Gait Energy Image-Based Human Attribute Recognition Using Two-Branch Deep Convolutional Neural Network

Gait is an attractive biometric identifier, playing an essential role in addressing the issue of identity and attribute recognition in surveillance for its non-invasive and non-cooperative features. In this study, we propose a two-branch deep convolutional neural network for gait-based attribute recognition, including age estimation and gender recognition. We improve the estimation module by predicting a joint distribution instead of two independent distributions. In addition, several improvements are also proposed for improving the final performance of human attribute recognition, including data augmentation methods and loss functions. We implement several gait-based attribute recognition experiments on the OULP-Age and OU-MVLP datasets. Experimental results show that the proposed method outperforms existing approaches. Finally, we elicit different body regions' contributions on attribute recognition tasks. Our conclusions can help improve the robustness of gait-based human attribute recognition systems in future.