4D unconstrained real-time face recognition using a commodity depth camera

Abstract

Robust unconstrained real-time face recognition still remains a challenge today. The recent addition to the market of lightweight commodity depth sensors brings new possibilities for human-machine interaction and therefore face recognition. This article accompanies the reader through a succinct survey of the current literature on face recognition in general and 3D face recognition using depth sensors in particular. Consequent to the assessment of experiments performed using implementations of the most established algorithms, it can be concluded that the majority are biased towards qualitative performance and are lacking in speed. A novel method which uses noisy data from such a commodity sensor to build dynamic internal representations of faces is proposed. Distances to a surface normal to the face are measured in real-time and used as input to a specific type of recurrent neural network, namely long short-term memory. This enables the prediction of facial structure in linear time and also increases robustness towards partial occlusions.