Recent years have witnessed the emerging development of single-antenna wireless respiration detection that can be integrated into IoT devices with a single transceiver chain. However, existing single-antenna-based solutions are all limited by the short sensing range within 2-4 m due to noise interference, which makes them difficult to be adopted in most room-scale scenarios. To deal with this dilemma, we propose a room-scale, noise-resistance, and accurate respiration monitoring system, named Wi-Cyclops , 1 which captures CSI changes induced by respiratory movements only via one antenna on commercial WiFi devices. To push the limits of effective sensing distance, we innovatively supply a new perspective to review the CSI samples along the sub-carrier dimension. From this dimension, we find that the interrelationship between sub-carriers with different timestamps still shows a high correlation even though the SNR decreases. Based on that, we analyze the noise characteristics along the sub-carrier dimension and correspondingly design a series of denoising schemes. Specifically, we carefully design a PCA-based denoising method to filter out ambient noises. After that, considering the low distribution densities of the AGC-induced noise, we then remove it by optimizing the DBSCAN denoising method with the K-Means-based adaptive radius search. Extensive experiments demonstrate that our system can work effectively in three typical family scenarios. Wi-Cyclops can achieve 98% accuracy even when the person is 7 m away from the transceiver pair. Compared with the start-of-art single-antenna-based approaches in real scenarios, Wi-Cyclops can improve the sensing range from 3 m to 7 m, which can meet the requirements of room-scale respiration monitoring. Additionally, to show the high compatibility with smart home devices, Wi-Cyclops is deployed on seven commercial IoT devices and still achieves a low average absolute error with 0.41 bpm.

The sensing capability of a Frequency Modulated Continuous Wave (FMCW) radar system is primarily determined by the hardware. However, we found that the mappings from high-resolution radar to low-resolution radar are not arbitrary, especially in the range-azimuth domain. We exploit such mappings to elevate the sensing power of an FMCW sensing system without modifying the hardware using a data-driven super-resolution approach. By training a super-resolution deep neural network with paired low-resolution range-azimuth maps and their corresponding high-resolution range-azimuth maps, we show that the trained super-resolution network can be used as a generic FMCW radar pre-processing module which improves the effective range resolution and the angular resolution of the raw inputs. The proposed super-resolution approach is not hardware dependent due to our proposed down-sampling data generation method. Downstream sensing tasks can be readily benefited from the elevated range-azimuth maps. We evaluated our approach on public and private FMCW human activity recognition (HAR) datasets and saw significant accuracy improvement.

Crowdsourcing is considered an efficient and promising paradigm for constructing large-scale signal fingerprint radio maps due to the proliferation of Wi-Fi-enabled devices. However, a crowdsourced indoor positioning system (IPS) has to handle diverse devices and the inherent heterogeneity in received signal strength (RSS) measurements. To address the device heterogeneity problem, differential fingerprinting methods have been explored, which mitigate the device characteristics that cause RSS from different commercial devices to report differently. In this article, we focus on mean differential fingerprinting (MDF) that produces the differential fingerprints by subtracting the mean RSS value of all access points from the original RSS fingerprints. We study the localization performance of the MDF method by means of the Cramér–Rao lower bound (CRLB) and show analytically that it outperforms another method that addresses device diversity. Furthermore, we evaluate the localization accuracy of existing solutions using real-life Wi-Fi RSS data sets collected by multiple consumer devices. The experimental results confirm our analytical findings and demonstrate the effectiveness of the MDF method to mitigate device diversity, as well as other factors that affect the RSS readings, including the device carrying mode and power control schemes of the Wi-Fi infrastructure, thus contributing to the wider adoption of crowdsourced IPS.

Predicting signals propagated indoors is the key to radio map building for indoor positioning. Conventional models allow limited learning from well-surveyed fingerprints and can not transfer to new areas. Inspired by recent advances in computer vision, this work frames the radio map building problem as an image generation problem and explores how to learn signal propagation from a massive amount of fingerprints collected from multiple floors of multiple buildings. We demonstrate a generative framework we call GenLoc to model signal propagation implicitly in a data-driven way. Ultimately, GenLoc provides a generalised model that can generate radio maps for new buildings readily. Meanwhile, GenLoc can optionally incorporate floor plan information to condition the generation of a radio map. Evaluation across multiple buildings shows that, compared to conventional methods, GenLoc achieves a decisively improved (15%) positioning accuracy at a lower generation cost.

Crowdsourcing radio signal fingerprints to build a radio map for indoor positioning system is an emerging alternative to conventional labour-costly manual survey. However, existing crowdsourced systems heavily rely on ground-truth location inputs or unrealistic constraints on the contributors, deterring a wider adaption of crowdsourced systems. Our work exploits three generic constraints of mobile data to retrieve the locations of the crowdsourced fingerprints and builds a completely passive indoor positioning system that assumes no manual intervention or unnatural constraints on the contributors. The proposed system was further evaluated in the Museum of Archaeology and Anthropology (MAA) with passively crowd- sourced data contributed by actual visitors while visitors can behave naturally without catering to crowdsourcing. Results show that the proposed system can achieve positioning accuracy comparable to traditional manual survey-based system with essentially no extra manual effort.

In indoor positioning, signal fluctuation is highly location-dependent. However, signal uncertainty is one critical yet commonly overlooked dimension of the radio signal to be fingerprinted. This paper reviews the commonly used Gaussian Processes (GP) for probabilistic positioning and points out the pitfall of using GP to model signal fingerprint uncertainty. This paper also proposes Deep Gaussian Processes (DGP) as a more informative alternative to address the issue. How DGP better measures uncertainty in signal fingerprinting is evaluated via simulated and realistically collected datasets.

In online customer service applications, multiple chatbots that are specialized in various topics are typically developed separately and are then merged with other human agents to a single platform, presenting to the users with a unified interface.Ideally the conversation can be transparently transferred between different sources of customer support so that domain-specific questions can be answered timely and this is what we coined as a Bot-Agent symbiosis.Conversation transition is a major challenge in such online customer service and our work formalises the challenge as two core problems, namely, when to transfer and which bot or agent to transfer to and introduces a deep neural networks based approach that addresses these problems.Inspired by the net promoter score (NPS), our research reveals how the problems can be effectively solved by providing user feedback and developing deep neural networks that predict the conversation category distribution and the NPS of the dialogues.Experiments on realistic data generated from an online service support platform demonstrate that the proposed approach outperforms state-of-the-art methods and shows promising perspective for transparent conversation transition.

Signal fingerprinting is considered to be potential as the general indoor positioning solution since it does not require extra infrastructure or hardware modification to current customer smart devices. However, due to the complex nature of radio signal propagation, fingerprints are highly susceptible to both temporal and spatial indoor dynamics, rendering the positioning outcomes unreliable. Most recent works dedicated to efficient ways of building radio maps during the offline phase yet overlooked the localisability of fingerprints collected rather opportunistically at the online phase. In this paper, we introduce the use of pseudo-measurements and propose a fingerprint anomaly detection method that effectively evaluates the localisability of fingerprints while positioning. Empirical evaluations demonstrate that filtering out untrustworthy fingerprints can significantly improve the positioning accuracy.