Crowdsensing Based Indoor Signal Map Reconstruction and Localization with Fusion Learning and Building Structural Information

Wireless signal fingerprint-based localization methods have gained significant attention for their superior performance, but their large-scale deployment is hampered by the high cost of building location-signal fingerprint databases. With the proliferation of mobile smart devices, crowdsensing-a novel Internet of Things (IoT)-driven paradigm-offers a low-cost solution for constructing such databases, promising to advance largescale localization. However, crowdsensing suffers inherent limitations: user trajectories, characterized by hot-spot aggregation and event-driven behavior, fail to fully cover target areas, leading to sparse or missing data in some regions. Additionally, indoor building structures significantly affect wireless signal propagation and energy attenuation. To address these challenges, this paper proposes a wireless signal map reconstruction method using limited observed samples, aiming to enable low-cost locationsignal fingerprint database construction and accurate indoor localization. The method innovatively embeds building structure information and integrates self-representation learning with generative adversarial learning, where self-learned features guide the adversarial process to achieve high-quality map inference under few-sample conditions. Finally, an indoor Bluetooth localization system was deployed in a campus teaching building for performance evaluation. Extensive comparative experiments with traditional methods validate that the proposed method exhibits significant advantages in both signal reconstruction quality and localization accuracy.