Towards the Adaptive Identification of Walkers: Automated Feature Selection of Footsteps Using Distinction-sensitive Lvq

We applied a method called Distinction-Sensitive Learning Vector Quantization (DSLVQ) to the classification of footsteps. The measurements were made by a pressure-sensitive floor, which is part of the smart sensing living room in our research laboratory. The aim is to identify walkers based on their single footsteps. DSLVQ is an extended version of Learning Vector Quantization (LVQ), and it can be used for automated feature scaling and selection during the training of an LVQ codebook. The method shows improvements in the classification accuracies compared to a standard LVQ. In addition, due to its capability of automated input pruning, discarding the non-informative features, it was able to detect automatically the most significant features from a large set of features. This is important in an adaptive identification system, where the informative features might change.