A low-complexity activity classification algorithm with optimized selection of accelerometric features

Activity classification consists in detecting and classifying a sequence of activities, choosing from a limited set of known activities, by observing the outputs generated by (typically) inertial sensor devices placed over the body of a user. To this end, machine learning techniques can be effectively used to detect meaningful patterns from data without explicitly defining classification rules. In this paper, we present a novel Body Sensor Network (BSN)-based low complexity activity classification algorithm, which can effectively detect activities performed by the user just analyzing the accelerometric signals generated by the BSN. A preliminary (computationally intensive) training phase, performed once, is run to automatically optimize the key parameters of the algorithm used in the following (computationally light) online phase for activity classification. In particular, during the training phase, optimized subsets of nodes are selected in order to minimize the number of relevant features and keep a good compromise between performance and time complexity. Our results show that the proposed algorithm outperforms other known activity classification algorithms, especially when using a limited number of nodes, and lends itself to real-time implementation.