Application of maximized inter-class variance for gender classification using RGB-depth camera

This study presents an approach for gender recognition problem in 3D space. Studies have shown that anatomic cues supply distinctive information for visual gender recognition. Vertical, horizontal and depth coordinates of 20 different joints of a walking pedestrian are extracted. Intra-class based means of the whole of 60 features are calculated. In next step, absolute differences between the mean values of male and female class features are found. The ones with differences higher than a predefined percentage are considered as distinctive features and used for class representation. We conducted our experiments and benchmark with Genetic Algorithm on a publicly available dataset UPCV gait captured with Microsoft Kinect, consisting of 5 gait sequences from 30 people. We used Multilayer Perceptron for training and testing. According to the results, our method shows a higher performance in less computation time.