SYNTHETIC WI-FI FINGERPRINT GENERATION AND INDOOR LOCALIZATION UNDER WI-FI SCAN THROTTLING CONSTRAINT

Abstract

This project addresses the growing demand for accurate indoor localization in environments where GPS is ineffective and Wi-Fi scan availability is constrained by modern smartphone operating systems. Specifically, it tackles the challenge introduced by Android's Wi-Fi scan throttling policies, which limit scan frequency and thereby degrade the performance of traditional Wi-Fi fingerprint-based positioning systems. The key objective of the project was to design and evaluate a hybrid indoor positioning system capable of operating under Wi-Fi scan throttling. The proposed solution combines synthetic Wi-Fi fingerprint data generation using a Conditional Denoising Diffusion Probabilistic Model (cDDPM) with Pedestrian Dead Reckoning (PDR) based on CNN+LSTM deep learning models for IMU sensor data. A fusion strategy then integrates these two modalities to deliver a robust indoor localization system. The methodology included: (1) generating synthetic RSSI data to augment real-world datasets, (2) constructing a localization pipeline using k-Nearest Neighbors (kNN) for Wi-Fi positioning, (3) building a displacement prediction model using CNN+LSTM for IMU-based tracking, and (4) implementing a throttling-aware fusion algorithm to simulate real-world constraints. The evaluation results showed that diffusion-generated synthetic data can significantly reduce localization errors—by up to 22% in low-data scenarios. The hybrid model maintained continuous trajectory estimates and partially mitigated PDR drift despite infrequent Wi-Fi corrections, validating the effectiveness of the fusion approach under Android scan throttling. This project exemplifies the design, implementation, and evaluation of a practical, computing-based solution to a real-world systems limitation.