MapICT: Unsupervised Radio-Map Learning From Imbalanced Crowd-Sourced Trajectories

The indoor localization based on WiFi fingerprints has attracted great attention, but the most significant and challenging work is the construction of the radio map without the need of site survey. With the rapid popularization of mobile and wearable devices, the method of collecting massive trajectory data by crowd-sourcing is becoming increasingly convenient and cost-effective, making unsupervised radio-map learning from unlabeled trajectories possible and promising. However, these crowd-sourced trajectories have low qualities and tend to present an imbalanced distribution due to the presence of hot spots, restricted areas in the building, as well as the random movement of participants. To address above challenges, MapICT scheme is proposed in this paper, using low-quality trajectory data to build radio map without site survey. First, Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm is suggested to extract target fingerprint vertexes, so as to balance the distribution of trajectory data. Second, target edges that represent relationship between vertexes are extracted by inertial data. Finally, a two-dimensional radio map can be constructed by vertexes and edges. Through simulation experiments and experiments in actual environment, the MapICT scheme proves to be feasible and effective, with the locating accuracy in teaching building and in mall being 11.98% and 7.38% respectively higher than that achieved by existing methods.

Automated summary

This paper proposes a MapICT scheme that utilizes low-quality trajectory data to construct a radio map. The proposed scheme employs the Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm to extract target fingerprint vertexes, balances the distribution of trajectory data, and extracts target edges using inertial data, thereby constructing a two-dimensional radio map. Simulation experiments and experiments in actual environments demonstrate the feasibility and effectiveness of the MapICT scheme, which achieves higher locating accuracy compared to existing methods, with a 11.98% improvement in a teaching building and a 7.38% improvement in a mall.