New approaches based on local binary patterns for gender identification from sensor signals

Abstract

Gender identification (GI) is to determine the sex of the individual based on the characteristics that distinguish between male and female. In this study, three different feature extraction methods are proposed for gender identification by using signals obtained from accelerometers, magnetometers and gyroscope sensors installed in 5 different body parts of the individuals. Feature extraction from signals is one of the most critical stages of GI. Because the success of GI depends on the features, different transformation methods have been applied to the signals obtained from sensors such as One Dimensional Local Binary Patterns (1D-LBPs), One Dimensional Robust Local Binary Patterns (1D-RLBPs) and Weighted One Dimensional Robust Local Binary Patterns (W-1D- RLBPs). By using these features, different machine learning methods (SVM, RF, ANN, Knn) were elaborated for classification. According to the results, it was seen that 1D-LBPs, 1D-RLBPs and W-1D-RLBPs methods provided useful features for GI. 96.04%, 96.72% and 97.28% success rates were perceived with the proposed methods, respectively. In the study, features, motions, sensor port, sensor type that affect GI success were determined. The achievements of the proposed methods have been found to be more successful than the feature groups provided by the frequency and time domains of the proposed feature extraction methods, which are also compared with the success of the feature groups derived from the same sensor signals from the time and frequency domains.