Slow Feature Preprocessing in Deep Neural Networks for Wearable Sensor-Based Locomotion Recognition

Lukas Günthermann, Lloyd Pellatt, Daniel Roggen

Abstract: This paper demonstrates how Slow Feature Analysis (SFA) can be used to transform sensor data before it is classified using a deep neural network. Slow features is concept originating from neurosciences which posits that a function can be regressed to fast changing sensory input to obtain a slower output, commensurate to the dynamics of the physical phenomena. Such dimensionality reduction is a common preprocessing technique. We explore this in the recognition of modes of locomotion and transportation from mobile phone motion sensors. Using a well-established deep learning activity recognition algorithm (DeepConvLSTM), we systematically investigate combinations of slow features as an additional layer, or substituting a layer in the deep network. The data used consists of 5s windows captured by a smartphone acceleration sensor carried at the hips. The results show that SFA can be used to improve the classification performance of the DeepConvLSTM network on an acceleration signal by up to 2.2%.