Most existing methods of identifying person except gait recognition require the cooperation of the subjects. Aiming at detecting the pattern of human walking movement, gait recognition takes advantage of the time-serial data and can identify a person distantly. The time-serial data, which is usually presented in video form, always has a limitation in frame rate, which intrinsically affects the performance of the recognition models. In order to increase the frame rate of gait videos, we propose a new kind of generative adversarial networks (GAN) named Frame-GAN to reduce the gap between adjacent frames. Inspired by the recent advances in metric learning, we also propose a new effective loss function named Margin Ratio Loss (MRL) to boost the recognition model. We evaluate the proposed method on the challenging CASIA-B and OU-ISIR gait databases. Extensive experimental results show that the proposed Frame-GAN and MRL are effective.
Xue, Wei; Ai, Hong; Sun, Tianyu; Song, Chunfeng; Huang, Yan; Wang, Liang
Gait recognition Generative adversarial networks Metric learning Deep learning