Dependable Person Recognition by Means of Local Descriptors of Dynamic Facial Features

In this work, a complementary approach that adds a dynamic component to face biometrics is proposed. The dynamic appearance and the time-dependent local features characterizing the face of an individual during speech utterance are indeed considered in their spatial and temporal components. Ultimately, the aim is to capture, represent and compare facial patterns related to speech utterance, to improve biometric system dependability thanks to an intrinsically difficult to forge descriptor. The proposed approach applies the concept of dynamic texture to the domain of person identification through dynamic facial patterns modeled by means of the Volume Local Binary Pattern (VLBP) descriptor, which effectively combines local features and movement. To the aim of improving the efficiency of this technique, only the occurrences of the Local Binary Patterns related to Three Orthogonal Planes (LBP-TOP) have been considered. A deep feed forward network has been trained and optimized on video samples from the XM2VTS database concerning utterance of a given sentence. The results obtained in the recognition task performed on test video sequences confirm that the proposed approach features state-of-the-art performances with regard to accuracy and robustness of the identification.