Large-Scale Indoor Localization via Outdoor Crowdsourcing Trajectories on Ride-Hailing Platform

Indoor positioning is critical for a variety of services, including ride-hailing. However, existing large-scale fingerprint-based indoor positioning systems face significant challenges due to high deployment costs, temporal instability and limited accessibility, making them impractical for widespread use. In this paper, we propose a novel approach to indoor positioning that leverages fingerprints only sampled outdoors, which can be collected through crowdsourcing within a ride-hailing platform. This approach significantly reduces deployment costs, enables timely updates to the fingerprint set, and provides unprecedented accessibility. We address three key challenges in this system, including using outdoor fingerprints to estimate indoor position, abnormal Access Points (APs), and existence of "blackholes" where overheard APs have no fingerprint. Our implementation, built on the DiDi ride-hailing platform, is evaluated through extensive experiments with 122 million orders across 13 million devices in multiple cities. The results demonstrate that our system achieves a significant reduction of 4.35m in pickup position error compared to existing efforts, showcasing its potential for large-scale adoption.