Find attachment

75734/Presentation.pptx
Title:
Human Activity Recognition and Monitoring Using Multiple Sensors.
Table of Contents:
Abstract
Introduction
Data Preparation and Pre-processing
Modelling
Generating Test Design
Evaluation and Result
Conclusion
Reference
Abstract:
Automatic detection of human physical activities could have huge impacts not only in health interventions, social networking, lifestyle, but also in targeted advertisement and corporate management.
We implemented and evaluated classification algorithm to detect four crucial human physical activities using five triaxle accelerometers worn concurrently on different parts of the body.
The accelerometer data were collected, cleaned, and pre-processed to extract features from 10 s window.
The time and frequency domain will be give the different features that were used with the random forest and k-nearest neighbour classifiers.
It will classify the subject activities.
Random Forest showed the best performance recognizing the activities with overall accuracy of 89 % for LOSO strategy for hip data.
Introduction:
We implemented and evaluated classification algorithm to detect four crucial human physical activities using five triaxle accelerometers worn concurrently on different parts of the body.
Accelerometers can be used as motion detectors as well as for body-position and posture sensing.
Their low power consumption, small dimensions, and light weight make them great candidates to make long-term activity monitoring, balance assessment, fall detection more practical.
These sensors can be worn on a single or multiple body sites.
We aimed at classifying activities for four classes using a LOSO and ten-fold strategy to evaluate algorithm performance.
Data Preparation and Pre-processing
Data Preparation:
The dataset consists of several files per participant, among these files, we are only interested in six files. 5 files are the different output files from each sensor placed at different locations and 6th file is the annotation interval file.
Pre-processing:
For pre-processing, we extracted the above mentioned files for each participant and created a merged file. This merged file was created on per participant basis. For merging, we took each wocket file with the timestamp and merged it with the annotation file by placing each time stamp into correct time interval.
Modelling:
Classification approach and algorithm:
For classification purpose, k-Nearest Neighbours and Random Forests were employed.
k-Nearest Neighbor is a supervised learning algorithm where the result of new instance query is classified based on 8 majority of K-Nearest Neighbor category.
A good k can be selected by various heuristic techniques, for example, cross-validation.
Random Forests are ensemble learning algorithms for classification that operate by constructing a multitude of decision trees at training time and outputting the class that is the mode of the classes.
We choose Random Forests as they correct the overfitting common in individual decision trees.
Generate Test Design:
Experiments design
The Experiments for this project are designed in several stages.
We started by visualizing the raw data i.e. the acceleration in x, y, z axes.
This visualization helped us to see and analyse the acceleration patterns for different activities.
Using this, we saw that how activity such as walking and climbing stairs have similar acceleration patterns and lying and sitting have similar acceleration patterns. This relation was expected.
Evaluation and Result:
For evaluating our model, we use two validation technique.
The first approach is 10-fold cross validation technique.
In this approach data are randomized and divided into 10 different subsets.
The algorithm is trained on n-1 subsets and tested on the remaining one.
The second approach was leave-one-subject-out cross validation.
Recognition models were trained on data from all subjects except one that is used for the test phase.
Conclusion:
We have identified the most suitable features required for classification of above mentioned activities: - Mean, St. deviation, Median, along with Frequency domain features extracted from signal magnitude vector.
Past experiments have proved that triaxle accelerometer data can be used to classify these type of human activities.
We performed LOSO and 10-fold cross validation techniques to validate our model using Random Forest and kNN, and for both, RF works better than k-NN.
The automatic detection of sensor placement sites could reduce the risk of people using wearable sensors inappropriately and...