Tree structure convolutional neural networks for gait-based gender and age classification

Gender classification and age estimation are tasks in which humans excel. If gender and age of human can be recognized automatically from images, it will be very helpful in many applications such as intelligent surveillance, micromarketing, etc. We propose a framework for gender and age classification through gait analysis. Gait-based recognition is a feasible approach as the gait of human subject can still be perceived at a long distance. The spatio-temporal gait features are concisely represented as Gait Energy Image (GEI), which is then input to a tree structure convolutional neural network (CNN). We train and test the first model on a single-view gait dataset. Based on the tree structure CNN framework, we propose a larger model for gender and age classification with the multi-view gait dataset. Our models can achieve gender classification accuracy of 97.42% and 99.11% on single-view gait and multi-view gait respectively. We then use our model to perform age group estimation and binary (young and elder groups) classification. Also, our models can achieve the best performance in specific age estimation in terms of various numerical measures than various recently proposed methods.