70%) and exhibited extreme testing instability on NTU-Fi_HAR despite high training accuracy. This suggests that the GAF image representation may not consistently preserve the discriminative signal characteristics required for robust generalization across diverse activity sets.

In conclusion, while baseline models exhibited dataset-dependent strengths and weaknesses—often suffering from overfitting (ResNet-18) or instability (LSTM, GAF-CNN)—the CNN-ECBAM framework demonstrated better adaptability. By integrating efficient channel and multi-scale spatial attention, the model effectively balances model complexity with feature selection, ensuring robust convergence and generalization across diverse sensing environments.

4.3 Confusion Matrix for Different Methods

This section evaluates the inter-class discriminative capabilities of the models using confusion matrices for both datasets, as visualized in Fig. 10 and Fig. 11.

Fig.10 Confusion matrices for all evaluated models on the UT_HAR dataset

Fig.11 Confusion matrices for all evaluated models on the NTU-Fi_HAR dataset

On the UT-HAR dataset, the proposed CNN-ECBAM demonstrated superior feature extraction, achieving near-perfect classification for distinct activities (fall, bend, run, sit down) and maintaining high precision (>95%) even for challenging transitional actions (e.g., stand up vs. walk). ResNet-18 also performed strongly but exhibited minor confusion between specific postures like standup and lie down. In contrast, ABLSTM and modified LeNet-5 struggled more significantly with dynamic transitions, while LSTM and GAF-CNN showed widespread misclassification, indicating a failure to capture discriminative features for complex human activities.

Conversely, on the NTU-Fi_HAR dataset, CNN-ECBAM achieved perfect classification across all six activities (box, circle, clean, fall, run, walk). Notably, GAF-CNN, which performed poorly on UT-HAR, achieved high accuracy here, suggesting its effectiveness is highly dataset-dependent. However, ResNet-18 failed to generalize, showing significant confusion across multiple classes (e.g., misclassify walk as box or clean), which aligns with the overfitting observed in the convergence analysis.

Qualitative examination of misclassified samples reveals that errors predominantly occur at motion boundaries where activity transitions are ambiguous (e.g., the initiation of standup resembling walk). While CNN-ECBAM significantly reduces these errors through its multi-scale attention mechanism, the persistence of minor boundary confusions suggests that incorporating extended temporal context could further enhance performance in future work.

4.4 Ablation Studies

Systematic ablation studies were conducted across eight configurations to validate the contributions of individual Enhanced CBAM components: 1) CNN without attention, 2) original CBAM(baseline), 3) Efficient Channel Attention (ECA) only, 4) Multi-Scale Spatial Attention (MSSA) only, 5) ECA combined with MSSA, 6) ECA with adaptive weights, 7) MSSA with adaptive weights, and 8) the full Enhanced CBAM framework. This progressive evaluation isolates the impact of each innovation, assessing their individual and synergistic effects on recognition performance. Table 2 details the performance metrics for each configuration.

The ECA mechanism yields a 0.2% accuracy gain on UT-HAR while reducing parameters by 4.3% compared with conventional MLP-based channel attention, highlighting its parameter efficiency. Multi-scale spatial attention adds 0.3%, underscoring the importance of capturing spatial patterns across multiple receptive field sizes, while learnable fusion weights further enhance performance by dynamically balancing channel and spatial attention. Overall, the enhanced CBAM achieves a 0.6% improvement over the standard CBAM, with synergistic rather than purely additive effects. All improvements are statistically significant (p < 0.05) across 10 independent runs.

5 Conclusion

A novel framework, CNN-ECBAM, is proposed for WiFi-based human activity recognition to address the inherent challenges of CSI-based sensing. First, a CSI-to-pseudo-color image conversion method is developed to preserve essential signal characteristics for effective CNN processing. Second, an enhanced CBAM architecture is introduced with improved channel and spatial attention mechanisms tailored to CSI data. Third, an integrated CNN framework is constructed in which these components are seamlessly combined. Extensive experiments on two public datasets demonstrate that CNN-ECBAM achieves competitive performance compared with existing methods, showing consistent accuracy improvements on both UT-HAR and NTU-Fi_HAR datasets. The enhanced attention mechanisms allow more effective capture of complex CSI patterns, resulting in higher recognition accuracy and robustness, thereby supporting deployment in real-world scenarios such as smart homes, healthcare monitoring, and building management.

Future work will focus on the design of lightweight architectures for edge deployment, enhancement of cross-domain generalization, and integration of multi-modal sensing to further improve recognition performance in diverse environments.

References

All Tables

All Figures

	Fig.1 The overview of human activity recognition
In the text

	Fig. 2 The overview of CBAM
In the text

	Fig.3 Main CNN architecture
In the text

	Fig. 4 UT_HAR generated RGB images
In the text

	Fig.5 NTU-Fi_HAR generated RGB images
In the text

	Fig.6 Enhanced CBAM architecture
In the text

	Fig.7 CNN-ECBAM architecture
In the text

	Fig.8 Accuracy convergence curves for all evaluated models on the UT_HAR dataset Note:   Each subplot shows the training accuracy (blue) and test accuracy (orange) over 200 epochs.
In the text

	Fig.9 Accuracy convergence curves for all evaluated models on the NTU-Fi_HAR dataset Note:   Each subplot shows the training accuracy (blue) and test accuracy (orange) over 200 epochs.
In the text

	Fig.10 Confusion matrices for all evaluated models on the UT_HAR dataset
In the text

	Fig.11 Confusion matrices for all evaluated models on the NTU-Fi_HAR dataset
In the text