Neuromorphic-inspired multi-view global-local fusion for IR-UWB radar dynamic gesture recognition

Dynamic gesture recognition using impulse radio ultra-wideband (IR-UWB) radar has attracted increasing interest for privacy-preserving and illumination-robust human-computer interaction. However, single-view radar perception is susceptible to occlusion and viewpoint-dependent information loss, while existing methods often struggle to jointly model fine-grained local motion patterns and long-range temporal dependencies in time-range (TR) representations.

To address these issues, this paper proposes a neuromorphic-inspired multi-view global-local fusion network for IR-UWB radar dynamic gesture recognition. Specifically, motion-enhanced TR maps from three complementary viewpoints are first integrated via early fusion to improve the spatial completeness of radar observations. A dual-branch architecture is then employed to capture local dynamic textures and global temporal structures in parallel. In addition, an adaptive fusion module combining gated first-order fusion and bilinear second-order interaction is introduced to enhance feature complementarity and representation discriminability.

Experiments on a public 12-class UWB gesture dataset under a subject-independent protocol show that the proposed method achieves an average accuracy of 98.29%, outperforming several representative baselines.

These results demonstrate the effectiveness of the proposed framework for robust multi-view radar-based dynamic gesture recognition.