Indicating eye contacts in one-to-many video teleconference with one web camera

In video teleconference systems, one drawback that affecting quality of experience (QoE) is the lack of eye contacts. Current methods on how to improve eye contacts in one-to-many video teleconference systems are far from commercial deployment since related hardware are very sophisticated and expensive. In this paper, we propose a novel appearance-based scheme to estimate gaze direction with accessible equipments. Differing from current geometrical fitting methods, machine learning algorithms are used in our scheme, making it more robust against the size and resolution of the eye segments appearing in images. To be specific, an adaptive filter is designed to smooth the fluctuated training data; meanwhile, support vector machine (SVM) is applied to get robust results under various conditions. The experiment results show that, under different illuminations and distances, our method can achieve better performance on gaze direction estimation than current methods in terms of accuracy.