Cross-View Gait Recognition Based on ViT and Convolution

The gait recognition draws a powerful magnetizing effect on biometric. It is easily affected by multiple covariant factors, especially clothing occlusion and view changing. To address such impact, this paper proposes a cross-view gait recognition hybrid framework, by integrate convolution and ViT into discriminative method. Taking the gait silhouettes as the original input, a multi-layer convolution and pooling are used to training gait features on different scales. Then, a ViT module is introduced to collect features from different angle, to reduce the influence of covariance of occlusion and view changes. The local details of pixels and global features of gait silhouettes are all concerned. Then, two features are fused to a horizontal pyramid, trained by a joint loss function, to enhance the discrimination and learning ability. Finally, evaluation is performed on the public dataset CASIA-B. Experiments show that the proposed method achieves average accuracy of 95.1%, 90.5%, 72.6% in three states. It is better in cross-view accuracy compared with current methods.