AI for Crowdsensing: Fine-Turned GPT2 Model for City-Level Wi-Fi Fingerprinting Based on Crowdsourced Public Mobility Data

City-level indoor localization (CIL) systems have long faced significant challenges in large-scale and diverse urban environments, primarily due to the low accuracy of navigation database generation, the inherent uncertainty in crowdsourced public mobility data, and the limited generalization capabilities of fingerprinting methods. To address these issues, this paper introduces the AI for crowdsensing framework, which leverages large AI models to fully account for the uncertainty present in crowdsourced mobility data from a wide range of public users, thereby enabling accurate building and floor detection tasks across varied indoor scenarios at the same time. The algorithm extracts key features related to users' indoor-outdoor transitions and vertical movements (such as taking elevator), facilitating the effective construction of building-level and floor-level navigation databases, respectively. Notably, we employ the Generative Pre-trained Transformer 2 (GPT-2) network as the foundational model and enhance its building and floor detection capabilities through a fine-tuning strategy based on generated building and floor datasets, taking into account complex wireless context and spatial-temporal characteristics of human mobility. This approach enables the matching of indoor mobility data without GNSS references, thereby improving the accuracy and reliability of the generated indoor navigation database. Extensive real-world experiments conducted across more than 30 buildings from 3 different cities demonstrate that the proposed AI for crowdsensing framework achieves an effective and accurate CIL, with the average building and floor detection precision of 95.74% and 96.35% respectively, which surpasses state-of-the-art (SOTA) physical and AI methods.