IMPROVED TWIN SUPPORT VECTOR MACHINE WITH GENERALIZED PINBALL AND APPLICATION ON HUMAN ACTIVITY RECOGNITION

Abstract

In this paper, we propose a new classifier termed as an improved version of twin support vector machine with generalized pinball (GPin-ITSVM). The primary advantage of GPin-ITSVM over GPin-TSVM is that avoids the singularity problem when solving the dual problems of GPin-TSVM. Motivated by the need to address this issue, we modify the GPin-TSVM by adding an extra regularization term to the objective function in primal problems of GPin-TSVM. Numerical experiments are carried out on 12 UCI benchmark datasets to investigate the validity of our proposed algorithm. The results show that the our GPin-ITSVM is superior to existing classifiers in classification accuracy. In addition, the use of this approach in Weizmann activity recognition applications is investigated, and the automatic feature extractor makes use of 5 types of Convolution Neural Network (CNN) models which are ResNet50, ResNet152V2, InceptionV3, InceptionResNetV2 and Xception.