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

Area of Science:

• Human-Computer Interaction
• Radar Signal Processing
• Machine Learning

Background:

• Impulse radio ultra-wideband (IR-UWB) radar offers privacy-preserving and illumination-robust human-computer interaction.
• Single-view radar systems face challenges with occlusion and viewpoint-dependent data loss.
• Existing methods struggle to integrate local motion and long-range temporal data in time-range (TR) representations.

Purpose of the Study:

• To develop a robust dynamic gesture recognition system using multi-view IR-UWB radar.
• To overcome limitations of single-view perception and improve spatio-temporal feature modeling.
• To enhance feature complementarity and discriminability for accurate gesture classification.

Main Methods:

• A neuromorphic-inspired multi-view global-local fusion network is proposed.
• Early fusion of motion-enhanced TR maps from three viewpoints enhances spatial completeness.
• A dual-branch architecture with adaptive fusion (gated first-order and bilinear second-order) captures global and local features.

Main Results:

• The proposed method achieved 98.29% average accuracy on a 12-class UWB gesture dataset.
• Performance was evaluated under a subject-independent protocol.
• The system significantly outperformed existing baseline methods.

Conclusions:

• The developed framework demonstrates high effectiveness for multi-view radar-based dynamic gesture recognition.
• The fusion strategy significantly improves robustness against occlusion and viewpoint variations.
• This work advances the capabilities of IR-UWB radar for sophisticated human-computer interaction.