Sudheer Reddy Bandi, A Gopi Suresh , Immanuel Alphonse / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.822-827 822 | P a g e Gait Based Gender Classification Using Silhouette Image Gait Database Sudheer Reddy Bandi 1 , A Gopi Suresh 2 , Immanuel Alphonse 3 1,2,3 Assistant Professor, Dept of Information Technology, Tagore Engineering College, Chennai Abstract A realistic appearance-based representation to discriminate gender from side view gait sequence is introduced here. This gait representation is based on simple features such as moments extracted from orthogonal view video silhouettes of human walking motion. The silhouette image is divided into two regions. The first region includes from the head to the torso region whereas the second region is from torso to the feet region and for each region features are extracted. Finally we employ different pattern classifiers like KNN (K- Nearest Neighbor) and SVM (Support Vector Machine) to classify the gender. The division of two regions is based on the centroid of the silhouette image. The experimental results show that SVM classifier gives better results when compared to other classifiers. The classification results are expected to be more reliable than those reported in previous papers. The proposed system is evaluated using side view videos of CASIA dataset B. Keywords: Appearance based features, binary moments, ellipse features, gait analysis, gender classification, and human silhouette. 1. INTRODUCTION Recently, biometrics has increasingly attracted the attention as a key technology for realizing a more secure and safer society. Although most of the studies on biometrics focus on person authentication, namely hard bio metrics, it is also important to promote the recognition of properties such as gender, age and ethnicity, namely soft biometrics. The perception of gender recognition determines social interactions. Humans are very accurate at identifying gender from a face, voice and through gait. Among biometric modalities, gait has several promising properties such as availability at a distance from a camera even without the cooperation of the subject; hence gait based hard biometrics [13][14][15] has been extensively studied with the aim of realizing wide area surveillance and assistance with criminal investigation. Furthermore, gait based soft biometrics are also an active research area (e.g., gender classification [7][9][10], age group classification [16][17]). There has been much work to classify gender from human faces. In early 1990s, various neural network techniques were employed for gender classification from a frontal face. In paper [1] Golomb et al. trained a fully connected two-layer neural network, SEXNET, to identify gender from face images. Brunelli and Poggio [2] developed HyperBF networks for gender classification in which two competing networks, one for male and the other for female, are trained using 16 geometric features. To sum up, some of these techniques are appearance- based methods and others are based on geometric features. In Moghaddam and Yang’s paper [3], nonlinear SVM was investigated in gender classification for low-resolution thumbnail face (21- by-12 pixels) on 1,755 images from the FERET database. Proceeding toward gender classification in emotional speech; in [4] Harb et al. proposed a method in which a set of acoustic and pitch features are used for gender identification. Concerning previous work on non-emotional speech, the system proposed by Zeng et al. [5] is based on Gaussian mixture models (GMMs) of pitch and spectral perceptual linear predictive coefficients. Nevertheless, when compared to face or voice, gait can be perceived at a distance. In addition to these, gait has various advantages like non-contact, non-invasive and in general, does not require subject’s willingness. This particular issue has stirred up the interest of the computer vision community in creating gait based gender recognition systems. A number of applications can be benefitted from the development of such systems. For example demographic analysis of systems, access control, biometric systems that use gender recognition to reduce the search space to half, etc. Most of gait research has been addressed to biometric identification, which consists of predicting the identity of a person according to his/her way of walk. However, few recent works have used gait analysis for other classification tasks, such as gender recognition or age estimation. Regardless of the purposes two different approaches to describe gait can be considered. Some works extract dynamic features from subject’s movements [12, 10], while others take static attributes related with the appearance of the subject [6, 7, 9]; what implicitly might contain motion information. There have been a number of appearance based methods to classify gender from gait. Lee and Grimson analyzed the motion of 7 different regions of a silhouette [9]. Features for discrimination of gender were selected using Analysis of Variation (ANOVA), and an SVM was trained to categorize gender.