Online-Adaptive Deep Learning for Multi-Sensor WiFi Indoor Localization

dc.contributor.advisorSohn, Gunho
dc.contributor.authorAhmad, Afnan
dc.date.accessioned2026-11-06T16:19:08Z
dc.date.available2026-11-06T16:19:08Z
dc.date.copyright2026-08-20
dc.date.issued2026-11-06
dc.date.updated2026-11-06T16:19:07Z
dc.degree.disciplineEarth & Space Science
dc.degree.levelDoctoral
dc.degree.namePhD - Doctor of Philosophy
dc.description.abstractIndoor positioning is constrained by the instability of WiFi signals, which degrade as building infrastructure changes and force the periodic re-surveys that limit practical fingerprinting deployment. This thesis develops an adaptive indoor localization approach through four contributions. An augmented reality framework built on the Microsoft HoloLens 2 automates centimeter-accurate ground-truth collection for WiFi fingerprinting, decoupling collection from runtime positioning and scaling across buildings without permanent infrastructure. The framework yields a longitudinal multimodal dataset spanning four York University buildings, covering 28,822 m² with 847 reference points and 239 access points, and combining WiFi, IMU, and magnetometer data collected over 2021–2022. A VAE-LSTM architecture jointly models spatial uncertainty and temporal movement dynamics from 20-second windows of sensor readings. The Online Feature Adaptation Module (OFAM) adapts the fusion weights of the WiFi, IMU, and magnetometer inputs during deployment from streaming sensor data, increasing the inertial contribution during WiFi degradation, while the network parameters remain fixed and no labeled position updates are required. With OFAM, the system achieves 0.54 m mean absolute error across the four York University buildings, a 68% improvement over WiFi-only baselines, with 5 ms adaptation overhead suitable for real-time deployment. Cross-year evaluation on the York University testbed shows 4% accuracy degradation compared with 22% without adaptation. The benefit reproduces on the independent IPIN 2016–2017 data, against which OFAM was never tuned, with 6% cross-year degradation against 29% for the static model.
dc.identifier.urihttps://hdl.handle.net/10315/44358
dc.languageen
dc.rightsAuthor owns copyright, except where explicitly noted. Please contact the author directly with licensing requests.
dc.subjectArtificial intelligence
dc.subjectGeographic information science and geodesy
dc.subjectComputer science
dc.subject.keywordsIndoor localization
dc.subject.keywordsWiFi fingerprinting
dc.subject.keywordsSensor fusion
dc.subject.keywordsOnline adaptation
dc.subject.keywordsVariational autoencoders
dc.subject.keywordsLong short-term memory networks
dc.subject.keywordsAugmented reality
dc.subject.keywordsMultimodal datasets
dc.subject.keywordsInertial measurement units
dc.subject.keywordsTemporal drift
dc.titleOnline-Adaptive Deep Learning for Multi-Sensor WiFi Indoor Localization
dc.typeElectronic Thesis or Dissertation

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
Ahmad_Afnan_2026_PhD.pdf
Size:
2.05 MB
Format:
Adobe Portable Document Format

License bundle

Now showing 1 - 2 of 2
Loading...
Thumbnail Image
Name:
license.txt
Size:
1.87 KB
Format:
Plain Text
Description:
Loading...
Thumbnail Image
Name:
YorkU_ETDlicense.txt
Size:
3.39 KB
Format:
Plain Text
Description: