Person Identification From Text Independent Lip Movement Using the Longest Matching Segment Method
The use of lipreading as a standalone modality for biometric classification continues to gain ground but is still presented with several real world challenges. The paper presents a novel form of video temporal modelling using the Longest Matching Segment (LMS) method on a given baseline training model. LMS uses a Vector Quantization (VQ) model to encode full training video dynamics by mapping it to a frame sequence of maximum likelihood codewords. The model is applied to person identification from text independent lip movement on segmented test sets of the CMU-PIE, VidTIMIT and XM2VTS talking datasets and identification is based on the class with the longest matching segment. The results show that LMS improves the conventional VQ models especially when combined with dynamic delta features. Combined with magnitude 2D-FFT (Mag-2D-FFT) features, the system delivers comparable accuracies to full face recognition.
Keywords: Lip Movement, Vector Quantisation, Longest Matching Segment, Person Identification
Download Full-Text
ABOUT THE AUTHOR
Paul C. Brown
Paul C. Brown graduated from the University of the West Indies with a BSc. Hons Degree in Electrical and Computer Engineering in 1993 and a MSc in Digital Systems in 2001. He graduated from Queens University Belfast in 2017 with a PhD in person and word recognition from lip motion dynamics.
Paul C. Brown
Paul C. Brown graduated from the University of the West Indies with a BSc. Hons Degree in Electrical and Computer Engineering in 1993 and a MSc in Digital Systems in 2001. He graduated from Queens University Belfast in 2017 with a PhD in person and word recognition from lip motion dynamics.