Gait Recognition Based on Attention Convolutional Networks and Part-based Feature Representation

Gait recognition systems are often affected by factors such as changes in viewing angle, clothing, carried items, and complex backgrounds, leading to significant intra-class variations in gait features and consequently reducing recognition performance. Therefore, how to construct a gait feature representation with strong discriminative power and good robustness has become a key issue in current gait recognition research. To address these issues, this paper proposes a gait recognition method that combines an attention-based convolutional network with block-based feature modeling. First, a gait representation is constructed using gait contour sequences, and global spatial features are extracted using a convolutional neural network. Concurrently, a Convolutional Block Attention Module (CBAM) is introduced to enhance the model’s ability to focus on key human motion regions (such as the legs and joints), thereby improving feature expressiveness. Subsequently, the gait feature map is structurally partitioned using a region-based chunking strategy to extract finer-grained local features. A multi-channel feature fusion strategy is then employed to integrate global and local information, thereby constructing a more robust gait feature representation. Finally, a classification network is used to perform identity recognition on the fused features, achieving high-precision gait recognition. We conducted extensive experiments on the CASIA-B gait dataset and systematically evaluated the model’s performance across different viewpoints and various walking conditions (normal walking, carrying objects, and wearing a jacket). The experimental results demonstrate that the proposed method outperforms multiple baseline models in metrics such as Rank-1 accuracy and mAP, and exhibits superior stability and generalization capabilities across different viewpoints.