CANet: Cross Alignment Network for Human Gait Recognition

Gait recognition has emerged as a crucial biometric technique because it enables unobtrusive and remote identification. However, silhouette-based methods are often affected by variations in clothing and carrying conditions. To address these challenges, we propose a novel multi-modal approach that integrates Gait Energy Images (GEI) and Skeletal Gait Energy Images (SkeGEI) to enhance recognition performance. GEIs capture appearance-based holistic motion patterns, while SkeGEIs, generated from normalized skeletal representations centred on the neck joint, provide robust structural features. Skeletal coordinates are extracted using HRNet for the CASIA-B dataset and AlphaPose for the OU-MVLP dataset, which are then used to generate SkeGEIs. Our composite fusion model effectively combines GEI and SkeGEI features through Cross-Alignment blocks (CAB), attention mechanisms, and Progressive Fusion Blocks (PFB). Extensive evaluations on the CASIA-B and OU-MVLP datasets demonstrate that the proposed approach achieves superior performance compared to state-ofthe- art methods across diverse conditions. However, the model struggles with crossdataset generalization, particularly in varied clothing conditions. Additionally, an ablation study highlights the effectiveness of each component, validating the superiority of our integrated model.