GaitFusionNet: a dynamic hierarchical feature fusion neural network for multimodal gait recognition

Gait recognition is an important biometric identification technology with broad applications in identity authentication. However, existing multimodal gait recognition methods fail to fully exploit the complementary advantages of different modalities and perform poorly in complex scenarios. We propose a dynamic hierarchical feature-fusion neural network, gait recognition fusion network (GaitFusionNet), which leverages the complementary strengths and distinct vulnerabilities of silhouettes and skeleton maps. We design a quality perception module (QPM) and a feature fusion module (FFM) to dynamically guide fusion, forming a multilevel perceptual adaptive fusion strategy that captures core discriminative information and integrates multimodal, multigranular features effectively. QPM quantitatively assesses silhouette and skeleton feature quality, employing 3D convolution and temporal attention to model gait dynamics and emphasize key frames. FFM preserves the original information through base fusion and further extracts unique gait patterns of different body parts based on the human body structure. The deep synergy between the two modules produces highly discriminative feature representations. Experiments show GaitFusionNet significantly boosts recognition performance and robustness in complex real-world environments, achieving rank 1 rates of 84.4%, 79.8%, and 87.0% on SUSTech1K, Gait3D, and GREW, respectively, surpassing traditional performance limits.