Xusheng Zhang

High-fidelity mmWave human point clouds are pivotal for applications like fall detection, people counting, and multi-person pose estimation. However, commercial radars are often hindered by multipath interference in cluttered indoor environments, which introduces ghost targets and distorts human structures. Prior attempts at mitigation either rely on temporal continuity, environmental priors, or data-hungry learning models, all of which suffer from latency, fragility, or high overhead. We present PathClean, the first system to eliminate multipath interference directly during point cloud generation instead of post-processing. PathClean leverages the inherent Angle of Arrival (AoA)-Angle of Departure (AoD) consistency of direct paths to instantly suppress multipath, while addressing the challenges of dense matching ambiguity, hardware imperfections, and real-time constraints. Implemented on commodity radars, PathClean removes 98.9% of multipath points, while retaining 97.6% of true human points. Extensive experiments across seven environments and three case studies—fall detection, people counting, and multi-person pose estimation—demonstrate that PathClean is robust, scalable, and practical for real-world deployment.

Millimeter-wave (mmWave) radar has been widely studied for ubiquitous human sensing, but its deployment in homes remains limited by the need for additional power, connectivity, and dedicated hardware. A formative user study with 236 participants suggests that integrating sensing capabilities into existing home appliances may offer a more acceptable deployment model: although many participants were reluctant to install standalone sensors, 92% expressed willingness to use appliance-integrated sensing services. Motivated by this finding, we present mmUbiSense, a system that integrates low-cost commercial radar chips into everyday appliances. This integration, however, exposes the radar to mechanical vibrations generated during appliance operation, which can substantially distort both micro-motion signals and macro-motion point clouds. Existing compensation methods typically rely on high-end radar hardware and are therefore difficult to apply under the cost constraints of appliance integration. mmUbiSense addresses the two sensing regimes separately through a physics-guided framework. For micro-motion sensing, we develop a spectral-matching estimator that models the interaction between appliance vibration and physiological motion in the complex radar signal, allowing respiration and heart rates to be recovered without relying on fine spatial resolution. For macro-motion sensing, we introduce a spatio-temporal clustering method that distinguishes vibration-induced ghost points from human motion by comparing Doppler responses with actual spatial displacement. We evaluate mmUbiSense on air conditioners, washing machines, and range hoods across vital sign monitoring, pose estimation, fall detection, and gesture recognition tasks. The results show consistent improvements over the corresponding baselines in the evaluated settings, supporting the feasibility of integrating reliable mmWave sensing into everyday household appliances using low-cost hardware.

Gait recognition serves as a foundation for numerous applications, ranging from early disease diagnosis to continuous user authentication. While traditional methods rely on cameras or wearables to achieve gait recognition, RF-based solutions have emerged in recent years that better preserve privacy and do not require dedicated wearables or sensors. Among the RF signals explored, WiFi-based solutions are particularly promising due to the ubiquitous deployment of WiFi infrastructure. However, existing WiFi-based methods either require multiple WiFi links or constrain users to walk along predefined trajectories. In addition, they typically require extensive data collection to train the recognition model. These limitations significantly impede the practical deployment and broader adoption of WiFi-based gait recognition systems. In this work, we present WiSiGait, a one-shot gait recognition framework that operates using only a single WiFi link and does not require users to follow any predefined trajectories. To realize WiSiGait, we extract trajectory-independent gait features that can be captured using a single WiFi link. For identification, we explore Siamese neural networks to realize one-shot gait recognition, eliminating the need for extensive pre-training data collection. Extensive experiments demonstrate that our system achieves robust gait recognition across arbitrary trajectories using only a single WiFi link. We believe this work represents a significant step toward the practical deployment of ubiquitous gait recognition. A demo video is available at: https://youtu.be/eTOkG5O3CCE.

Continuous, fine-grained cardiac monitoring during sleep is vital for cardiovascular health. Emerging as a promising alternative to cumbersome wearables, mmWave radar offers a contactless, privacy-preserving modality capable of capturing minute skin vibrations. However, transitioning mmWave radar from labs to real-world bedrooms remains challenging. Our rigorous empirical evaluation of state-of-the-art systems reveals a severe reliability gap: valid Inter-Beat Intervals (IBI) are captured for only 24%-42% of the night within an error threshold of 50 ms. Since clinical HRV analysis requires millisecond-level precision (e.g., RMSSD changes), such coarse and intermittent tracking renders existing methods medically insufficient. We attribute this failure to two fundamental hurdles: (i) instability driven by overwhelming respiratory interference and postural diversity, and (ii) signal entanglement in co-sleeping scenarios where angular separation defies hardware resolution. To bridge this gap, we present DuBCG. First, to tackle instability, we introduce a Head-Facing configuration to capture the Radar Ballistocardiogram (R-BCG). By leveraging the body's longitudinal recoil, this approach achieves orthogonal respiratory suppression and posture-invariant robustness. Second, to resolve entanglement, we propose the Multi-point Scattering Spatial Selectivity Model, theoretically proving that signal separability is achievable beyond physical resolution limits. Extensive real-world benchmarks demonstrate that DuBCG achieves a median IBI error of 8.4 ms—an 8× improvement over existing methods. This precision bridges the gap from existing intermittent monitoring to long-term, clinically comparable continuous HRV assessment.

MmWave radar has emerged as a promising technology for human sensing. While large bandwidths and multiple antennas enable high-quality point clouds for sensing applications, low-cost 24 GHz radars are restricted to 250 MHz bandwidth by regulations, leading to limited range resolution. In addition, the limited number of antennas, constrained by cost and device size, results in low angular resolution. Together, these limitations produce sparse point clouds that impede fine-grained human sensing. In this work, we present RaPoint, the first system to produce high-density point clouds from 24 GHz radars under regulatory bandwidth limits. RaPoint introduces a joint super-resolution modeling framework that leverages the MUSIC algorithm to jointly exploit super-resolution range and angle dimensions, together with the Doppler dimension, for distinguishing reflection points. To overcome the high computational complexity problem of the MUSIC algorithm, we convert the eigenvalue decomposition from the noise subspace to the sparse signal subspace, drastically reducing computation. We implement RaPoint on three representative mmWave radars. Extensive experiments demonstrate that RaPoint improves point cloud density by 9.45× and 24.10× over the state-of-the-art research and industrial baselines, respectively, and substantially boosts downstream sensing performance. We believe RaPoint takes a significant step toward fine-grained human sensing with low-cost, bandwidth-limited 24 GHz radars, enabling affordable and ubiquitous radar sensing.

WiFi sensing has garnered significant attention from both academic and industrial communities, largely due to the widespread deployment of WiFi infrastructure. However, most existing WiFi sensing works rely on Channel State Information (CSI), which can only be extracted from very few commercial WiFi devices. The widespread adoption of new WiFi protocols, such as IEEE 802.11ac and 802.11ax, presents a valuable opportunity to leverage the widely available Beamforming Feedback Information (BFI) for WiFi sensing. Several studies have explored the potential of BFI-based WiFi sensing. However, these works are limited to a specific number of antennas and cannot achieve fine-grained BFI sensing across an arbitrary number of antennas. In this work, we design and implement FreeBFI, the first BFI-based WiFi sensing system that can work with an arbitrary number of antennas. FreeBFI fully exploits the channel information and the SNR information contained in BFI to establish the relationship between BFI and target motion across arbitrary antenna counts. Furthermore, to extract fine-grained motion information from the established relationship, FreeBFI smartly fuses the information from multiple antennas and proposes a novel optimization algorithm to enhance the motion signal. To showcase the sensing capability of FreeBFI, we select two representative WiFi sensing applications: respiration monitoring and gesture recognition. We conduct comprehensive experiments covering a wide range of antenna counts and test the performance of FreeBFI on various WiFi devices. Experimental results demonstrate that FreeBFI not only delivers accurate and robust sensing performance under arbitrary antenna counts but also enhances sensing accuracy as all antennas are utilized. For respiration monitoring, FreeBFI significantly extends the sensing range from 4 m to 8 m. For gesture recognition, FreeBFI improves complex gesture recognition accuracy by over 20%. We believe this work marks a significant step toward the broader adoption of WiFi sensing on next-generation WiFi devices.

Couples sharing a bed is a common practice in households, making sleep monitoring in multi-person scenarios a significant area of research. The high sensitivity of mmWave makes it a promising tool for monitoring vital signs during sleep. However, implementing mmWave-based solutions for such settings poses two major challenges. First, the range and angular resolution of commercial mmWave radar are still inadequate for handling such close-proximity scenarios where individuals share the same bed. Second, even with sufficient angular resolution, aligning the main lobe with one target usually faces strong interference from other targets located in the sidelobes. To address these challenges, we leverage antenna pattern nulls---directions with zero gain---to mitigate interference and overcome resolution limitations. Implementing null steering on practical mmWave radar, however, is challenging due to the human body being an extended target that spans a range of angles, as well as hardware imperfections that render traditional algorithms ineffective. In this paper, we present mmNull, the first null steering system designed for commercial mmWave radars. mmNull models the human body as an extended target rather than a single point, derives a closed-form solution to maximize target gain while suppressing interferences, and employs a heuristic algorithm to address hardware imperfections. Comprehensive experiments show that mmNull achieves a more than 5x improvement in accuracy compared to current state-of-the-art (SOTA) methods for monitoring respiration rate, heart rate, and heart rate variability (HRV) in multi-person scenarios.