Abstract:
Prevention of forged identities via human identification is crucial for any
authentication system. Electrocardiogram (ECG) is the graphical representation of
the heart surface potential as a function of time. ECG is an emerging Biometric
trait which can overcome the limitation of forgery of traditional Biometrics. ECG
based system can detect the presence of person in living form and thus proves the
aliveness of that individual. For ECG based human identifi cation, apart from
template matching approach, methods involving features categorized as non-fiducial
and fiducial have been reported in the literature. Although, some of these methods
offer high identification accuracy, methods capable of highly accepting heartbeats
authentic to the training database and that of truly rejecting heartbeats external to
the training database thus providing higher authentication accuracy, have been
limitedly reported. Therefore, the development of an ECG based human
identification method capable of effectively identifying true identity remains a
challenging task. Exploiting the fact that approximate coefficients of discrete
wavelet transform (DWT) of ECG signal contain the low frequency components
which provide intrinsic varying details from person to person thus can be
employed as a feature in human identification. Considering the system pole
preserving property of autocorrelation, approximate coefficients of DWT employed
on the autocorrelation sequence of ECG signal is found to be more eff ective than
that obtained from the ECG signal. Unlike the DWT coefficients, the discrete
curvelet transform coefficients have directional p a r a m e t e r s and are more
efficient in representing curve-like edges that differ in ECG signals of diff erent
per- sons. Therefore, cross-correlation of adjacent columns and mean of column
elements of discrete curvelet transform coefficients are utilized to form sets of
features that can offer even a better attractiveness in identifying different humans.
In order to reduce feature dimension, cross-correlation and MC features are
reduced based on dominant energy bands and by employing PCA. The proposed
feature sets when fed to each of the euclidean distance based classifier are shown
to able to identify humans with higher authentication and identification accuracy.
Evaluation of simulation results indicate that the proposed methods provide
superior efficacy compared to that obtained from some of the state-of-the-art
methods of human identification.
