An Efficient ORB based Face Recognition framework for Human-Robot Interaction

Abstract

In Human-Robot Interaction (HRI), quick and efficient FR techniques are often required in service robots. In a real time scenario, it is absolute that face image patterns observed by robots depends often on variations such as pose, light conditions, location of the robots (view point), etc. In addition to these constraints, the service robots are expected to be quick enough for FR so that they can be deployed in applications such as counting people, security and surveillance, directing humans, etc. In this paper, ORB, a computational expensive and quick feature extraction technique is used, which has been a panacea for the above mentioned constraints. One of the dimensionality reduction techniques called PCA (a tool which reduces high dimensional data to lower dimension while keeping most of the data) with its sublime advantages of reduction of storage and time is often used. But, in the FR system, the linear uncorrelated components of PCA doesn't consider the non-linear factors such as occlusion and in such cases PCA fails to find the good representative direction. Kernel PCA (KPCA) which uses kernel methods considers even the non-linear factors and is proven to be more suitable than PCA, thus producing better results. By considering all these factors, our paper proposes a novel technique ORB-KPCA for FR along with Threshold Based Filtering (TBF). The proposed technique is proven to be efficient in both time and space by experimenting on three benchmark datasets (ORL, Faces96 and Grimace).