A multi-view architecture for the SHL challenge

Abstract

To recognize locomotion and transportation modes in a user-independent manner with an unknown target phone position, we (team Eagles) propose an approach based on two main steps: reduction of the impact of regular effects that stem from each phone position, followed by the recognition of the appropriate activity. The general architecture is composed of three groups of neural networks organized in the following order. The first group allows the recognition of the source, the second group allows the normalization of data to neutralize the impact of the source on the activity learning process, and the last group allows the recognition of the activity itself. We perform extensive experiments and the preliminary results encourage us to follow this direction, including the source learning to reduce the phone position's biases and activity separately.