Print
Univis
Search
 
FAU-Logo
Techn. Fakultät Willkommen am Institut für Informatik FAU-Logo

Projects

Deep Learning in Sensor-based Gait Analysis
  • Exemplary network architecture for a deep convolutional neural network estimating stride-specific biomechanical parameters

    This project is part of the EFI Moves initiative.

    Journal Articles
    Hannink, Julius; Kautz, Thomas; Pasluosta, Cristian; Gaßmann, Karl-Günter; Klucken, Jochen; Eskofier, Björn
    Sensor-Based Gait Parameter Extraction With Deep Convolutional Neural Networks
    IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 1, pp. 85-93, 2017 (BiBTeX, Who cited this?)
SensorDataToolbox

The SensorDataToolbox is a collection of algorithms and workflows that are used in a variaty of mobile inertial sensor data analysis applications. Implementations in the toolbox are mainly MATLAB code with occasional C/C++ implementations for performance reasons. The purpose for this project is a common code-base for the group that can also be used to archive implementations created in the various projects. Currently, the SensorDataToolbox contains generally applicable methods for

  • Various preprocessing steps (calibration, normalization, resampling, filtering, etc.)
  • Segmentation of sensor data (mulit-dim. subsequence dynamic time warping, sliding window)
  • Double-integration of inertial sensor data to arrive at sensor trajectories over time
  • Extraction of statistical parameters from segments of data e.g. defined by the segmentation

In the case of inertial sensor data from human gait, the prototypical application the toolbox is developed by, a set of biomechanical events (toe-off, heel-strike and mid-stance) can be extracted from the segmentation of continuous sensor data into strides. Furthermore, several biomechanical parameters can be extracted from the sensor trajectories for each stride (stride length, width etc.) or from the timings of individual gait events (stride, swing and stance time). In this respect, the toolbox also constitutes the code-base for the EFIMoves project and allows efficient deployment of graphical user interfaces that can be used in a clinical setting for advanced gait analysis.

This project is part of the EFI Moves initiative.

 

 

An example of a sequence of straight strides from inertial sensor data of human gait extracted with the SensorDataToolbox.
An example of events detected in a gait sequence: Toe-off (TO), heel-strike (HS) and mid-stance (MS)
Foot motion over one exemplary stride rendered based on the trajectory extracted with the SensorDataToolbox from inertial sensor data, click to play.
The analysis tool that is currently being developed in the EFIMoves project for advances gait analysis
Gait Dynamics

In this project we try to quantify the dynamics in human gait from mobile inertial sensor data captured at the feet. The measures we apply are generally drawn from nonlinear time series analysis, chaos theory and the analysis of dynamical systems.

As an exemplary application, we investigate the gait dynamics in Parkinson's disease patients. Depending on the progression of the disease, these patients may have problems of keeping the pace or experience motor fluctuations.

In particular, we try to quantify memory length scales in the signals as a measure with direct physical meaning as it is entangled with the patient's neuromuscular control system and can give insight into its function. We employ the reshape-scale method by Zandiyeh and Von Tscharner (“Reshape scale method: A novel multi scale entropic analysis approach,” Phys. A Stat. Mech. its Appl., 2013) and compute an entropy profile for a set of time scales within the signal. As a primary features, the entropic half life (EnHL) is extracted from these transitions: Short EnHL encodes a fast transition from order to chaos while longer EnHL is representative of a slower transition (see Fig. 1).

Preliminary results on 40 patients show correlations of EnHL to disease state (H&Y scale) and clinical, questionaire-based quality of life measures (SF-12). However, these findings have to be critically discussed and confirmed on larger populations.

 

This project is part of the EFI Moves initiative.

 

Figure 1: Entropic half life (EnHL) definition as the mid-point in the transition from ordered signals on very short time-scales to random signals on longer timescales. Shorter EnHL encodes a faster transition while longer EnHL is a feature of slower transition from order to chaos.

Partners and Funding Agencies