Wi-Fi-based continuous activity sensing is of great importance to personal healthcare, security monitoring, and healthy lifestyle assessment. However, it remains challenging to understand multiperson continuous activities as the Wi-Fi signals reflected by each person are mixed up in the Wi-Fi channel state information (CSI). To this end, we present MuSense, the first Wi-Fi-based system that enables multiperson continuous activity segmentation and recognition using commodity devices. In MuSense, we design a Wi-Fi network interface card (NIC) combination and calibration (NICCC) algorithm to construct a high-resolution receiving array and calibrate the CSI measurement noises of this array. On this basis, we propose a multiperson reflection signal separation (MPRSS) algorithm to completely and practically separate each person’s Wi-Fi reflection signals, obtaining the amplitude attenuation and phase shift corresponding to each person’s activities on the subcarrier dimension. Finally, we design an unsupervised adversarial continuous activity sensing network (UACAS-Net), with a two-stage adversarial training method to achieve generalized multiperson continuous activity sensing. Through two-stage adversarial training, UACAS-Net can capture distinguishing features of continuous activities from separated reflection signal streams in the source and target domains while minimizing the feature domain discrepancies to segment and recognize each person’s continuous activities in different domains. Intensive experiments are conducted under three different scenarios and the results demonstrate the effectiveness and practicality of MuSense for multiperson continuous activity sensing.

Previous works have verified the feasibility of WiFi-based human pose estimation (HPE). However, their crucial limitations lie in requiring favorable placement of Wi-Fi devices and only sensing human poses in a small area. To address these issues, we propose WiLink, a Wi-Fi-based 3D HPE system that selectively uses several existing Wi-Fi links to achieve accurate HPE everywhere indoors. We find that the Channel State Information (CSI) fluctuations caused by human pose changes over different Wi-Fi links are various. According to the effectiveness of Wi-Fi links for human pose sensing, we classify the links as Noise-Dominated Links, Most-Effective Links and Redundant Links. Then we propose a Dynamic Link Selection (DLS) mechanism to adaptively select Most-Effective Links for HPE. This process maximizes the importance and minimizes the redundancy of the selected links. Finally, we feed the CSI samples corresponding to Most-Effective Links into a neural network to estimate human poses. Intensive experiments are conducted, and the results show that WiLink achieves a good performance in the scenario with multiple available Wi-Fi links.

Gesture awareness plays a crucial role in promoting human–computer interface. Previous works either depend on customized hardware or need a priori learning of wireless signal patterns, facing downsides in terms of the privacy concern, availability, and reliability. In this article, we propose CentiTrack, the first centimeter-level passive gesture-tracking system that works with only three commodity WiFi devices, without any extra hardware modifications or wearable sensors. To this end, we first identify the channel state information (CSI) measurement error sources in the physical-layer process, and then denoise CSI by the complex ratio between adjacent antennas. Principal component analysis (PCA) is further adopted to separate the reflected signals from noises. Benchmark experiments are conducted to verify that the phase changes of denoised CSI are proportional to the length changes of the dynamic path reflected off the hand. In addition, we adopt the multiple signal classification (MUSIC) algorithm to estimate the Angle-of-Arrivals (AoAs) of dynamic paths, and then locate the initial position of hands with triangulation. We also propose a novel static componnets elimination algorithm for tracking correction by eliminating the components unrelated to motion. A prototype of CentiTrack is fully realized and evaluated in various real scenarios. Extensive experiments show that CentiTrack is superior in terms of tracking accuracy, sensing range, and device cost, compared with the state-of-the-arts.

Daily activity monitoring is essential to healthy lifestyle assessment and personal healthcare, among which Wi-Fi-based solutions have attracted increasing attention due to their no-intrusive and privacy-protected characters. However, related researches are based on the assumption that there is an interval between two activities, during which the target is thought to be static. This assumption falls short of reality as human activities are performed continuously in daily life. Therefore, this article aims to design a nonintrusive and privacy-protected system, namely, Wi-Monitor, to monitor human activities in daily life. In Wi-Monitor, we first fragmentize Wi-Fi channel state information (CSI) streams into CSI bins and design a feature extraction network to extract activity fragmentation features (AFFs) from these CSI bins. From the extracted AFFs, a temporal convolutional network (TCN) is further used to capture activity continuity features (ACFs), which are used as distinguishing characteristics of continuous activities. Finally, Wi-Monitor utilizes these distinguishing characteristics to segment and recognize human activities in daily life simultaneously to achieve daily activity monitoring. In addition, we design an over-segmentation suppression mechanism with two training stages in Wi-Monitor to overcome the over-segmentation issue and enhance the activity monitoring accuracy. Intensive experiments are conducted in three different scenarios and the results demonstrate the effectiveness and practicality of Wi-Monitor for daily activity monitoring.

WiFi-based passive gesture tracking plays a crucial role in promoting human-computer interface, due to its pervasive availability and cost-effectiveness. Prior works focus on tracking gestures on 2D sensing plane by aggregating multiple WiFi links (typically 2), with Uniform Linear Arrays (ULAs). However, gestures actually contain 3D spatial information instead of just 2D, thus interpreting the 3D traces as 2D may lead to enormous tracking errors. This paper aims at exploring the possibility of passively tracking 3D hand traces likewise leveraging two WiFi links with standard 3-element ULAs, and presents a generic 3D centimeter-level passive gesture tracking system, called CentiTrack-3D. To this end, we make two key observations: (1) The ULA-resolved angle contains the integrated information of the azimuth and elevation in 3D space, despite its inability to estimate azimuth and elevation separately. (2) The radial motion deviating from the sensing plane also leads to length variations of paths. Motivated by the observations, a 3D tracking model namedChaosis elaborately designed to deduce the hand 3D coordinates, so as to track the traces. Extensive experiments yield that CentiTrack-3D achieves an overall median tracking granularity of 2.5 cm in 3D space in case of diverse users and environment conditions.

Recently, commodity Wi-Fi devices have been shown to be able to construct human pose images, i.e., human skeletons, as fine-grained as cameras. Existing papers achieve good results when constructing the images of subjects who are in the prior training samples. However, the performance drops when it comes to new subjects, i.e., the subjects who are not in the training samples. This paper focuses on solving the subject-generalization problem in human pose image construction. To this end, we define the subject as the domain. Then we design a Domain-Independent Neural Network (DINN) to extract subject-independent features and convert them into fine-grained human pose images. We also propose a novel training method to train the DINN and it has no re-training overhead comparing with the domain-adversarial approach. We build a prototype system and experimental results demonstrate that our system can construct fine-grained human pose images of new subjects with commodity Wi-Fi in both the visible and through-wall scenarios, which shows the effectiveness and the subject-generalization ability of our model.

Collecting a large amount of labeled data is crutial for training deep neural network, which is a limitation for medical image classification because it necessarily involves expert knowledge. To mitigate this problem of insufficient labeled medical data, in this work, we propose a novel semi-supervised framework for medical image classification. For unlabeled data, we apply the consistency-based strategy to produce high-quality pseudo label, which encourages model to output the same predictions under different perturbations. In addition, we present a novel mixed sample data augmentation CamMix to effectively exploit the relation between samples, mixing pairs of input data and labels according to the class activation map mask. We have evaluated our proposed method on two public medical image datasets, interstitial lung disease dataset and ISIC 2018 skin lesion analysis dataset. The results demonstrate superior performance of our method over other existing methods on the two datasets. Meanwhile, our proposed CamMix performs better than the current mixed sample data augmentation methods.

In this work, we address the issue of Wi-Fi-based human activity recognition (HAR) system in densely deployed Wi-Fi environments. With the benefit of sufficient information provided by Wi-Fi channel state information (CSI), HAR based on Wi-Fi has become an active research area in recent years. Traditional Wi-Fi CSI-based HAR applications usually focus on utilizing one Wi-Fi transmitter and one or several Wi-Fi receivers to extract the activity-related features, ignoring the communication among multiple Wi-Fi devices in the real world. In this paper, we present a novel Wi-Fi link selection model on the basis of continuous state decision-making process in which CSI is modeled as a part of the state. The model, referred to as WiAgent, takes an action of selecting one Wi-Fi link according to current state, and then updates the state for the choice of the next action. From extensive experiment results, our method performs better than other solutions in a given environment where multiple Wi-Fi transmitters exist.

Device-free gesture recognition plays a crucial role in smart home applications, setting human free from wearable devices and causing no privacy concerns. Prior WiFi-based recognition systems have achieved high accuracy in a static environment, but with limitations in adapting changes in environments and locations. In this paper, we propose a fine-grained deep adaptation networks based gesture recognition scheme (DANGR) using the Channel State Information (CSI). DANGR applies wavelet transformation for amplitude denoising, and conjugate calibration to remove CSI time-variant random phase offsets. A Generative Adversarial Networks (GAN) based data augmentation approach is proposed to reduce the large consumptions of data collection and the over-fitting risks caused by incomplete dataset. The distribution of CSI in various environments may be biased. In order to shrink these domains discrepancies in environments, we adopt domain adaptation based on multikernel Maximum Mean Discrepancy scheme, which matches the mean-embeddings of abstract representations across domains in a reproducing kernel Hilbert space. Extensive empirical evidence shows that DANGR yields mean 94.5% accuracy of gesture recognition confronting environmental variations, providing a promising scheme for practical and long-run implementation.

In the era of Internet of Things, human sensing, which detects and interprets human motions including gestures or postures, has emerged as a challenging problem in areas such as assisted living and remote monitoring. Besides conventional sensing methodologies that rely on wearable devices and camera systems, WiFi-based technologies are evolving as a promising solution for indoor monitoring and activity recognition recently. In this paper, we propose WiRoI, a device-free human sensing scheme which is able to precisely detect and interpret human motions within certain spatial region of interest (RoI) using only commodity WiFi devices. To this end, the channel state information (CSI) data in PHY layer is obtained directly by upgrading the firmware. To make the system robust enough, we propose a method to eliminate the noises caused by other humans between TX and RX. And particularly, we are the first to explore spatial region of interest towards accurate and robust human sensing. Experiments were conducted in a common office and the results demonstrated that WiRoI is able to accurately detect and interpret gestures (In order to test the performance of the proposed method, we build a gesture recognition system.) with the accuracy lowered by less than 5% when there are other humans walking or standing between the TX and RX.

Audio represents one of the most appealing yet least exploited modalities in wireless sensor networks, due to the potentially extremely large data volumes and limited wireless capacity. Therefore, how to effectively collect audio sensing information remains a challenging problem. In this paper, we propose a new paradigm of audio information collection based on the concept of audio-on-demand. We consider a sink-free environment targeting for disaster management, where audio chunks are stored inside the network for retrieval. The difficulty is to guarantee a high search success rate without infrastructure support. To solve the problem, we design a novel replication algorithm that deploys an optimal number of$O(\sqrt{n})$replicas across the sensor network. We prove the optimality of the energy consumption of the algorithm. We implement a sink-free audio-on-demand (SAoD) WSN system, and conduct extensive simulations to evaluate the performance and efficiency of our design. The experimental results show that our design can provide satisfactory quality of audio-on-demand service with short startup latency and slight playback jitter. Extensive simulation results show that this design achieves a search success rate of 98 percent while reducing the search energy consumption by an order of magnitude compared with existing schemes.

Audio represents one of the most appealing yet least exploited modalities in wireless sensor networks, due to the potentially extremely large data volumes and limited wireless capacity. Therefore, how to effectively collect audio sensing information remains a challenging problem. In this paper, we propose a new paradigm of audio information collection based on the concept of audio-on-demand. We consider a sink-free environment targeting for disaster management, where audio chunks are stored inside the network for retrieval. The difficulty is to guarantee a high search success rate without infrastructure support. To solve the problem, we design a novel replication algorithm that deploys an optimal number of O(√n) replicas across the sensor network. We prove the optimality of the energy consumption of the algorithm, and use real testbed experiments and extensive simulations to evaluate the performance and efficiency of our design. The experimental results show that our design can provide satisfactory quality of audio-on-demand service with short startup latency and slight playback jitter. Extensive simulation results show that this design achieves a search success rate of 98% while reducing the search energy consumption by an order of magnitude compared with existing schemes.