RPT-Mamba: A Range-Aware Physical Token Mamba Network for Far-Field mmWave Radar Gesture Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Radar Signal Processing

Background:

• Millimeter-wave (mmWave) radar offers privacy-preserving contactless sensing.
• Sparse radar point clouds degrade with distance, causing distribution shifts and hindering model generalization.
• Existing methods overlook distance-related degradation in radar point clouds for gesture recognition.

Purpose of the Study:

• To develop a range-aware network for robust mmWave radar point cloud gesture recognition.
• To explicitly model and mitigate distance-induced degradation in radar sensing data.
• To improve the generalization of gesture recognition models from near-range to far-field scenarios.

Main Methods:

• Introduced RPT-Mamba, a range-aware physical token Mamba network.
• Constructed physical point tokens using spatial coordinates, Doppler velocity, echo intensity, and range information.
• Employed a range-aware stochastic degradation strategy and context-guided attribute reconstruction during training.
• Utilized a bidirectional Mamba temporal encoder for modeling long-range gesture dynamics.

Main Results:

• RPT-Mamba achieved 92.09% accuracy and 92.04% Macro-F1 on the mTransSee dataset (random split).
• Demonstrated strong performance under a challenging near-to-far protocol with 85.34% accuracy and 84.77% Macro-F1.
• Outperformed existing point-cloud, radar-gesture, Transformer, and Mamba baselines.

Conclusions:

• RPT-Mamba effectively addresses range-induced degradation in mmWave radar point clouds.
• The proposed range-aware approach significantly enhances generalization for contactless gesture recognition.
• This work provides a robust solution for mmWave radar-based human-computer interaction in varying distances.