Gait recognition based on multi-feature representation and temporal modeling of periodic parts

• DOI: 10.1007/s40747-023-01293-z
• Publication Date: 2023-12-11T07:02:16Z
• Authors: Zhenni Li; Shiqiang Li; Dong Xiao; Zhengmin Gu; Yue Yu
• Keywords: Feature (linguistics); Convolution (computer science); Representation

Abstract Despite the ability of 3D convolutional methods to extract spatio-temporal information simultaneously, they also increase parameter redundancy and computational storage costs. Previous work that has utilized 2D convolution method approached problem in one two ways: either using entire body sequence as input global features or dividing into several parts local features. However, tends overlook detailed specific each part, while fails capture relationships between regions. Therefore, this study proposes a new framework for constructing representations, which involves extracting fusing novel manner. To achieve this, we introduce multi-feature extraction-fusion (MFEF) module, includes branches: branch extracts individually, after are fused multiple strategies. Additionally, gait is periodic action different contribute unequally recognition during cycle, propose temporal feature modeling (PTFM) from adjacent frame complete based on Furthermore, fine-grained our utilizes parallel PTFMs correspond with part. We conducted comprehensive experimental widely used public dataset CASIA-B. Results indicate proposed approach achieved an average rank-1 accuracy 97.2% normal walking conditions, 92.3% carrying bag walking, 80.5% wearing jacket walking.