ENHANCED WI-FI-BASED INDOOR
LOCALIZATION
SYSTEM USING MACHINE LEARNING

Dataset: UJIIndoorLoc

Collected at Universitat Jaume I (2013) across ~110,000 m² campus with 3 multi-floor buildings. Features 19,937 training samples and 1,111 validation samples, captured with 25 Android devices by 20+ users in diverse environments including corridors, offices, labs, and classrooms.

Gaussian Kernel Fusion

Each neighbor i is assigned weight wᵢ = exp(−dᵢ²/2σ²) where σ is computed adaptively as max(median(d₁…dₖ), 0.5). By assigning exponentially higher weights to the nearest fingerprints in RSSI signal space, the system accurately reflects the spatial correlation between adjacent locations — preserving the natural relationship that physically close positions share similar signal patterns.

Strengths

• Interpretable, mathematically explainable predictions
• Tolerant of sparse WAP coverage where 55 WAPs have 0% coverage
• Rare high-variance WAPs leveraged as localization landmarks
• Per-building/floor models eliminate cross-floor confusion

Limitations & Future Work

• Limited generalization to new environments
• RSSI instability from walls, people, and multipath propagation
• KNN distance computation grows with database size
• Advanced feature engineering for WAP selection using mutual information and ANOVA F-score ranking
• Generalization of the Model and Algorithmic Comparison by trying different algorithms such as RF, SVM, and Neural Network
• Real-time implementation and adaptive fingerprinting on mobile devices with live Wi-Fi measurements