Interns: Haoyan ZengUG
Advisors: Aggelos Bletsas, Anastasios, Kleniatis, Georgios Andreadis
Develop a device-free indoor localization system that estimates human location (or occupancy) using passive RFID sensing. The project combines real-world RFID measurements, wireless propagation simulation, and machine learning to improve localization accuracy and generalization across different environments.
Week 1 Slideshow This Week's Work: -Introduced academic background and the Magic Room project, which uses batteryless RFID tags and machine learning for device-free human sensing. Explained the project’s sensing principles, its potential for localization and movement tracking, and the next step of developing an ML-ready dataset.
Week 2 Slideshow This Week's Work: -I examined how device-free localization is modeled in a physical room, including RF propagation, grid-based representations, ellipse models, and linear inverse methods. I also reviewed the limitations of existing localization approaches and proposed using machine learning to map RFID signal patterns to the number and locations of people.
Week 3 Slideshow This Week's Work: -I learned the practical details of RFID data collection, including coordinate-system calibration, time-window averaging, and the trade-off between collection speed and signal stability. I also studied how a MIMO-OFDM JRC system performs communication and radar sensing through DFT- and MUSIC-based range-angle processing.
Week 4 Slideshow This Week's Work: -My main goal this week was to implement the MUSIC algorithm for high-resolution range-angle estimation in the JRC system. I learned how to replace the existing DFT processing chain with covariance and MUSIC-spectrum blocks, and how to implement these as custom GNU Radio modules. I also explored using raw timestamped RFID logs to support more flexible machine-learning-based localization and people counting.
Week 5 Slideshow This Week's Work: -I successfully implemented and integrated the MUSIC algorithm into the MIMO-OFDM JRC processing chain. By comparing its range-angle heatmap with the original DFT method, I verified that MUSIC produced a sharper target peak and better localization resolution. I also developed a cascaded framework for the Magic Room project that combines room-specific RFID patterns with live raw-log measurements for more robust localization.
Week 6 Slideshow This Week's Work: -I studied the Geo2SigMap paper and explored how its cascaded signal-mapping framework could be adapted to the Magic Room project. I also examined the collected RFID dataset structure, including antenna–tag–frequency links, calibration coverage, and data availability.
Week 7 Slideshow This Week's Work: -I investigated a similar link-voting framework based on calibration reliability, observed signal anomalies, and physical relevance. However, applying it directly to the real dataset proved difficult because of sparse and inconsistent measurements, so I identified simulated data as a promising alternative for developing and validating the approach.
Week 8 Slideshow This Week's Work: -I ran two simulation methods and compared their results with real-world measurements. I then designed a link-voting score that combines calibration reliability, observed signal anomalies, and physical relevance derived from Sionna propagation paths to generate a candidate-location heatmap.
Week 9 Slideshow This Week's Work: -Rehearsed my final presentation and consolidated the existing simulation results, real measurements, and project findings. I standardized and compared baseline-relative RSSI and phase data from Sionna, MATLAB, and real experiments, while investigating scale alignment between simulated and measured signals. I also proposed future directions, including evaluating link reliability, refining the simulated signal scale, and modeling missing RFID channel effects without overfitting to a specific room.
Week 10 Slideshow This Week's Work: -