Hien Vu

We developed a mmWave radar system that continuously monitors respiration of freely moving subjects. The core challenge is that subject motion destroys temporal correspondence in radar point clouds, making traditional approaches fail when targets are mobile. Our system uses a four-stage feedforward pipeline: signal processing for angularly-accurate 3D point cloud extraction from a low-cost radar, dual-branch pose tracking combining PointNet-based estimation with ConvLSTM-UNet trajectory prediction, anchor-based spatial encoding for temporal correspondence, and spatial-temporal deep learning for millimeter-scale respiration extraction. The system achieves less than 5 cm pose tracking error and less than 3 BPM respiration error on moving targets, with 3x to 4x improvement in downstream task performance over threshold-based detection. This work is currently under review.

MmCows is a multimodal sensing platform for continuous behavior monitoring in uncontrolled environments. The system integrates nine synchronized sensor modalities including UWB localization, IMU, multi-view RGB cameras, and environmental sensors. We developed centimeter-accurate UWB 3D tracking with dynamic anchor selection for NLOS mitigation, multi-view triangulation with automatic outlier rejection, and a multimodal fusion pipeline for real-time activity classification. The dataset spans 14 days with 4.8M frames. Published as a Spotlight paper at NeurIPS 2024.

eTag is an energy-autonomous wearable IoT sensor that operates indefinitely without battery replacement. It uses a two-hop communication topology (LF backscatter to LoRa to cloud) achieving greater than 99% uplink reliability. A shared-coil RF architecture multiplexes RFID scanning and wireless charging on a single antenna, reducing hardware size by 70%. The embedded firmware reduces energy consumption by 15x, and a closed-loop wireless charging controller handles 20 to 70 cm variable coupling distances. Seven prototypes ran continuously for over 2,000 hours. Published at ACM MobiCom 2023.

We implemented an Extended Kalman Filter on a RISC-V microcontroller to fuse data from a 9-DOF inertial measurement unit (accelerometer, gyroscope, magnetometer), achieving real-time, drift-corrected 3D pose estimation. The system was integrated with Parallel Tracking and Mapping (PTAM) for visual-inertial odometry in augmented reality applications. The EKF implementation was optimized for the constrained computational resources of the RISC-V platform and was subsequently integrated into a custom chip design.

We developed a high-energy wearable power system in a flexible, conformal form factor, capable of delivering 840W in repeated 5-second discharge pulses. The system uses a cascade multiphase DC-DC converter with a bidirectional buck-boost control policy that maintains uninterrupted power delivery under rapidly changing load conditions. The design was taken through the full product lifecycle from initial prototype through DoD qualification testing, ultimately passing all requirements and achieving successful technology transfer to a defense contractor. Several critical performance specifications were exceeded during qualification.

We designed and implemented a high-speed storage system consisting of an FPGA controller and 512GB of SLC-NAND flash memory across 16 ICs, meeting timing closure and signal integrity requirements for over 1,000 components. The project also included a dynamic voltage-scaling power backup system to protect data integrity during power loss events, extending capacitor lifespan for reliable operation.