GAIT based Automatic Human Gender Prediction

Abstract

As automatic gender prediction of a person is important in many social activities, a study and analysis of human gait for automatically predicting the gender of a person is presented in this paper.  The main objective of this work is to develop a system to identify the gender of individuals derived from a video sequence of their walking (GAIT). Unlike other human traits such as finger print, face, palm etc., acquiring the GAIT of a person does not require his cooperation and hence GAIT is used in this study to predict gender.  People belonging to same age but different genders have different gait appearance and in this work, appearance-based gait feature namely gait energy image (GEI) is chosen for predicting the gender. The GEI image is partitioned into five regions such as head & neck, chest & back, waist &buttocks, leg and foot from which features are extracted. Different weights are assigned to different body components according to their gender discriminating power. Since gender classification is a two-class problem, equal number of subjects belonging to male and female class is taken from the CASIA Gait Database (Dataset B) to avoid bias. The gait data for the 62 subjects consists of 31 male and 31 female subjects.  Out of 31 male subjects 21 samples is used for training and 10 is used for testing. Similarly, it is done for training and testing the female subjects too. Three -fold cross-validation is employed to evaluate the performance of the SVM classifier and the overall accuracy, specificity and sensitivity of the proposed gender detection system are 91.537%, 91.38% and 91.11% respectively.