1 INTRODUCTION
HAR (Human Activity Recognition) can be defined as analyzing and identifying systematic of
the information such as the types of observation and the patterns of behavior, and the recogni-
tion results can be described in a natural language(Aggarwal. 2011). HAR system can perceive
the intention of users, so it has broad application prospects in intelligent video surveillance,
medical diagnosis, motion analysis and human-computer interface(HCI), which has become a
research hot spot in the field of artificial intelligence and pattern recognition(Ni et al. 2013,
Yang et al. 2012, Bulling et al. 2008). At present, EOG-based HAR has become a new research
spot. In the EOG-HAR system, identifying saccade signals is a critical step, some algorithms of
detecting eye movement have been proposed. Among them, Clement proposed a method which
use the visual angle of original EOG signals to identify and recognize eye movement signal
endpoint(Clement. 1991). Aungsakun utilized the characteristics that the eye movement of EOG
signals changed faster to extract eye movement characteristic parameters(Aungsakun et al.
2011). Besides, Antrobus also suggested to use the statistics of eye movement signals and the
characteristics of time domain(Antrobus. 1973). The approaches above mainly focus on the
time-domain characteristics of EOG signals. Obviously, it is difficult to depict the original EOG
signals correctly in some noise environment (such as the movement of the electrode position,
channel distortion, etc.),which cannot be avoided when applied in the real environment, hence
feature parameters of saccade signal based on time-domain analysis revealed a poor robustness.
On the other hand, it is common to use multi-channel to acquire EOG data in order to obtain
A novel saccade signals detection algorithm for EOG-based
human activity recognition
Zhao Lv
The Key Laboratory of Intelligent Computing & Signal Processing, Anhui University, Hefei, China
Co-Innovation Center for Information Supply & Assurance Technology, Anhui University, Hefei, China
Jinning Guan
School of Computer Science and Technology, Anhui University, Hefei, China
Bengyan Zhou
Co-Innovation Center for Information Supply & Assurance Technology, Anhui University, Hefei, China
Xiaopei Wu(*corresponding author)
The Key Laboratory of Intelligent Computing & Signal Processing, Anhui University, Hefei, China
ABSTRACT: A research on the relationship between eye movements and human behavior is a
hot topic in the field of Human Activity Recognition (HAR). In this paper, a novel saccade sig-
nals detection algorithm for EOG-based HAR, which aims to improve the performance of HAR
system, was proposed. In the proposed algorithm, Common Spatial Pattern (CSP) was utilized
to build spatial filters, and then use it to process original multi-channel EOG signals. Conse-
quently, feature parameters of different saccade signals can be acquired. To valid the perform-
ance of the proposed algorithm, a linear Support Vector Machine (SVM) was chosen. In lab en-
vironment, four types of saccade signals corresponding to up, down, left and right were used as
analysis objects. Experimental results show that the accuracy recognition ratio is about 97.7%,
which reveal that the proposed algorithm has a good classification performance in saccade sig-
nals analysis.
KEYWORDS: EOG; saccade signals; common spatial pattern(CSP); joint approximate diago-
nalization; support vector machine (SVM)