Real-Time, Non-Intrusive Fall Detection via Wi-Fi CSI: A Comparative Study of CNN-LSTM, GNN, and Transformer Models

Research Gaps

A comprehensive analysis of previous studies in CSI-based HAR reveals several persistent limitations that hinder their scalability and real-world applicability. One major concern is the prevalent use of small, non-representative datasets, which heightens the risk of overfitting and undermines the models’ ability to generalize across diverse home layouts, occupant behaviors, and environmental conditions. For example, Lowe et al. [12], Moshiri et al. [14], and Forbes et al. [15] relied on datasets ranging from 420 to 1,100 samples, which limited the robustness of their models. Zhuravchak et al. [13] demonstrated that even state-of-the-art architectures like InceptionTime could only achieve limited accuracy when trained on small-scale data. Recently, Hu et al. [17] introduced a mobile Wi-Fi-CSI receiver system for construction workers, yet their dataset was limited to a small-scale simulated site, raising questions about scalability and generalizability. Youm and Go [18] explored lightweight Neural Architecture Search (NAS)-pruned models, but real-time performance in dynamic, uncontrolled environments remains untested. Bayad et al. [19] proposed a 2D-CNN system that generalizes to unseen environments and new participants, but their evaluation was still limited to a constrained indoor dataset with specific environmental variations, highlighting the need for broader, multi-site validation.

This issue is further compounded by the lack of evaluation in heterogeneous spatial settings, which ignores the variations introduced by different building materials, furniture arrangements, and human-induced dynamics, all of which affect the propagation of the Wi-Fi signal. Studies such as Yan et al. [16], Forbes et al. [15], Hu et al. [17], Youm and Go [18], and Bayad et al. [19] validated their systems in only one or two environments, limiting the generalizability of their models across different deployment scenarios.

Another critical gap lies in the insufficient preprocessing of raw CSI data. Many studies, including those by Moshiri et al. [14] and Lowe et al. [12], directly utilize unfiltered amplitude or phase information, which is often noisy and susceptible to phase shifts, multipath effects, and hardware-related artifacts. Without proper cleaning and transformation of the signal, these inconsistencies degrade the performance of the model and contribute to poor reproducibility. Even recent mobile receiver, NAS-pruned, and adaptive 2D-CNN models [17–19] often rely on preprocessed datasets or offline analysis, limiting insights into end-to-end real-time deployment.

Equally significant is the gap in real-time system deployment. Despite the growing need for continuous and passive monitoring, especially in applications such as eldercare, most proposed systems are validated only under controlled lab conditions, with no integration into real-time pipelines. For example, although Yan et al. [16] used a relatively large dataset, the lack of real-time classification limited the system’s utility for real-world applications. Similar limitations in online processing were observed in the works of Zhuravchak et al. [13], Forbes et al. [15], and Moshiri et al. [14]. Recent studies [20, 21] have advanced representation learning and cross-domain adaptation in HAR. However, significant research gaps remain: real-time HAR under robust Wi-Fi CSI preprocessing, end-to-end deployment in dynamic or unseen environments, and integration of large-scale self-supervised pretraining for practical scenarios are still largely unsolved problems. Youm and Go [18] improved efficiency via NAS-pruning, but real-world tests were not conducted. Bayad et al. [19] demonstrated adaptability to unseen environments and individuals but primarily in small-scale indoor settings, leaving large-scale deployment unexamined.

In summary, existing studies frequently suffer from: (1) limited and environment-specific datasets [12, 14, 15, 17–21]; (2) inadequate preprocessing of CSI signals [12, 14, 17–19]; (3) absence of real-time activity recognition [13, 16–21]; and (4) narrow experimental validation in realistic settings [15–19]. These gaps point to the urgent need for holistic solutions that combine diverse data collection, rigorous signal preprocessing, and real-time deep learning pipelines to ensure robust and scalable performance.

Our study directly addresses these challenges by incorporating a more diverse and realistic dataset, applying tailored preprocessing techniques to refine raw CSI data, and developing an end-to-end deep learning model capable of real-time activity classification. Compared to prior work—including mobile receiver systems [13], NAS-pruned lightweight models [18], and adaptive 2D-CNNs for unseen environments [19]—our approach integrates real-time deployment, robust preprocessing, and evaluation across multiple environments, demonstrating both practical applicability and improved generalizability in Wi-Fi-based HAR. This approach aims to bridge the gap between laboratory-based experimentation and real-world deployment, ultimately improving the reliability and practicality of Wi-Fi-based HAR systems.

Our Contribution

This research presents a practical and robust approach to real-time fall detection using Wi-Fi-based HAR, addressing limitations in prior studies [12, 15, 18–22] that often rely on offline analysis, preprocessed datasets, simulations, or expensive hardware. Unlike these works, our system directly processes raw CSI in realistic indoor settings, achieves end-to-end real-time performance, and is validated on low-cost, widely available hardware. Each design choice emphasizes practicality, robustness, and reproducibility for real-world deployment.

• Low-cost, real-time HAR system: uses only an off-the-shelf Wi-Fi router and Raspberry Pi 4, eliminating the need for expensive sensors or specialized hardware. Previous studies [17, 19] often require multiple antennas, mobile devices, or costly experimental setups, limiting accessibility for everyday applications.

• End-to-end solution: integrates data collection, CSI preprocessing, model training, and deployment in one seamless pipeline. Many works [12, 21, 22] rely on offline datasets or simulations, making them unsuitable for real-time monitoring.

• Beacon-frame CSI acquisition: uses only passive Wi-Fi beacon frames to generate CSI, avoiding interference with existing network traffic. Other approaches often require active transmissions or additional devices [15, 18], which increases complexity and deployment cost.

• Robust dataset and preprocessing workflow: captures both Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) scenarios and uses subcarrier selection, Hampel and Savitzky–Golay (SG) filtering, and normalization to improve signal quality. Prior studies [20, 21] often rely on preprocessed or idealized datasets, limiting generalization to realistic indoor environments.

• Validated lightweight deep learning architectures (CNN-LSTM, 2D-CNN): CNN-LSTM learns spatial CSI patterns and temporal dependencies simultaneously, which is essential for complex activity recognition. Transformer and GNN approaches [18, 19] are often unverified on real CSI data, raising doubts about practical utility.

• Sliding window mechanism: overlapping segments improve prediction smoothness and reduce errors during activity transitions. Most related works do not explicitly handle temporal overlap, which can lead to misclassification at activity boundaries [6, 23].

• Feature enhancement with PCA: reduces dimensionality and improves feature clarity, particularly for activities with similar CSI patterns (e.g., falling vs. lying). Many previous methods ignore architecture-specific preprocessing [13, 17], leading to lower accuracy under NLOS conditions.

• Continuous real-time pipeline: automates packet capture, conversion, preprocessing, and classification with Python threading and watchdog tools to maintain low-latency operation. Other approaches typically focus only on model evaluation offline, without demonstrating full real-time deployment [12, 20].

• Foundation for practical applications: the validated, low-cost, and end-to-end tested system enables deployment in smart homes, healthcare, and assisted living environments. Prior works provide algorithmic or simulation results [21, 22], but rarely demonstrate practical applicability in realistic indoor settings.

Mathematical Formulation of CSI

CSI captures the frequency response of a wireless channel, describing how transmitted signals are affected by multipath propagation and environmental changes, including human movements. Let the transmitted signal be , the received signal can be expressed as

where is the channel impulse response (CIR), denotes convolution, and is additive noise. The CIR can be further represented in discrete form across subcarriers as

where and denote the amplitude and phase of the -th subcarrier, respectively. CSI captures these parameters, reflecting the effects of multipath propagation, attenuation, and changes in the environment, such as human motion. By monitoring variations in CSI over time, human activities can be inferred in a non-intrusive manner, forming the basis of Wi-Fi-based HAR systems.

System Architecture and CSI Acquisition Setup

We develop a practical and cost-effective Wi-Fi-based HAR system using a Raspberry Pi 4, a standard Wi-Fi router (TP-Link Archer AX23), and a laptop. This setup balances affordability, portability, ease of deployment, and sufficient processing capability, making it suitable for real-world applications.

The Raspberry Pi acts as the CSI receiver. Its built-in Broadcom Wi-Fi chip (BCM43455C0), compact design, and General-Purpose Input/Output (GPIO) support make it ideal for edge sensing in constrained environments. Since the default operating system lacks native CSI access, we install DragonOS Pi64 (kernel 5.4.0-1033-raspi), a Linux distribution tailored for signal processing, and integrate Nexmon, an open-source firmware patch that enables CSI extraction by modifying Wi-Fi drivers. Nexmon also allows beacon frame filtering (frame control ), and Media Access Control (MAC)-based capture, ensuring only relevant, high-quality data is collected.

The TP-Link router is configured to operate on the 5 GHz band, channel 36, with an 80 MHz bandwidth, broadcasting beacon frames at a fixed 50 ms interval (~20 Frames Per Second (FPS)). This setup ensures regular, passive CSI sampling without generating additional traffic, minimizing interference and simplifying deployment.

Each CSI segment comprises 100 consecutive beacon frames (~5 seconds), capturing both transient and steady-state features of human activities. This window length balances responsiveness and pattern richness for classification.

Data are offloaded to a laptop for processing and model training, leveraging its superior computational power. This edge-backend separation improves real-time performance, reduces latency, and allows the use of more complex models without overloading the embedded hardware.

Overall, the hybrid architecture—lightweight at the edge, powerful at the backend—provides a scalable and flexible solution for CSI-based HAR, aligning with the growing need for affordable and deployable sensing systems in smart home and healthcare environments.

Data Preprocessing

To ensure the reliability and effectiveness of machine learning models trained on raw CSI data acquired from Raspberry Pi devices, rigorous preprocessing is essential. Raw CSI measurements are inherently noisy and may contain irrelevant or redundant information that can negatively impact classification accuracy. Through a structured preprocessing pipeline, the CSI data can be transformed into a stable and informative representation, thereby enhancing the robustness and predictive performance of subsequent models.

Classification

Classifying human activities from CSI data is a vital part of building a real-time HAR system. After collecting and preprocessing the data, we explore different deep learning models to identify the best trade-off between accuracy and real-time performance for practical use. We focus on the following three architectures: CNN-LSTM, Graph Neural Networks (GNN), and Transformers, because each offers unique strengths tailored to the complex spatio-temporal nature of Wi-Fi CSI signals.

CNN-LSTM models combine two powerful components: convolutional layers and recurrent units. The convolutional layers act as spatial feature extractors, scanning the CSI matrix to detect local patterns across subcarriers, which are crucial for identifying characteristic signal changes caused by different human activities. The LSTM layers model the temporal dependencies by remembering sequences of extracted features over time, capturing how activities unfold dynamically. This combination makes CNN-LSTM highly suitable for our study, as it balances spatial detail with temporal context, which is essential for recognizing complex movements such as falls or transitions between actions.

By contrast, GNNs treat CSI data as a graph structure, where nodes represent different signal points and edges reflect their relationships. This framework allows GNNs to learn spatial dependencies in a flexible non-Euclidean domain, which is especially useful for capturing subtle interactions in multipath and cluttered environments typical of indoor Wi-Fi signals. The message-passing and aggregation mechanisms of GNNs help the model adapt to complex topologies and spatial correlations that traditional CNNs might miss, making them a promising choice for nuanced activity patterns in realistic settings.

Transformers rely on a self-attention mechanism that computes weighted dependencies between all elements in a sequence simultaneously. This architecture excels at modeling long-range temporal relationships without the limitations of sequential memory in RNNs or LSTMs. Transformers’ multi-head attention enables the model to focus on different parts of the CSI sequence in parallel, efficiently extracting relevant features for activity classification. Their ability to handle high-dimensional, sequential data with global context makes Transformers an attractive candidate for HAR, especially as activity signals often involve long and complex temporal patterns.

By studying these models, we aim to benchmark their classification accuracy and computational efficiency, identifying the architecture that best fits real-time HAR with Wi-Fi CSI. This thorough comparison highlights how each model’s internal mechanisms align with the unique challenges of CSI data—spatial complexity, temporal dynamics, and noise robustness—and help us select the most practical and effective approach for deployment in real-world scenarios.

Equipment Setup

The implementation of the HAR system required a carefully designed setup on both the transmitter and receiver sides to enable the acquisition of high-quality CSI data. On the transmitter side, a Wi-Fi router was configured to operate in the 5 GHz frequency band, specifically using channel 36 with an 80 MHz bandwidth. The Beacon Interval was set to 50 milliseconds to ensure consistent and periodic transmission of control frames.

On the receiver side, a Raspberry Pi 4 was employed and installed with the Nexmon firmware to enable low-level access for monitoring Wi-Fi signals and capturing CSI data. The setup began by generating a CSI parameter string using the following command:

makecsiparams -c 36/80 -C 1 -N 1 -m {router MAC address} -b 0 × 80

This command configured the system to collect CSI data on channel 36 with 80 MHz bandwidth, utilizing the first processing core and spatial stream of the Raspberry Pi. It also ensured the capture of Beacon frames by specifying the identifier and filtering packets based on the router’s MAC address.

To ensure that the system operated exclusively in monitor mode and remained unaffected by standard client-side processes, the wpa_supplicant service—responsible for managing Wi-Fi connections—was terminated with the following command:

pkill wpa_supplicant

Subsequently, the wireless interface was reactivated to confirm its operational state:

ifconfig wlan0 up

Next, the Nexutil tool was configured to initiate CSI extraction using the generated parameter string:

nexutil -Iwlan0 -s500 -b -l34 –v{parameters from makecsiparams}

To facilitate passive Wi-Fi monitoring, a dedicated monitor interface was created with the following commands:

iw phy $(iw dev wlan0 info | gawk’/wiphy/ {printf“phy”$2}’)

interface add mon0 type monitor

ifconfig mon0 up

This complete configuration enabled the Raspberry Pi to continuously monitor the wireless medium in passive mode, thereby allowing uninterrupted and high-fidelity collection of CSI data without interference from active client communication.

Experimental Setup

The CSI dataset used in this study was collected in two distinct environments designed to simulate various real-world conditions for HAR. Figure 2 provides the floor plans of these environments, showing the layout of the transmitter, receiver, and the paths along which human activities were performed. The selection of these environments was intentional, aiming to capture a wide range of activity types in different spatial configurations. By employing distinct environments, the dataset represents diverse scenarios and challenging the HAR system to generalize well across varying environmental conditions.

Two environmental setups were used in this study: (1) the same-room setup and (2) the different-room setup. In the first setup, the environments depicted in Figure 2 were selected to capture the nuances of human activity in both controlled and dynamic conditions. This configuration places both the Wi-Fi router and the Raspberry Pi device in the same room. Within this setup, we conducted two distinct scenarios: one with a clear LOS between the devices, and another with an NLOS condition introduced by placing obstacles between them. The setup also considers factors such as signal strength, potential interference, and the presence of obstacles—all of which can significantly impact the quality and consistency of the collected CSI data. These elements are essential for evaluating the robustness of the HAR system and its ability to classify activities accurately in real time.

Figure 3 illustrates the floor plan of the second setup used for data collection. In this configuration, the Wi-Fi router was placed in one room, while the Raspberry Pi device responsible for collecting CSI was located in a separate room where the test subject performed various activities. This arrangement represents an extreme NLOS scenario, designed to emulate realistic situations commonly found in daily life, for example, when a Wi-Fi router is installed in a living room or bedroom and the user is in another area such as a bathroom. Such a setup allows for evaluating the system’s capability to detect human presence or gestures even when the LOS is completely obstructed, making it highly applicable for non-intrusive activity monitoring in smart home environments.

Data Collection

We continuously captured CSI data in batches of 100 frames using a Python script that repeatedly executed the following command:

tcpdump -i wlan0 dst port 5500 -c 100

The collected CSI data was represented as complex numbers. With an 80 MHz bandwidth, each CSI frame consisted of 256 subcarriers. The number of CSI samples collected for each activity class is shown in Table 2.

Data Preprocessing

The preprocessing of the raw CSI data involved several stages of filtering and transformation, as illustrated in Figure 4. Initially, we removed null, pilot, and low-value subcarriers to retain only informative parts of the signal. Next, we applied the Hampel filter with a Gaussian scaling factor of 1.4826, a window length of five samples, and a threshold of three standard deviations. This process helped to eliminate outliers while preserving the signal structure. The result is shown in Figure 5. Subsequently, we applied the SG filter with a window length of 7 and a polynomial order of 3 to smooth the signal. The smoothed data are shown in Figure 6. The smoothed data were then normalized to ensure a consistent scale across all subcarriers. The normalized output is depicted in Figure 7. Finally, we applied PCA to reduce dimensionality by selecting the top 10 principal components. The PCA-transformed data are visualized in Figure 8.

Model Training and Performance Evaluation

Following data preprocessing, we proceeded to the model training phase, evaluating three machine learning architectures: CNN-LSTM, GNN, and Transformer. Each model was trained with hyperparameter tuning to optimize its performance and minimize overfitting.

To assess and compare performance, we tracked accuracy and loss over training epochs and generated comprehensive classification reports. These reports included accuracy, precision, recall, and F1-score for each activity class. Furthermore, we visualized confusion matrices to gain deeper insight into misclassification patterns and class-specific challenges.

CNN-LSTM Training Procedure

The CNN-LSTM training pipeline starts by organizing the dataset into folders by activity classes: falling, lying down, staying still, transitioning, and walking. Each folder contains CSV files representing individual activity instances, from which CSI values are extracted. Label encoding converts categorical activity names into numeric labels, a necessary step for neural network classification. The dataset is then stratified and split into 80% training and 20% testing sets, ensuring balanced representation of each activity.

Before training, the data are reshaped to fit the CNN-LSTM model by treating each sample as a sequence of CSI frames, with dimensions (sequence length, features, width, and channels). The CNN layers process each frame individually to extract spatial features, and then the LSTM layers analyze these features over time to capture temporal dependencies. This combination enables the model to learn both the detailed spatial characteristics of each CSI snapshot and the evolving patterns across the activity sequence, which is essential for accurately recognizing human movements. The detailed CNN-LSTM model architecture and training parameters are summarized in Table 3.

Training uses the Adam optimizer with a learning rate of 0.0005, optimized with sparse categorical cross-entropy loss since the targets are integer labels. Early stopping halts training if validation loss does not improve for 10 epochs, preventing overfitting. The model trains for up to 150 epochs with a batch size of 16, balancing computational efficiency with performance. This design ensures robust learning of complex activity patterns in real-time scenarios.

Performance Evaluation of CNN-LSTM

The model exhibits a strong downward trend in training and validation losses, as illustrated in Figure 9, and reaches a validation accuracy of 92.65%, indicating strong generalization capability.

Table 4 presents the classification performance across each class. High precision and recall are achieved overall, with minor confusion noted for the lying down class. The normalized confusion matrix in Figure 10 shows occasional confusion between lying down and falling and between transitioning and walking.

Performance Evaluation of CNN-LSTM on PCA-Transformed Data

When training on PCA-transformed CSI data, the CNN-LSTM model achieves even higher test accuracy of 94.85%, with stable convergence as shown in Figure 11. This suggests that PCA effectively reduces dimensionality while retaining meaningful variance.

The classification report in Table 5 shows perfect F1-score for transitioning and excellent performance for other classes. However, lying down remains the most challenging class, with a precision of 0.84 and recall of 0.70, indicating some confusion with falling and walking as visualized in Figure 12.

GNN Training Procedure

Similar to the previously discussed methods, the GNN training pipeline begins by loading the CSI dataset from separate directories, each corresponding to a specific activity class. Labels for each activity class are numerically encoded using a label encoder.

Next, we transform the CSI data into graph-structured data suitable for input into GNN. Each CSI sample is converted into a graph where each node corresponds to each subcarrier, and the features associated with each node consist of a time-series signal of length 100 (representing the sequence length). A predefined chain graph structure with self-loops is created to connect these nodes, effectively modeling relationships between adjacent subcarriers.

The dataset is then encapsulated into a custom PyTorch Geometric InMemoryDataset class (CSIGraphDataset), creating graph objects containing node features, edge information, and corresponding labels.

The model itself is defined using PyTorch Geometric’s graph convolutional layers. The architecture consists of two graph convolutional network convolution layers, each followed by a Rectified Linear Unit (ReLU), to extract spatial relationships across subcarriers within each CSI graph. A global mean pooling operation is applied to aggregate node-level features into a graph-level representation, which is then passed through a linear fully connected layer to predict activity classes. The model uses a log-softmax activation function for output, suitable for multiclass classification. The model architecture and training parameters of GNN are presented in Table 6.

Performance Evaluation of GNN

The performance evaluation of the GNN model shows promising yet slightly lower results compared to the CNN-LSTM models. The loss and accuracy curves, as displayed in Figure 13, indicate good convergence, with validation accuracy reaching 89.41%. The classification report, as presented in Table 7, highlights that the model performs well on activities such as falling, still, transitioning, and walking. However, the model struggles with the lying down class, yielding the lowest F1-score of 0.65, due to both lower precision (0.67) and recall (0.64). This is further reflected in the confusion matrix, where 21% of lying down samples were misclassified as falling, and 14% as walking, as visualized in Figure 14.

Performance Evaluation of GNN on PCA-Transformed Data

The performance evaluation of the GNN model trained with the PCA dataset shows that while the overall accuracy is 85.59%, there are clear weaknesses in classifying some activity classes. As shown in the loss and accuracy curves in Figure 15, we can see that the model generally converged well within 30 epochs, though the validation loss plateaued earlier than the training loss. According to the classification report in Table 8, the lying down class suffers the most, with a precision of 0.46, recall of 0.41, and F1-score of 0.44, indicating that the model has difficulty distinguishing it. Other classes, such as falling, still, transitioning, and walking, perform significantly better, all with F1-scores above 0.87. The normalized confusion matrix, illustrated in Figure 16, also highlights that lying down is frequently confused with falling and walking, which suggests overlapping features or PCA removing critical discriminative information.

Performance Evaluation of Transformer

The performance evaluation of the Transformer model indicates moderate effectiveness in classifying human activity using CSI data. The model yielded a test accuracy of 71.32%, with the training and validation loss showing consistent convergence, as depicted in Figure 17. However, validation accuracy plateaued below the training accuracy, suggesting slight overfitting.

From the classification report displayed in Table 9, the still class performed best with an F1-score of 0.85, while lying down and transitioning suffered poor recall scores of 0.15 and 0.17, respectively. These two classes are often confused with walking, as displayed in the normalized confusion matrix in Figure 18. For instance, 52% of lying down samples was misclassified as walking, and 45% of transitioning was misclassified similarly.

This implies that although the Transformer model is capable of identifying more stationary actions like still and falling, it struggles with nuanced or ambiguous transitions such as lying down and transitioning, which share overlapping CSI characteristics with walking.

Performance Evaluation of Transformer on PCA-Transformed Data

The evaluation of the Transformer model trained on the PCA-reduced CSI dataset indicates moderate performance, yielding a test accuracy of 83.09%. As shown in the training curves in Figure 19, both training and validation loss decrease consistently, following a smooth downward trajectory. The model’s accuracy also improves progressively across the 75 training epochs. However, a slight overfitting trend emerges near the end, evidenced by the growing disparity between training and validation accuracy.

The classification report, as highlighted in Table 10, reveals that falling, still, transitioning, and walking are well recognized with F1-scores above 0.84, whereas lying down is significantly underperforming with an F1-score of only 0.37. The confusion matrix, as visualized in Figure 20, confirms that lying down instances are often misclassified as falling, still, or walking, which compromises the overall reliability for anomaly detection.

Despite the decent average performance, this result suggests that PCA-based dimensionality reduction may not preserve all the discriminative features necessary for distinguishing subtle activities like lying down, indicating a possible trade-off between model efficiency and recognition accuracy.

Among the three models evaluated, CNN-LSTM, GNN, and Transformer, each demonstrated unique strengths and sensitivities to PCA-transformed data. The CNN-LSTM model performed consistently well both with and without PCA, indicating its robustness in handling high-dimensional raw CSI data. A slight accuracy improvement after PCA implies that the model benefits from noise reduction and feature selection while retaining essential temporal dynamics. On the other hand, the GNN model yielded better results without PCA. This suggests that PCA likely removed key spatial relationships among CSI subcarriers—relationships that GNNs are specifically designed to leverage. In contrast, the Transformer model experienced significant performance gains with PCA, demonstrating its preference for lower-dimensional inputs to avoid complications associated with processing complex raw data.

The variation in performance across models can be attributed to architectural suitability. CNN-LSTM excels at learning both spatial and temporal features, making it well equipped to process sequential CSI data in its raw form. GNNs, which are structured to model relationships between nodes (e.g., subcarriers), rely heavily on the original structure of the input. Applying PCA alters this structure and likely impairs the GNN’s ability to fully capture spatial dependencies. The Transformer, which depends on self-attention mechanisms, benefits from dimensionality reduction since lower-dimensional inputs simplify attention computations and reduce noise, thereby improving its ability to generalize.

PCA proved to be a double-edged sword. While it enhances training efficiency by reducing feature space and filtering noise, it can also eliminate subtle but crucial features necessary for accurate classification. This trade-off was particularly evident in the GNN and Transformer models, where performance on certain classes—especially the lying down class—dropped after applying PCA. This suggests that important discriminative features may have been lost during dimensionality reduction, impairing class-wise sensitivity. Therefore, while PCA can be beneficial for simplifying data, its impact must be carefully assessed based on the architecture and task.

Real-Time System Architecture

To enable seamless HAR in dynamic environments, we developed a real-time processing pipeline that leverages a sliding window mechanism. This technique allows continuous packet capture, data preprocessing, and activity classification, ensuring timely and accurate system responses. The pipeline operates on an end-to-end automated framework, integrating real-time packet acquisition, signal transformation, and predictive modeling, as illustrated in Figure 21.

Specifically, Wi-Fi packet captures are initiated remotely using tcpdump over Secure Shell (SSH), allowing efficient and scalable deployment across edge devices. Each capture batch consists of 100 packets, stored into timestamped Packet Capture (PCAP) files at predefined intervals. Importantly, the sliding window mechanism creates overlapping capture sessions, which ensures no loss of information and facilitates uninterrupted feature extraction—a critical requirement for real-time classification systems.

Upon collection, each PCAP file is instantly converted to CSV format using a custom parsing script. This intermediate format simplifies downstream processing by structuring the raw CSI in a tabular format. Preprocessing is then performed on each CSV file to enhance data quality and uniformity. Non-relevant subcarriers are excluded to reduce noise and computational overhead. Subsequently, a Hampel filter is applied to eliminate outliers, while the SG filter is employed to smooth temporal fluctuations in the CSI signals. These signal processing techniques are chosen for their robustness in handling real-world, noisy wireless data. Finally, min-max normalization is used to scale the data features between 0 and 1, ensuring consistent input ranges for the deep learning model.

The normalized data are subsequently reshaped to fit the input dimensions required by our pre-trained CNN-LSTM model. The model, which combines spatial feature extraction via convolutional layers and temporal sequence modeling through LSTM units, is used to predict the human activity class in real time. This architecture is well suited for CSI data, where both spatial variations across subcarriers and temporal dynamics across packet sequences are significant for accurate classification.

To manage multiple tasks such as packet capture, file conversion, preprocessing, and inference concurrently, the system employs Python’s threading module in combination with the ThreadPoolExecutor. This multi-threaded design ensures that each pipeline component runs in parallel without introducing latency. Moreover, real-time file system monitoring is implemented using the watchdog library, enabling immediate detection and processing of new PCAP files as they are generated.

Real-Time Performance Validation

We validated the real-time performance of the CNN-LSTM model with four activity classes: falling, staying still, transitioning, and walking. Each class had 500 samples, and the model achieved an overall accuracy of 95.48%. Each activity was performed 20 times. The results are shown in Table 11.

After confirming good performance with four activities, we introduced a fifth class, lying down. Real-time validation was conducted again with and without PCA under LOS and NLOS conditions.

Testing was conducted across five activities: falling, lying down, staying still, transitioning, and walking under both LOS and NLOS conditions. Each activity was performed 20 times. Results for CNN-LSTM with and without PCA are shown in Tables 12 to 15.

A recurring issue was the confusion between falling and lying down across LOS and NLOS conditions. This is likely due to the postural similarity reflected in similar CSI patterns. While PCA helped reduce this confusion, it did not eliminate it.

PCA improved model performance in NLOS scenarios by filtering out irrelevant variance and highlighting key features, particularly improving classification of staying still, transitioning, and walking.

Signal obstruction under NLOS increased the falling vs. lying down confusion, but the model remained robust for other classes. Overall, PCA enhanced robustness, but additional techniques may be needed to resolve overlapping patterns in similar activities.

The classification results for the CNN-LSTM model in the second setup, where the Wi-Fi router and Raspberry Pi were placed in different rooms with the person present for testing, are shown in Table 16. This setup, depicted in Figure 3, simulates an extreme NLOS scenario, reflecting a common real-world situation, where the Wi-Fi router and monitoring devices (like the Raspberry Pi) are located in separate rooms, and the person is in a different room.

The system performs well across all activity categories, even in this challenging extreme NLOS setup. It correctly identifies 16 out of 20 Falling instances, 19 out of 20 Staying Still instances, and 19 out of 20 Walking instances, with only a few misclassifications. The system is particularly effective at detecting minimal movement and static activities, such as Staying Still, and can reliably identify dynamic actions like Walking. For Transitioning, the system achieved 17 correct classifications but had a few more misclassifications, likely due to the subtle nature of the movement in NLOS conditions.

Overall, the results demonstrate that the CNN-LSTM model is robust in recognizing human activities, even in environments where environmental separation and potential signal degradation could be challenging. These results highlight the system’s suitability for practical applications in smart home monitoring systems, showing its ability to classify activities effectively even in non-ideal conditions.

Real-Time Performance Validation in a Hospital Environment

To evaluate the performance of the proposed CNN-LSTM model in a real-world setting, we performed real-time experiments in a real hospital room. The model was trained to classify four types of human activities—falling, staying still, transitioning, and walking—based on temporal-spatial features derived from sensor data.

As shown in Table 17, the model achieved promising classification performance in a complex and dynamic hospital environment. The most critical activity class—falling—was correctly classified in 12 out of 20 cases, resulting in a true positive rate of 60%. Misclassifications were primarily observed with staying still (five instances) and walking (three instances), which may be attributed to similar postural signatures or transitional motions immediately following a fall.

The staying still class exhibited strong recognition accuracy with 15 correct predictions out of 20, although four samples were misclassified as walking. This suggests that the model may sometimes conflate motionless periods at the end of a walking sequence with intentional stillness.

Transitioning activities were recognized correctly in 15 of 20 cases, with occasional confusion with staying still (three instances) and walking (two instances), likely due to the overlapping temporal nature of those activities.

Walking was the most reliably detected class, with 16 correct predictions. Only four samples were misclassified—two as falling, and one each as staying still and transitioning. This demonstrates the robustness of the model in identifying dynamic movement patterns associated with walking.

Overall, the results demonstrate that the CNN-LSTM model can effectively distinguish between critical human activities in real time within a hospital environment, which is inherently noisy and filled with unpredictable human motion. While further tuning is necessary to improve fall detection sensitivity, the model shows strong potential for integration into intelligent patient monitoring systems.

Generalization Performance

Assessing how well a model generalizes to unseen individuals is essential for Wi-Fi CSI-based HAR and fall detection. In real-world deployments, models must reliably identify activities across a variety of people, environments, and postural behaviors. However, due to resource constraints, our current dataset includes only two participants.

To approximate subject-independent evaluation, we adopt a leave-one-participant-out-like strategy: the models are trained exclusively on data from one participant and then tested on the combined data from both participants. While this approach does not constitute a full Leave-One-Subject-Out (LOSO) protocol, it provides preliminary insight into the model’s ability to capture subject-invariant features from CSI data and its potential for generalization beyond the training individual.

The results indicate that, although performance slightly decreases compared to within-subject evaluations, the CNN-LSTM model maintains strong discriminative capability across the unseen participants. This suggests that the model is able to extract meaningful patterns from CSI that are not overly dependent on the specific characteristics of a single participant.

We acknowledge that this evaluation is limited by the small participant pool and does not fully reflect real-world variability. Future work will involve a rigorous LOSO protocol with a substantially larger and more diverse cohort of participants, incorporating variations in age, body type, and movement style. Expanding the dataset and testing across diverse scenarios will enable a more thorough assessment of generalization performance, which is crucial for deploying CSI-based fall detection systems in healthcare and assisted living environments.

Comparison With State-of-the-Art

To contextualize the contribution of our system, we benchmarked it against recent state-of-the-art methods in CSI-based HAR published between 2021 and 2025. These works include small-scale and offline approaches [12, 14, 15], architecture-focused models such as InceptionTime [13], and more recent efforts employing mobile receivers, NAS-pruned networks, adaptive convolutional systems, and self-supervised or cross-domain learning [17–21].

As Table 18 shows, most existing approaches are constrained by small datasets, controlled environments, and offline evaluation pipelines. Even recent advances that improve efficiency or adaptability, such as NAS-pruned or adaptive CNN models, still lack validation in real-time and real-world scenarios. Self-supervised and cross-domain methods improve representation learning, but they remain restricted to offline analysis without deployment considerations.

In contrast, our work distinguishes itself in three critical ways. First, it employs a substantially larger and more diverse dataset collected across multiple indoor settings, including both LOS and NLOS scenarios. Second, it integrates a robust preprocessing pipeline, including subcarrier selection, Hampel and SG filtering, normalization, and PCA, to enhance signal quality and reduce overfitting. Third, and most importantly, it delivers an end-to-end real-time HAR system on low-cost hardware, demonstrating practical deployability rather than remaining confined to simulation or offline experiments. This positions our system as a step forward toward bridging the gap between academic prototypes and scalable real-world HAR solutions.

Limitations and Future Work

Despite its promise, the study has several limitations that point to directions for future research:

• Dataset Size and Diversity: The dataset was limited in participant number and activity samples, with 700 falling events and 2,700 normal activity events (lying down, staying still, transitioning, and walking). This constrains generalization, particularly for distinguishing similar postures such as falling vs. lying down. Expanding the dataset to include more subjects, varied environments, and activity patterns will improve robustness.

• Generalization to Unseen Individuals: Subject-independent evaluations were preliminary, with models trained on one participant and tested on others as a proof-of-concept. A comprehensive LOSO evaluation with a larger cohort is necessary to ensure real-world applicability.

• Model Interpretability: CNN-LSTM models act as black boxes, which can limit trust in healthcare applications. Future work should explore explainability techniques, such as saliency maps or attention visualization, to understand which CSI features influence classification decisions.

• Data Augmentation: Rare but critical classes, especially falling and lying down, should be enriched with diverse synthetic or real samples to reduce confusion under LOS and NLOS conditions.

• Multi-Sensor Fusion: Integrating Wi-Fi CSI with other modalities—such as accelerometers, gyroscopes, or vision-based sensors—may enhance accuracy and robustness, particularly in complex environments like hospitals.

• Lightweight Deployment: Optimizing models for edge devices such as Raspberry Pi using Tiny Machine Learning or efficient architectures (e.g., MobileNet, EfficientNet) would improve portability, energy efficiency, and practical usability.

• Multi-Person Recognition: Extending the system to simultaneously recognize multiple individuals remains a challenging yet important goal for smart home and hospital applications.

In summary, CSI-based fall detection with CNN-LSTM demonstrates high accuracy (up to 95.5% for four-class scenarios with PCA), low latency, and resilience in both controlled and real-world hospital environments. The sliding-window pipeline, multithreaded Python processing, and real-time tcpdump capture confirm practical feasibility. Addressing limitations, particularly data set diversity, fall sensitivity, interpretability, and lightweight deployment, will be key to realizing its full potential as a scalable, privacy-preserving monitoring solution for healthcare and eldercare.