TSGait: Simple and Effective Two-Stream Network for Gait Recognition

Current gait recognition work ignores the importance of motion information. In this paper, we propose a novel two-stream gait recognition framework called TSGait, which fuses the motion and appearance information to obtain more discriminative feature representations. Specifically, we build a feature extraction module called Dynamic and Static Feature Extractor (DSFE), which can obtain more comprehensive gait information. Moreover, we propose a new distance metric method called Weighted Euclidean distance (WEuc), which aims to adaptively adjust the scale of the Euclidean distance by imposing the direction constraint on it. The WEuc distance is applied to triplet loss and the evaluation system, which can significantly improve the recognition performance of the model. The experiment results indicate that the average accuracy of our method is $\mathbf{9 2. 7 \%}$ on CASIA-B dataset, which outperforms state-of-the-art methods.