TriFusionNet: Comprehensive Adaptation to Viewpoint Changes and Physiological Information Integration in 3D-Based Person Re-Identification

This study introduces an innovative person re-identification method, TriFusionNet, which cleverly combines 3D reconstruction image features with data augmentation techniques, significantly enhancing performance. We incorporate a “Physiological Information Feature Extraction Branch” to comprehensively capture both the body structure and motion information of persons. The primary challenge addressed is the longstanding negative impact of viewpoint changes on accuracy in person re-identification. Our approach enables the synthesis of more robust and identifiable features in new views. Furthermore, to comprehensively capture person features, we introduce a physiological information feature extraction branch that extracts 15 authentic physiological features from action sequences. These features not only contribute to accuracy but also deepen the understanding of person body status and motion characteristics. The uniqueness of this research lies in going beyond traditional data augmentation methods. Through 3D reconstruction and feature fusion, we address occlusion issues and, through keyframe selection and multi-angle feature fusion, diversify training data more effectively, preserving the body structure and motion information of persons. In summary, TriFusionNet introduces innovative thinking to person re-identification by integrating 3D reconstruction, feature fusion, and physiological information feature extraction. It not only improves performance but also establishes a foundation for more comprehensive extraction of person features, addressing challenges faced by traditional methods.