Deep Capsule Network Design Method for Gait Recognition

The objective of gait recognition is to identify persons by walking styles. Deep learning methods have presented great potential in gait recognition tasks. However, existing deep networks for gait recognition are mainly constructed by stacking convolutional layers and pooling layers. Since convolutional layers share weights across different local regions, and pooling layers easily cause gait information loss, the performances of existing gait recognition methods based on CNNs are still limited. To address this problem, there have appeared some gait recognition methods by capsules. However, since a capsule represents a single feature activation by a vector rather than a scalar activation, existing deep capsule networks still easily suffer from overfitting risks and the performances are still limited on large gait datasets. In this paper, we study the designing method of deep capsule network for gait recognition. We propose to extract low-level gait dynamic features by temporal module, then design capsule layers based on human body alignment module to extract high-level gait features, then design harmonization module to reduce overfitting risks by combining global and local gait features. The proposed method achieves state-of-the-art recognition results on CASIA-B and OUMVLP dataset, and also achieves 100% accuracies in many experimental settings. This phenomenon demonstrates that capsules are quite useful in improving gait recognition performances, and designing deep capsule networks can provide a feasible direction for achieving high gait recognition accuracies, even achieving 100% accuracies.