AdaptiveGait: adaptive feature fusion network for gait recognition

Gait recognition is a biometric approach used to identify people based on their walking patterns at long distances and low resolutions. Most advanced gait recognition methods based on silhouettes employ the focal convolution module. However, experiments have demonstrated that the horizontal segmentation method used in this module causes information loss at the feature map demarcation line. In this paper, we propose an adaptive feature fusion block (AFFB) for feature extraction that utilizes comprehensive global features to compensate for the lost local features, significantly reducing feature loss caused by local convolution. Additionally, we introduce a feature expansion module (FEM) to enrich the temporal information of gait features and adaptively balance the body detail information extracted by the model with the overall body information . We evaluated our model on CASIA-B and OU-MVLP datasets and compared it to other gait models using RANK-1 accuracy. The experimental results show that our model can represent gait features better than other models and achieved high accuracy in gait recognition across perspectives and various walking conditions.The source code will be available on https://github.com/Lentia/AdaptiveGait.