GaitPart‑based Cross‑view Gait Recognition Method

Abstract

[Purposes] To address the problem of sharp decrease in recognition accuracy caused by covariate factors such as camera view or pedestrian occlusion in gait recognition, an improved feature enhancement GaitPart cross-view gait recognition method IFE-GaitPart (An Improved Feature Enhancement GaitPart) is proposed. [Methods] In the proposed method, the network model was improved into a two-path parallel form containing a spatial feature extraction branch and a significant temporal modeling branch. First, a convolutional network was used to extract shallow features from the original input sequence as the input of the two-path network. Then, a non-uniform convolutional approach was proposed to extract fine-grained information of gait in spatial feature extraction and global features were fused to improve the information capacity of spatial features, while in the significant temporal modeling branch, an adaptive multi-scale temporal feature extraction module was proposed on the significant time modeling branch to obtain the short-term dependence and global time cues of the gait in the time dimension, and the complete temporal information was obtained after stitching. Finally, the temporal features were stitched in the channel dimension and the gait features were output by using a fully connected layer. [Results] The experimental results show that the effectiveness of the method is demonstrated by achieving 97.6%, 94.5%, and 81.1% Rank-1 accuracy on the CASIA-B dataset with normal walking, carrying luggage, and wearing outerwear, respectively.