A Hybrid Model Based on Constraint OSELM, Adaptive Weighted SRC and KNN for Large-Scale Indoor Localization

In this paper, a novel hybrid model based on the constraint online sequential extreme learning machine (COSELM) classifier with adaptive weighted sparse representation classification (WSRC) and k nearest neighbor (KNN) is proposed for the WiFi-based indoor positioning system. It is referred to as A Fast-Accurate-Reliable Localization System (AFARLS). AFARLS exploits the speed advantage of COSELM to reduce the computational cost, and the accuracy advantage of WSRC to enhance the classification performance, by utilizing KNN as the adaptive sub-dictionary selection strategy. The understanding is that the original extreme learning machine (ELM) is less robust against noise, while sparse representation classification (SRC) and KNN suffer a high computational burden when using the over-complete dictionary. AFARLS unifies their complementary strengths to resolve each other’s limitation. In large-scale multi-building and multi-floor environments, AFARLS estimates a location that considers the building, floor, and position (longitude and latitude) in a hierarchical and sequential approach according to a discriminative criterion to the COSELM output. If the classifier result is unreliable, AFARLS uses KNN to achieve the best relevant sub-dictionary. The sub-dictionary is fed to WSRC to re-estimate the building and the floor, while the position is predicted by the ELM regressor. AFARLS has been verified on two publicly available datasets, the EU Zenodo and the UJIIndoorLoc. The experimental results demonstrate that AFARLS outperforms the state-of-the-art algorithms on the former dataset, and it provides near state-of-the-art performance on the latter dataset. When the size of the dataset increases remarkably, AFARLS shows that it can maintain its real-time high-accuracy performance.