A novel biometric recognition method based on multi kernelled bijection octal pattern using gait sound

• Published in: Applied acoustics Vol. 173; p. 107701
• Main Authors: Aydemir, Emrah, Tuncer, Turker, Dogan, Sengul, Unsal, Musa
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2021
• Subjects: Biometrics; Gait recognition; Information fusion; Multi kernelled bijection octal pattern; Sound recognition; Biometrics; Multi kernelled bijection octal pattern; Gait recognition; Information fusion; Sound recognition
• ISSN: 0003-682X; 1872-910X
• DOI: 10.1016/j.apacoust.2020.107701

•A novel sound based biometric recognition method is presented.•A novel Multi Kernelled Bijection Octal Pattern is proposed.•A hybrid feature selector (RFINCA) is used to select dscriminative features.•Two cases were defined. 98.0% and 94.90% classification accuracies were obtained. Many gait based methods have been presented about biometric identification in the literature. Gait recognition methods have generally used images and sensors signals. In this work, a novel gait based biometric recognition method is presented. A novel Multi Kernelled Bijection Octal Pattern (MK-BOP) is presented in this study. The main aim of the proposed MK-BOP is to extract distinctive and comprehensive features from a signal (gait sound). By using the proposed MK-BOP, a novel biometric recognition method is proposed. Gait sounds are collected, and two novel datasets are collected. The first dataset is a noisy and heterogeneous dataset. The second dataset is a clear and homogenous dataset. A multileveled method is presented to authenticate subjects from these datasets. One dimensional discrete wavelet transform (1D-DWT) is applied to sound signal with Symlet 6 (sym6) filter, and levels are calculated. The proposed MK-BOP generates features from each level signals, and the generated features are concatenated. A hybrid feature selector (RFNCA) selects the most discriminative feature, and selected most discriminative features are forwarded to classifiers. 0.980 and 0.949 success rates were achieved for clear and noisy datasets, respectively.