Gender Classification from Gait Energy and Posture Images Using Multi-stage Network

Abstract

Gait-based gender classification from an image sequence captured at a distance from human subjects can provide valuation information for video surveillance. One common approach is to adopt machine learning for the prediction of the gender class. Algorithms perform gender classification based on spatio-temporal feature, e.g., Gait Energy Image (GEI), extracted from the video. Although GEI can concisely characterize the movements over a gait cycle, it has some limitations. For instance, GEI lacks photometric information and does not exhibit a clear posture of the subject. To improve gender classification, we think that more features must be utilized. In this paper, we propose a gender classification framework that exploits not only the GEI, but also the characteristic poses of the walking cycle. The proposed framework is a multi-stream and multi-stage network that is capable of gradually learning the gait features from multiple modality inputs acquired in multiple views. The extracted features are fused and input to the classifier which is trained with ensemble learning. We evaluate and compare the performance of our proposed model with a variety of gait-based gender classification methods on two benchmark datasets. Through thorough experimentations, we demonstrate that our proposed model achieves higher gender classification accuracy than the methods that utilize only either GEI, or posture image.