Multi-Floor Indoor Localization Scheme Using a Seq2Seq-Based Floor Detection and Particle Filter With Clustering

Abstract

In this study, we present an infrastructure-independent multi-floor indoor localization scheme that uses a deep learning (DL)-based floor detection method and a particle filter with clustering. To implement localization with limited measurement data, we incorporate the user's vertical motion information to initialize and optimize the system. This method assumes two prerequisites: the capability for rapid floor detection and extraction of vertical motion features. These ensure a correlation between vertical movement and two-dimensional location and can be efficiently integrated with map information. The proposed scheme has several notable features. First, we utilize the strong feature extraction capability of the sequence-to-sequence (Seq2Seq) model for sequential data to implement real-time step action prediction. We also develop a floor decision algorithm to extract vertical movement information from the step action sequence. The proposed floor detection method can track the floor regardless of the user's activities. Second, we configure calibration nodes (CN) on the map based on prior knowledge from the environmental information. By combining CNs with DL-based floor detection, we not only extend the particle filter to three-dimensional applications but also achieve calibration and repair of the localization. Third, we introduce a clustering method to improve localization accuracy and reduce computational complexity in uncertain measurements. The experimental results show that the Seq2Seq model has good robustness to noisy data, the proposed DL-based floor detection achieved an average floor number accuracy of 93.42% without restricting user behavior, and all the floor transitions were successfully recognized. Moreover, under the long path multi-floor scenario, our scheme achieved a localization accuracy of over 96% within a 2m error boundary.