Comprehensible people recognition using the Kinect's face and skeleton model

In recent years, significant progress has been made in facial recognition. New systems achieve remarkable recognition rates even in unconstrained environments and overcome humanlevel performance. However, these systems need millions of images to learn facial features. As a consequence, the learning process is computationally expensive. Although the resulting face representation consists of very few features only, it can hardly be understood as it is the result of a complex learning process. In an attempt to overcome these issues, we limit our approach to comprehensible features with clear interpretation. In this paper, we propose a set of comprehensible and stable geometric features based on the Kinect's face and skeleton model, that allow for the use of simple classifiers such as k-NN and LDA. These features include the estimated distances between facial feature points in space that are obtained from the Kinect's face model and geometric distances between joints of the Kinect's skeleton model. In order to obtain more stable features, we propose to aggregate sequences of these measures into a more stable and still comprehensible representation. We test the quality of the proposed features with two classification methods, k-Nearest Neighbour (k-NN) and Linear Discrimination Analysis (LDA) and report a recognition rate of up to 88% for k-NN and up to 89% for LDA on a data set with 37 individuals. The results show that the proposed features can be used to contribute to reliable and comprehensible people recognition.