Hybrid and Convolutional Neural Networks for Locomotion Recognition

This paper explores the relevance of an approach based exclusively on deep neural networks for locomotion recognition. This work is done within the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge as team Power of Things. Provided data used during the experiments is part of the SHL dataset for which we emphasize the adaptability to different applications of the ubiquitous computing. This quality emerges from the broad spectrum of modalities that this dataset encompasses, they are 16 in total. More than 500 different convolutional and hybrid architectures are evaluated, and a Bayesian optimization procedure is used for hyper-parameters space exploration. The influence of these hyper-parameters on performances is analyzed using the fANOVA framework. Best models achieve a recognition rate of about 92% measured by the f1 score.