ANN Feature Scaling based k-Nearest Neighbor Algorithm for Indoor Localization

The Radio Frequency (RF) fingerprint based localization system has become one of the most competitive solutions in the field of indoor localization. The k-Nearest Neighbor (kNN) algorithm is popularly used in the localization system, due to its simplicity for implementation. However, traditional kNN algorithm might fail to consider the variability of correspondence between geometrical location and the similarity of received-signal-strength (RSS) vectors using Euclid distance. In the light of feature scaling kNN (FS-kNN) and continuous feature scaling kNN (CFS-kNN) algorithms, a novel indoor localization algorithm is proposed in this paper. The algorithm maps the similarity of each pair of RSS vectors to their geometrical distance without the ambiguous boundary that might yield in FS-kNN. To avoid the ambiguous boundary, the proposed algorithm determines the weights continuously by an artificial neural network (ANN) that can be trained via some samples. The conditions to guarantee the validity of this algorithm are discussed as well. The proposed algorithm can be used in eigher gridless or grid based RSS matching methods of indoor localization. In our experiment, the performance of the proposed algorithm is better than RADAR, FS-kNN and CFS-kNN, in term of accuracy. Moreover, ANN is more flexible and reasonable than other fixed structure schemes to determine the weights used in the feature-scaling based algorithms. The algorithm is easy to realize in engineering as well.