Gait recognition based on multimodal fusion of silhouette and pose heatmap

As a non-contact biometric recognition technology, gait recognition demonstrates significant application value in intelligent security monitoring, medical rehabilitation assessment and other fields, thanks to its characteristics of long-distance perception and the difficulty of disguising behavioral patterns. However, existing technologies suffer from significantly limited recognition accuracy under complex environments (such as abrupt lighting changes and dynamic occlusions), pose variations (such as turning and speed changes), and cross-view imaging conditions. This paper proposes a multimodal gait recognition method fusing human silhouette and pose heatmap, which realizes adaptive adjustment of modal fusion weights through the collaborative modeling of deep convolutional neural network (CNN) and Transformer architecture, combined with a dynamic scene gating network (DSGN). Experimental results show that on the public datasets CASIA-B and GREW, this method improves the recognition accuracy by 4.5%-10.1% compared with traditional single-modal technologies, verifying the effectiveness of multimodal feature complementarity and the dynamic gating mechanism.