Unsupervised Phase Learning and Extraction from Repetitive Movements

Abstract

Phase extraction from repetitive movements is one crucial part in various applications such as interactive robotics, physical rehabilitation, or gait analysis. However, pre-existing automatic phase extraction techniques are specific to a target movement due to some handcrafted-features. To make it more universal, a novel unsupervised-learning-based phase extraction technique is proposed. A neural network architecture and a cost function are designed to learn the concept of phase from records of a repetitive movement without any given phase label. The method is tested on a rat's gait cycle and a human's upper limb movement. The phases are successfully extracted at the sample level despite the variations in movement speed, trajectory, or subject's anthropometric features.