Deep convolutional neural networks for motion instability identification using kinect

Evaluating the execution style of human motion can give insight into the performance and behaviour exhibited by the participant. This could enable support in developing personalised rehabilitation programmes by providing better understanding of motion mechanics and contextual behaviour. However, performing analyses, generating statistical representations and models which are free from external bins, repeatable and robust is a difficult task. In this work, we propose a framework which evaluates clinically valid motions to identify unstable behaviour during performance using Deep Convolutional Neural Networks. The framework is composed of two parts; 1) Instead of using the whole skeleton as input, we divide the human skeleton into five joint groups. For each group, feature encoding is used to represent spatial and temporal domains to permit high-level abstraction and to remove noise these are then represented using distance matrices. 2) The encoded representations are labelled using an automatic labelling method and evaluated using deep learning. Experimental results demonstrates the ability to correctly classify data compared to classical approaches.

1 Introduction

There has been significant interest in digital analysis methods for detection and quantification of human motion for use in electronic health interventions [1]. This, in part, is due to the increased availability of low-cost multi-modality marker-less capturing devices. Sensor technology (e.g. Microsoft Kinect) offers new dimensions by harnessing multiple techniques such as feature extraction and encoding. This has been observed in the work of Bigy et al. [2], they proposed a technique for recognising posture and Freezing of Gait in those with Parkinsons disease to aid in detecting trips and falls within the home. Yang et al. [3] implemented a framework that extracts both depth and colour image data from the Kinect to assess the posture of participants when performing standing balance, the framework allowed for detection of subtle directional changes such as postural sway.