A Gait Recognition Network With Attention-GaitNet

In this paper, we propose a novel network structure GaitNet, which treats gait as a set of disordered sequences, compares different sequences to extract features, and introduces attention to gait recognition, thereby achieving a good recognition score under different angles and carrying situations of different clothes/bags. Experiments show that our method shows its advantages when the cross-view angle is larger than thirty-six angle. The average rank-1 accuracy rate on the CASIA-B gait dataset is greater than ninety-five angle, and it is in the OU-MVLP gait. The accuracy of the average rank-1 on the data set is greater than eighty-six angle, and our method shows excellent robustness in the case of missing frames and incomplete silhouettes of some images. At the same time, a gait database based on the OU-MVLP gait data structure was established. The video includes long-distance and short-distance from zero angle to two hundred and seventy angle, recording a video every fifteen angle, with two sequences, the average recognition rate of rank-1 is over seventy-eight angle.