Artificial Neural Networks in Motion Analysis-Applications of Unsupervised and Heuristic Feature Selection Techniques.

Marion Mundt,Franz Bamer,Bernd Markert,Sina David,Arnd Koeppe

doi:10.3390/s20164581

Abstract

The use of machine learning to estimate joint angles from inertial sensors is a promising approach to in-field motion analysis. In this context, the simplification of the measurements by using a small number of sensors is of great interest. Neural networks have the opportunity to estimate joint angles from a sparse dataset, which enables the reduction of sensors necessary for the determination of all three-dimensional lower limb joint angles. Additionally, the dimensions of the problem can be simplified using principal component analysis. Training a long short-term memory neural network on the prediction of 3D lower limb joint angles based on inertial data showed that three sensors placed on the pelvis and both shanks are sufficient. The application of principal component analysis to the data of five sensors did not reveal improved results. The use of longer motion sequences compared to time-normalised gait cycles seems to be advantageous for the prediction accuracy, which bridges the gap to real-time applications of long short-term memory neural networks in the future.

Highlights

The analysis of human motion is of great interest to many different applications: it can be used to identify injury risk or to increase performance during sports-related tasks [1], and in many clinical applications [2,3]
We hypothesised that the use of three instead of five sensors placed on the pelvis and lower limbs would not decrease the prediction accuracy of the neural network significantly and that the application of a Principal component analysis (PCA) would have a regularising effect on the network
The strongest correlations between the sensor readings and the first PCA component were found for the pelvis angular rate around the medio-lateral (z) axis (r = −0.716) and vertical (y) axis (r = 0.746), both thigh sensors’ vertical (y) axes and both shank sensors’ medio-lateral (z) accelerations

Summary

Introduction

The analysis of human motion is of great interest to many different applications: it can be used to identify injury risk or to increase performance during sports-related tasks [1], and in many clinical applications [2,3]. With regard to daily life, wearable technology has shown its feasibility in both sports [4] and clinical applications [5]. Using these systems might help to prevent injuries or the onset of motion related diseases. For this purpose, easy-to-use feedback systems are necessary to inform the user of potential risks [6,7]. Lebleu et al [8] stated an error of less than 5◦ to be acceptable for clinical gait analysis, but the errors reported using IMU systems to range between

Objectives

Results

Discussion

Conclusion