Modeling and Optimization of Support Vector Machine (SVM) for Detection of Human Eye and 2DPCA as Feature Extraction

Abstract

Eye tracking has become one of the most important processes in the human computer interaction. Eye tracking has also been shown to be useful in various applications. Specific applications included in this system are the system in handwriting reading, music reading, human activity recognition, perceptions of advertising , sport playing, Human Computer Interaction (HCI), especially for people with disabilities, medical research and other areas. Eye detection is the most important thing to do as an initial basis in the application of eye tracking. By getting the position of the eye, eye tracking will be easier to do at a later stage (Bhoi and Mihir 2010) . In general, eye detection is done in two steps. It is determine the location of the face to extract the eye area then eye detection from the eye area. The purpose of this research is perform the modeling and optimization of kernel parameters of SVM with GA for human eye detection with 2DPCA as feature extraction. The research methodology consists of several stage. These are data collection, data preprocessing, feature extraction, modeling the human eye detection with SVM and GA, manufacture training and testing system for eye detection, analysis, and research report writing. Experiment conducted with feature extraction using the 2DPCA. The results obtained then validated using the Euclidean distance. The experiment results gained 68.97% accuracy on the dimension (d) = 2. The next experiment was performed using the SVM algorithm for the classification algorithm. Kernel used in SVM is linear, RBF, and polynomial. The results of the three kernel maximum accuracy is obtained with the RBF kernel parameters C = 3 and log2 Gamma = -1. This research is made using Matlab R2008b version 7.7.0.471. Eye detection model feature extraction using SVM with 2DPCA produce the highest accuracy on the RBF kernel with value Log2 Gamma = -1 and C > 0 with a value of 99.97 % on the training data and 88.16 % on the testing data.