Uncertainty based active learning with deep neural networks for inertial gait analysis
Alexander Vaith, Bertram Taetz, Gabriele Bleser · 2020
Inertial measurement units (IMUs) enable the capture of human motion in-field. This can be used in various analysis applications ranging from the medical domain over sports to daily activities. Manual data labelling for classification or regression tasks is often time-consuming and cumbersome, in particular when it comes to larger datasets. Active learning algorithms try to reduce the labelling cost, for instance via suggesting samples with high prediction uncertainty that should be explicitly labelled. In this work, we apply two probabilistic deep learning approaches on different state-of-the-art deep neural network structures for timeseries data, with uncertainty based measures to actively query sample labels. This is applied to gait phase classification using IMU data as inputs. We demonstrate the performance on a newly captured dataset, where we obtained high accuracy (up to 96%) with up to 43% fewer samples as compared to random sampling in an online setting. In an offline setting, we could extract heel strike and toe-off foot events with an accuracy of 99.9% using active learning strategies with up to 58% fewer samples.