Transcript Online classifier construction algorithm for human activity

Online classifier construction
algorithm for human activity detection
using a tri-axial accelerometer
Yen-Ping Chen, Jhun-Ying Yang, Shun-Nan Liou ,
Gwo-Yun Lee, Jeen-Shing Wang
Presenter : Yi-Che Liu
From： Applied Mathematics and Computation
Citation : 2
I.F. :
0.961
Introduction





Purpose :
Recognizing different types of human daily
activities using a tri-axial accelerometer.
Online add new training samples.
Online add additional classes.
Online delete an existing class.
Method

Dynamic linear discriminant analysis
Update within-class scatter matrix and betweenclass scatter matrix.
Online classifier construction
algorithm
Data preprocessing

segment the acceleration data into windows with 50%
overlap.
Feature extraction








Mean
Correlation between axes
Energy
Interquartile range
Mean absolute deviation
Root mean square
Variance
Standard deviation
Fuzzy basis function classifier
Experimental design
and results


The acceleration data used in our
experiments was collected using the
MMA7260Q tri-axial accelerometer
The fundamental requirements of the
acceleration device:lightness,
sensing, and wireless transmission



The accelerometer’s sensitivity was
set from -4.0g~+4.0g
The output signal of the
accelerometer is sampled at 100Hz
by a 10-bit ADC
Experiments were performed on
windows XP OS,with P4 2.4GHz CPU
and 512 MB memory

Eight common domestic activities:
standing, sitting, walking, running,
vacuuming, scrubbing, brushing teeth,
working at a computer.

We gathered acceleration data from
seven normal, healthy subjects(4
females, 3males;age 24.1 ± 1.8 years)
in a controlled laboratory environment.



A single tri-axial accelerometer
module was mounted on the
dominant wrist,which is better for
discriminating activities involving
upper body movements.
All were asked to perform each
activity for two minutes.
Sampling frequency is 100Hz.So the
total number of the acceleration
samples for each activity is 12,000.
In the experiments, we applied the
proposed online classifier
construction scheme to the three
cases :
1.Adding additional new samples to
the existing activities
2.Adding new human activities
3.Deleting existing activities
Adding additional new samples to
the existing activities
This graph shows the average recognition rates of consecutively adding additional 10%
new samples to the eight activity classes.
This graph shows the execution times of the dynamic LDA and the conventional LDA.
This graph shows the dynamic LDA only requires a fixed size of memory to store the
statistical information of the currently available training data for the update of the scatter
matrices.
Adding new human activities
This graph shows the average recognition rates of different numbers of activities.
This is reasonable because the recognition task becomes more difficult as the number of
classes increases.
This graph shows the execution times of the dynamic LDA and the conventional LDA.
This is because when a new activity is added, the conventional LDA has to re-compute
“SW” and “SB” using the whole training data.
This graph shows the memory requirements.
Because the conventional LDA has to re-compute “SW” and “SB” too.
Deleting existing activities
This graph shows the execution times of the dynamic LDA and the conventional LDA.
For the dynamic LDA, we used Eqs. (16)–(18) to update the scatter matrices without any
computation from the training data.
For the conventional LDA, we re-computed the scatter matrices by excluding the data of
removed activities from the training data.
This graph shows the memory requirements.
Because the conventional LDA has to re-compute “SW” and “SB” too.
Conclusions



The proposed dynamic LDA is capable of
online updating the scatter matrices with
the same results as those obtained by the
conventional LDA operated in an offline
mode.
The storage of the complete training
dataset is not required for the dynamic
LDA.
The memory usage and computational
efficiency is greatly improved.