Indoor Corner Detection and Matching from Crowdsourced Movement Trajectories

Indoor landmarks, like corners, staircases and etc, play an important role in crowdsourcing-based indoor localization systems. This paper studies the problem of indoor corner detection and matching from crowdsourced movement trajectories. For corner detection, we adopt a machine learning approach by training a corner detector with both time and frequency features. For corner matching, we first apply the multidimensional scaling technique for matrix dimensionality reduction and then propose an improved K-means algorithm to obtain an intermediate matching result for each feature dimension. We also propose a voting algorithm to obtain the final matching result for each corner sample based on its all intermediate dimension matching results. Experiment results show that the machine learning-based corner detection can achieve much better detection performance, compared with the existing algorithms based on signal change detection. For corner matching, the proposed scheme can achieve high matching accuracy and the constructed corner fingerprints can achieve the nearest distance with their respective reference corner fingerprints.