Skip to content
forked from tpimentelms/WHARF

Wearable Human Activity Recognition Folder. Machine Learning framework for training models and running inference of human bimanual gestures with data from smart watches

Notifications You must be signed in to change notification settings

tricostume/WHARF

 
 

Repository files navigation

WHARF - Wearable Human Activity Recognition Folder

alt text

  1. What is it?

WHARF - Wearable Human Activity Recognition Folder is a public repository of code and data sets for Human Activity Recognition systems based on information provided by wearable sensors.

WHARF is composed of:

  • WHARF Data Set, a collection of labelled accelerometer data recordings (obtained by a single wrist-worn tri-axial accelerometer) to be used for the creation and validation of acceleration models of simple human activities.

  • WHARF Code, a collection of MATLAB functions allowing for the creation and validation of the models and trials in WHARF Data Set. Detailed descriptions of the proposed system and its performance can be found at:

    1. Bruno, B., Mastrogiovanni, F., Sgorbissa, A., Vernazza, T., Zaccaria, R.: Analysis of human behavior recognition algorithms based on acceleration data In: IEEE Int Conf on Robotics and Automation (ICRA), pp. 1602--1607 (2013)

    2. Bruno, B., Mastrogiovanni, F., Sgorbissa, A., Vernazza, T., Zaccaria, R.: Human motion modelling and recognition: A computational approach In: IEEE Int Conf on Automation Science and Engineering (CASE), pp. 156--161 (2012)

Abstract

One of the most challenging fields of research is concerned with the task of understanding the degree of independence of a person by analysing the characteristics of several daily life tasks carried out by the subject. The article expands a framework dedicated to the recognition of single-hand gestures by means of inertial measurements with the additional capability of recognising bimanual gestures. Modelling is performed through Gaussian Mixture Modelling (GMM) with K-means optimised selection of Gaussians and Gaussian Mixture Regression (GMR) and the validation of tests; through Mahalanobis distance comparison and punctual Gaussian a priori probabilities. Two kinds of models are obtained for 5 activities, namely the 4 × 4D and the 2 × 7D models with implicit and explicit correlation among devices respectively and validations carried out with each approach for comparison. Validation can be performed twice, once with the previously mentioned approach and once with the Dynamic Time Warping Method (DTW). With the whole picture of this work, three out of five activities could be recognised in the majority of trials.

Code

The code is separated in two folders: WHARF Code and WHARF Data Set. The first one contains all functions necessary for training and inference and their appropriate functioning depend on the existence of the data set in the second folder. As an example from the activity Open/Close curtains, the model extracted from the activity "Open/Close Curtains (OCC)" can be seen here: alt text

Installation

This code is written in Matlab and it remains the only requisite to be able to run properly.

Results

Five actions could be classified with an overal accuracy of 82% as concluded by a K-fold cross validation analysis as seen in this confusion matrix and further explained in our paper. alt text

About

Wearable Human Activity Recognition Folder. Machine Learning framework for training models and running inference of human bimanual gestures with data from smart watches

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • MATLAB 100.0%