Background

Wi-Fi–based sensing — extracting Channel State Information (CSI) from commodity radios — has matured rapidly and moved from laboratory demos toward standardized support and commercial products. The IEEE 802.11bf amendment (WLAN Sensing) is step in that direction - accelerating interoperable sensing features in Wi-Fi hardware and protocols. Commercial vendors (notably Huawei) are already offering CSI-based sensing solutions for smart-building energy saving and intrusion detection, illustrating practical viability outside academia. AI/ML advances have improved detection and classification performance for tasks such as occupancy, fall detection, gesture recognition, and vital-sign (breathing) monitoring; however, robustness in uncontrolled public spaces remains challenging due to interference, multipath, and domain shift.

This project focuses on reproducing and validating results on publicly available CSI datasets, for an industrial MSc thesis. In general, objective is to do a rigorous comparison of preprocessing and ML model choices, evaluation of cross dataset generalization, and delivery of reproducible code and baselines.

Aim

Evaluate the feasibility and limits of Wi‑Fi CSI sensing for relevant tasks for example in automotive using open datasets. Develop reproducible benchmarks and ML methods that assess the feasibility of this approach for the intended application and provide concrete guidance for future in‑field pilots. Specifically, validate prior results on open datasets, evaluate preprocessing and model choices to validate detection of presence, basic activity (sit/stand/fall-like events), and breathing signatures.

Objectives

• Literature survey and standards review (IEEE 802.11bf implications).
• Survey and catalog relevant public CSI datasets and summarize their metadata (device, carrier, sample rate, tasks).
• Implement a modular preprocessing pipeline (amplitude/phase sanitization, alignment, denoising, spectrogram conversion).
• Reproduce and benchmark 2–3 representative papers for intended use cases (occupancy, activity recognition, vital‑sign estimation) on available open datasets.
• Compare model families: 1D CNN / RNN on raw timeseries, and 2D CNN on spectrograms/time‑frequency transforms.
• Quantify cross‑dataset generalization and perform domain adaptation experiments (fine‑tuning, transfer learning).
• Deliver an open, well‑documented codebase and a final thesis report with reproducible notebooks and recommendations.
• Present the final findings internally at the company.

Study design

• Phase 1 — Survey & selection: choose 2–3 open datasets covering activities, presence, and breathing; extract metadata (AP/NIC type, subcarriers, sampling rate, recording layout).
• Phase 2 — Data Preprocessing & model implementation: implement reusable modules for amplitude/phase sanitization, per‑subcarrier filtering, windowing, spectrogram generation, and data augmentation. Implement ML models: e.g., 1D CNN, GRU, 2D CNN on spectrograms.
• Phase 3 — Reproduction & benchmarking: reproduce selected results, run standardized experiments with consistent splits, compute detection accuracy metrics.
• Phase 4 — Generalization & analysis: cross‑dataset transfer experiments, ablation of preprocessing choices, and sensitivity analysis to metadata differences (receiver count, subcarriers, band).
• Deliverables: Artifacts such as code repository with dataset preprocessing functions, ML model development, Jupyter Notebooks for reproduced experiments, and the final thesis document.

Suitability

• 1-2 Master thesis students with engineering background
• Good programming skills, statistics and ML
• Familiarity with a suitable programming language preferably Python
• Knowledge in ML & Deep-learning methods
• Experience with signal processing and analysis of time-series data

Questions? Please contact Jawwad Ahmed, Senior Data Scientist, jawwad.ahmed@autoliv.com