The Wi-Fi sensing technology introduces a paradigm shift in violence detection. Wi-Fi sensing, an emerging technology, uses Wi-Fi signals to detect changes in surroundings, motion, and presence in a specific area. Similar to short-range passive radar systems, this innovative technique tracks locations and activities by utilising the interactions between Wi-Fi waves and movement. Employing current Wi-Fi infrastructure, Wi-Fi sensing detects and analyses disturbances caused by movement and environmental changes.
3.1. Activity Recognition
Liu et al. [
16] introduce a method for detecting intense human motion in enclosed surroundings based on CSI obtained from Wi-Fi signals. The study tackles the requirement for passive human detection without the need for any devices in a variety of contexts, including automated homes, geriatric care, protecting assets, and detection of intrusion. The authors propose a system that utilises the physical layer information in wireless signals, specifically the CSI, to identify strong human motion involving numerous targets. The proposed method in the paper involves several key steps to achieve device-free, passive detection of intense human propulsion by means of CSI from wireless transmissions. Initially, the system focuses on line-of-sight (LOS) and non-line-of-sight (NLOS) identification by analysing the skewness of channel impulse response (CIR) distribution. To mitigate the sensitivity of wireless signals to environmental changes, signal preprocessing techniques are applied, including the use of a band-pass filter to filter out unrelated frequency components of a signal. Subsequently, features such as standard deviation, median absolute deviation, interquartile range, and signal entropy are calculated from CSI phase and amplitude difference to effectively characterise various human motions.
For classification, the method employs a one-class support vector machine (OSVM) classifier, utilising training sets containing only one positive sample type, namely intense human motion, without incorporating other sample types. The system’s performance is evaluated through experiments conducted in both enclosed and semi-closed spaces. The evaluation encompasses LOS/NLOS identification and detection of intense human motion. Furthermore, the paper explores the impact of factors such as transmitter–receiver height, sample size, and distance between the receiver and transmitter on the overall effectiveness of the proposed method. Experiments are conducted in semi-closed and confined areas, simulating real-world indoor scenarios. The collected data include various human activities in LOS and NLOS conditions, helping train and validate the proposed detection system. The paper reports promising results for LOS/NLOS identification, maximising the rate of detection while minimising the rate of false alarms. For intense human motion detection, the system exhibits notable performance, with sensitivity (true positive rate) and specificity (true negative rate) reaching satisfactory levels, especially when considering both LOS and NLOS conditions.
The proposed system presents advantages, leveraging existing Wi-Fi access points to achieve device-free, passive detection of human motion without the need for additional equipment. Notably, privacy concerns related to camera-based methods are effectively addressed, making them well suited for deployment in private spaces. The utilisation of CSI enhances the system’s capabilities by providing fine-grained information, enabling the detection of complex human motions. However, some shortcomings exist. The system’s performance is influenced by the height of the transmitter–receiver pair, with optimal results obtained at specific heights. Additionally, distinguishing between intense human motion and regular activities may pose challenges, potentially leading to misclassifications.
Furthermore, the effectiveness of the system diminishes in semi-closed spaces compared to enclosed spaces, highlighting considerations for specific deployment scenarios. Potential applications for the proposed system include smart buildings, elderly support, safeguarding assets, and theft prevention, among others. Its ability to operate without additional devices makes it suitable for deployment in environments where privacy concerns or the inconvenience of wearable sensors exist. The focus on intense human motion detection positions it for applications in security monitoring and incident identification in complicated indoor scenarios. In conclusion, the paper presents a robust framework for device-free, passive detection of intense human motion using CSI from Wi-Fi signals. The experimental results demonstrate its effectiveness in real-world scenarios, highlighting its potential for applications in diverse domains.
Gu et al. [
17] discuss a novel approach for recognising human activities using Wi-Fi ambient signals. The authors highlight the challenges posed by the increasing number of sensors in devices for the Internet of Things (IoT) applications, particularly in terms of volume and energy issues. They propose utilising wireless ambient signals, specifically radio frequency (RF) transceivers, as an alternative source of information for activity recognition. The authors conducted an empirical study to understand the impact of various activities on RSSI values, a key parameter in Wi-Fi communication. They explore the stability of wireless signals in indoor environments and propose a method to recover physical activities by analysing signal fluctuations caused by human movements.
The research builds upon previous works that primarily focused on the utilisation of wireless transmissions for services that use location. However, the paper introduces a novel approach by connecting pervasive wireless signals with the recognition of human activities, a concept referred to as “WiSee”. The paper acknowledges pioneering works such as WiSee and Wi-Vi but emphasises the need for solutions compatible with commodity Wi-Fi devices. The authors explore the details of two categories of solutions: CSI-based and RSSI-based. While CSI offers unique advantages, its limitations in terms of hardware and complexity lead the authors to favour RSSI as a more practical trade-off between efficiency and cost, particularly for smartphones and tablets.
The empirical investigation of Wi-Fi characteristics is delineated by the authors, who emphasise the influence of diverse activities on RSSI. The research findings indicate that activities do indeed influence the signal, with distinct activities producing unique patterns or fingerprints on Wi-Fi RSSI. The architecture of the recognition system is presented, involving three layers: access points as signal sources, smart devices for data collection, and a server for storing, analysing, and recognising activities. Real-world applications can utilise the online fingerprint-based activity recognition system without requiring adjustments at access points or mobile entities due to its adaptability and flexibility. The paper details the preprocessing module, aimed at filtering abnormal samples, and the classification module, which selects suitable features for activity recognition. The selection of recognition features involves a case study, focusing on the mean (μ) and standard deviation (σ) of RSSI values.
The authors introduce a fusion algorithm, combining k-nearest neighbours (k-NN) classification with a classification tree, to enhance recognition precision. Six distinct activities are prototyped into the system as part of the performance evaluation process. The proposed fusion algorithm outperforms other classifiers, such as k-NN, naïve Bayes, and bagging, in terms of recognition precision. The paper acknowledges the importance of carefully selecting the group size for optimal performance. Advantages of the proposed system include its flexibility, adaptability, and the ability to work with commodity Wi-Fi devices. The fusion algorithm enhances recognition accuracy by combining features and addressing challenges related to similar footprints of different activities. The system requires no modifications at the ends of access points or mobile entities, making it suitable for real-world applications.
However, the paper acknowledges that the size of data groups during recognition is crucial, and the system’s performance is impacted by the choice of this parameter. Additionally, while the proposed system shows promising results, it is essential to consider scalability and potential challenges in diverse environments. In conclusion, the paper highlights the contributions of Passive Human Activity Recognition Based on Wi-Fi Ambient Signals, including empirical results, an online fingerprint-based architecture, and extensive evaluations. The proposal establishes the system as a viable remedy for dormant activity by humans detection via Wi-Fi ambient signals, with potential applications in various real-world scenarios.
The research on Wi-Fi sensing for smart residential environments with real-time activity monitoring by Sahoo et al. [
18] addresses the broad applications of human activity recognition (HAR). The study identifies traditional sensors’ limitations, particularly the intrusive nature of proximity, pressure, and light sensors. Instead, it proposes harnessing RF signals, specifically Wi-Fi signals, to achieve real-time activity detection. The core concept revolves around utilising CSI derived from Wi-Fi signals. The proposed system introduces a two-layer architecture employing low-cost IoT devices, specifically the ESP32 micro-controller. The architecture consists of an IoT layer responsible for data generation and an edge layer for real-time data processing, filtering, and visualisation.
The methodology encompasses data collection using ESP32 devices, data transfer between IoT and edge layers, and processing CSI data to extract amplitude and phase information. The novel two-layer architecture focuses on edge computing technology for real-time activity detection. The IoT layer comprises physical devices, while the edge layer collects, processes, and visualises data almost in real-time. The paper concentrates on the ESP32 micro-controller for both IoT and edge layers, highlighting the flow of CSI data from IoT to edge for further analysis. The proposed system’s advantages include simplicity, low cost, and real-time processing capabilities.
The ESP32 micro-controller is positioned as a superior alternative to traditional Wi-Fi sensing tools. However, the paper acknowledges a 70% accuracy in activity classification using the lightweight SVM algorithm, indicating room for improvement. Additionally, the limitations of the ESP32, such as its low processing capability, are recognised. In terms of applications, the paper suggests its relevance in smart home environments, where real-time activity detection can enhance the user experience and enable various services. The proposed system can find utility in monitoring daily activities, providing security, and facilitating personalised human-computer interactions within smart homes.
Yang et al. [
19] introduce a novel approach for HAR in enclosed environments using commercial Wi-Fi devices. Traditional HAR systems face challenges related to inconvenience, environmental restrictions, and the need for special equipment. In contrast, Wi-Fi-based systems, leveraging CSI, offer a device-free solution with potential applications in smart homes, security monitoring, medical assistance, and more. The proposed system, termed WiTA, focuses on three technical challenges. Firstly, it addresses the separation of effective signals with multipath from the original CSI. An introduced method for the precise extraction of multipath signals utilises the propagation delays of different multipath signals. Secondly, the paper tackles the segmentation of effective motion fragments, avoiding the need for re-adjustment of parameters when the signal environment changes.
Thirdly, it places significant emphasis on the efficient extraction of correlation features from the initial CSI. WiTA comprises two main modules: CSI data processing and neural networks. Involved in the CSI data processing module are the extraction of features, multipath separation, noise elimination, and feature acquisition. The neural network module integrates temporal–frequency attention through the utilisation of an attention mechanism integrated into a multi-layer LSTM network architecture. The proposed algorithm in the paper offers several advantages in the context of human activity recognition. Firstly, it effectively addresses the challenge of separating multipath signals, ensuring accurate recognition of human activities. This is a critical aspect, particularly in complex environments where signal interference is common. Additionally, the end-to-end method employed by the algorithm minimises information loss during data processing, enhancing the overall efficiency of the recognition system.
The introduction of a temporal–frequency attention mechanism in WiTA further contributes to its effectiveness. This attention mechanism improves the system’s ability to focus on relevant features, ultimately leading to enhanced recognition accuracy. However, the algorithm comes with certain shortcomings that warrant consideration. One notable challenge is the complexity of training, especially due to the interaction between four sub-networks during the training process. This complexity may necessitate a step-by-step training method to ensure optimal performance. Another limitation is the acknowledgement of the system regarding the processing capacity, particularly in the ESP32 micro-controllers utilised in the experimental setup. This recognition of processing limitations suggests that the algorithm’s practical implementation may face constraints in certain hardware environments.
Overall, while the algorithm presents significant advancements in human activity recognition, careful consideration of these advantages and shortcomings is crucial for its successful implementation. The authors present experimental results using the WiAR dataset and their own collected data, demonstrating an average recognition accuracy of 94%. In comparison to other prevalent recognition algorithms, the proposed algorithm demonstrates superior accuracy. Additionally, the effect of the quantity of training samples on recognition accuracy is assessed, indicating a positive relationship.
Feng et al. [
20] explore the development of a system, named Wi-Multi, for human activity recognition based on CSI obtained from commercial wireless devices. The application areas targeted include healthcare, security, and IoT. Traditional approaches for activity recognition often involve sensors or cameras, but these methods face issues such as inconvenience for users wearing sensors and privacy concerns with camera-based systems. The proposed system leverages CSI extracted from commercial Wi-Fi devices, eliminating the need for additional equipment and addressing cost concerns. The paper’s deep learning network selects high-level characteristics independently, with no pre-processing modules required. The results of the tests indicate that the system might successfully strike a balance between efficiency and accuracy at various stages, with Wi-Multi achieving an average accuracy of 96.1%.
One key contribution is the introduction of a three-phase system tailored for various stages of system setup. When there are few samples in the profile, Phase 1 is used, based on principal component analysis (PCA), discrete wavelet transform (DWT), and distance-based classification. Using the SVM algorithm, features are extracted from the time and frequency domains and classified in the second phase. Phase 3 employs LSTM in a deep learning network when a large number of samples are available. The research highlights a new activity extraction technique that can locate an activity’s beginning and ending even in noisy settings with several subjects. The algorithm employs outlier filtering, differential algorithms, and eigenvalue comparison to achieve robust activity sample extraction.
While the paper highlights several advantages, such as the use of existing Wi-Fi devices, the autonomous feature selection of the deep learning network, and the effectiveness of the activity extraction algorithm, it does not explicitly address potential shortcomings or limitations. The experimental results demonstrate the system’s performance in terms of accuracy, efficiency, and the ability to recognise the number of subjects present in a situation. The Wi-Multi system offers a comprehensive approach to human activity recognition using CSI from Wi-Fi signals. The three-phase design caters to different stages of system deployment, showcasing adaptability and robustness in various environments. The proposed system presents a promising direction for applications in healthcare, security, and IoT, providing a balance between accuracy and efficiency.
3.2. Violence Detection
Zhang et al. [
21] introduce WiVi, a pioneering passive violent behaviour detection system utilising commercial Wi-Fi infrastructure. Addressing the escalating global concern of school violence impacting youth’s physical and mental health, WiVi leverages CSI from Wi-Fi signals to recognise violent activities. The existing challenges in violence detection, including the limitations of traditional methods and privacy concerns, prompt the need for innovative solutions. WiVi leverages the widespread implementation of Wi-Fi infrastructure, in particular the CSI offered by the physical layer, to expose multipath characteristics impacted by human movements and actions. The study highlights the underuse of CSI by current techniques and suggests a unique method that combines correlated information generated from the combination of distinct CSI subcarriers’ time series data. A PCA-based feature fusion technique is created to integrate both time series and correlated information adaptively in various contexts, while a Gabor filter-based feature extraction method is presented for the automatic extraction of correlated features.
The authors highlight the significance of WiVi’s ability to precisely identify complex violent activities, even in changing operating environments. They address the limitations of existing methods, which often concentrate on pre-defined actions and struggle to adapt to diverse scenarios. The proposed feedback adjustment method ensures adaptability for environmental changes by readjusting model parameters and potentially retraining the model when necessary. WiVi’s prototype, implemented on commercial Wi-Fi devices, demonstrates impressive performance with a 93.46% accuracy in detecting violent activities and a 6.43% false alarm rate. The paper concludes by summarising the key contributions, emphasising the combination of time series and correlated features, the Gabor-filter and PCA-based methods, and the successful implementation and evaluation of WiVi in various environments.
In conclusion, WiVi represents a significant advancement in violence detection systems, offering a reliable and adaptive solution with potential applications in ensuring the safety of educational environments. The authors provide a comprehensive overview of their methodology, results, and contributions, laying the groundwork for future research in the field.
Zhou et al. [
22] introduce Wi-Dog, a non-invasive system for monitoring physical assaults in smart schools using commercial Wi-Fi devices. The research addresses the increasing concern of school violence and leverages Wi-Fi signals to detect and alert in real-time instances of physical assault. The motivation for this innovation stems from the limitations of traditional monitoring systems, such as wearable sensors and camera-based approaches, which either compromise comfort or raise privacy concerns. Wi-Dog employs Wi-Fi signals to capture distinct features in the CSI at the physical layer (PL). Unlike conventional approaches, Wi-Dog focuses on the irregular and unpredictable nature of assault events. The system overcomes three critical challenges: extracting motion data from CSI dynamics with noise, detecting abnormal transitions during human interactions, and differentiating actual attacks from acts resembling assaults.
Wi-Dog employs a multifaceted approach to non-invasive physical assault monitoring through commercial Wi-Fi devices. In relation to motion information extraction, the system leverages spatial diversity, analysing variations in CSI waveforms across different antennas for accurate and abundant motion cues. Noise reduction steps are strategically implemented to eliminate irrelevant interferences while retaining crucial motion data. Wi-Dog anomalous transition detection differs from traditional methods in that it makes use of the target frequency band’s signal complexity to monitor the amount of irregularity and intensity change. The system further enhances location-independent detection through cross-correlation of adjacent subcarriers. For the differentiation of assaults, novel features such as Doppler frequency shifts are extracted to represent intensity levels. Employing a SVM classifier in local analysis, Wi-Dog reduces false alarms by considering irregularity and continuity over longer time durations. During the experimental study, when two volunteers simulate real physical collisions where drastic conflicts exhibit significant intensity cues such as speed, acceleration, and kinetic energy, the fine-grained spectrogram resulting from distinct power variations in frequency bands is shown in
Figure 4.
This comprehensive methodology showcases Wi-Dog’s ability to effectively and innovatively monitor physical assaults in real-time, offering a promising solution for enhancing safety in smart school environments. The research validates Wi-Dog through experiments involving imitated physical attacks in both LOS and NLOS environments, utilising real-time CSI measurements from commercial Wi-Fi devices. The results demonstrate high true detection rates (0.94 in LOS, 0.85 in NLOS) and low false alarm rates (0.08 in LOS, 0.11 in NLOS) across various scenarios and parameter changes, showcasing the system’s robustness. Wi-Dog offers several advantages, including non-invasiveness, addressing privacy concerns associated with camera-based monitoring, and its potential for ubiquitous deployment through existing Wi-Fi infrastructures. However, it may be limited in generalisation to specific environments and necessitates further adaptation for broader applications. Despite its promising results, consideration of false positives in complex real-world scenarios remains a point of discussion. The application of Wi-Dog extends beyond school violence prevention, demonstrating its potential for general emergency detection. The proposed technology’s versatility is underscored by potential applications, such as injury rescue, elderly healthcare, and terrorist threat warning.
Hsu et al. [
23] address the increasing threat of cybercrime, particularly through attacks on public Wi-Fi networks, focusing on the notorious “evil twin” phenomenon. Evil twins are rogue access points that mimic legitimate ones, putting users at risk of data theft. The authors highlight the limitations of existing solutions and suggest a proactive client-side remedy, the wireless packet forwarding detector (WPFD), designed for Wi-Fi users without requiring specialised equipment or system management. The paper outlines the challenges posed by evil twin attacks, emphasising the need for user-friendly and effective solutions. The authors provide WPFD, an active client-side solution that makes use of monitoring techniques based on the transmission control protocol/internet protocol (TCP/IP). To identify the existence of a malevolent twin, Monitoring the retransmission behaviour of the corresponding synchronisation/acknowledgment (SYN/ACK) packets, WPFD transmits SYN packets to well-known websites.
The proposed solution aims to be user-friendly, requiring only the TCP/IP header of wireless packets for detection. WPFD is designed to operate in real-time, sending probe packets to significant websites for detection. The WPFD method presents several advantages as a user-friendly solution for Wi-Fi users, eliminating the need for specialised equipment or system management. WPFD claims to provide accurate and fast detection of evil twins through its active approach, offering independence from specific network information, such as legal access points/internet protocols or training data. Notably, WPFD sets itself apart by seamlessly integrating both utilising one WNIC in both passive monitor mode and active probe packets.
However, the paper has shortcomings, including a lack of experimental details, as specific datasets or real-world experiments validating WPFD’s effectiveness are not provided. Additionally, the limited depth of the security discussion raises the need for a more thorough analysis of potential vulnerabilities and countermeasures to strengthen the overall robustness of the proposed solution. The primary application of WPFD is to enhance the security of Wi-Fi users by actively detecting and preventing evil twin attacks. It offers a practical and efficient solution for individuals using public Wi-Fi networks, where the risk of evil twin attacks is prevalent. In conclusion, the paper introduces WPFD as a user-side solution to address the security concerns associated with evil twin attacks. While it emphasises user-friendliness and effectiveness, further validation through detailed experiments and a more thorough security analysis would contribute to establishing the robustness of WPFD in real-world scenarios.
Liu et al. [
24] present an inventive strategy for identifying violent behaviour by leveraging CSI from wireless signals, specifically Wi-Fi. The authors highlight the severe psychological impact of violence on victims, including depression, anxiety, and suicide, and address the challenges associated with monitoring unpredictable incidents in diverse settings. Critiquing existing methods for their costliness and privacy concerns, the paper proposes a cost-effective and privacy-friendly alternative based on CSI. Emphasising the advantages of wireless technology, particularly the resilience of CSI-based methods to environmental factors, the authors argue for the accessibility and affordability of their approach. They position their method in the broader context of CSI-based human behaviour recognition, citing previous research on applications like fall detection and gesture recognition. This places the proposed violent behaviour detection method within the evolving landscape of CSI-based technologies. The proposed method leverages the advantages of wireless technology for human behaviour recognition, specifically CSI.
The authors contend that CSI-based techniques are advantageous for maintaining privacy because they are unaffected by external variables like light and temperature. The method involves extracting features from the frequency domain, time domain, and image domain, adopting a multi-domain approach to capture more information from CSI. The model considers the complex indoor wireless signal propagation environment, emphasising the effects of human behaviour on received signals, including Doppler frequency shifts. The experiments employ an emerging class of commercial Wi-Fi equipment based on orthogonal frequency division multiplexing—multiple input multiple output (OFDM-MIMO) to collect fine-grained CSI information. The data collection process aims to measure changes in CSI resulting from distinguishable propagation paths, including direct paths, reflections, diffractions, and refractions. The experimental findings underscore the efficacy of the suggested approach in identifying violent conduct.
Specifically, the average recognition accuracy is reported as 97.3% in darkroom scenarios and 92.7% in laboratory scenarios. These outcomes emphasise the system’s robust performance and its capacity to function optimally using existing Wi-Fi infrastructure. This characteristic positions the method not only as an accurate but also a cost-effective and scalable solution. One advantage of the proposed method lies in its reliance on wireless signals. By avoiding the use of cameras, it minimises privacy concerns that are often associated with camera-based methods. This feature enhances the acceptability and ethical considerations of the system. The utilisation of existing Wi-Fi infrastructure is a key cost-saving aspect. It eliminates the need for additional, potentially expensive hardware, making the solution economically viable for widespread adoption in various settings. The system demonstrates impressive accuracy in detecting violent behaviour across diverse scenarios.
The consistently high recognition accuracy underscores its reliability in identifying potential threats, contributing to its overall efficacy. The study candidly acknowledges potential challenges in distinguishing between violent and regular activities. This recognition prompts a call for further refinement, suggesting an avenue for improvement to enhance the system’s capability in accurately discerning different behavioural patterns. The paper notes a decrease in the system’s effectiveness in semi-closed spaces. This limitation implies that the method might encounter challenges in environments with specific structural characteristics, potentially influencing its applicability in certain settings. The proposed CSI-based violence detection method holds promise for enhancing security in various domains, especially in public spaces where the risk of violent incidents exists. Its advantages, including privacy preservation, cost-effectiveness, and high accuracy, position it as a promising solution for real-world deployment. The adaptability of the method to existing Wi-Fi infrastructure makes it suitable for widespread use, effectively addressing the limitations associated with previous hardware-dependent approaches.