Physics-aware generative fusion via Complex-Flow Mamba enables robust 4D radar-vision perception.

Area of Science:

• Autonomous Perception
• Signal Processing
• Deep Learning

Background:

• Conventional deep learning models for 4D imaging radar suffer from the 'point cloud fallacy,' treating echoes as discrete points and losing crucial phase coherence.
• This loss of phase information makes systems vulnerable to interrupted sampling repeater jamming (ISRJ) and degrades performance on sparse targets at long ranges.

Purpose of the Study:

• To propose Complex-Flow Mamba, a novel framework that operates at the signal-physics layer to preserve complex-domain phase coherence in 4D imaging radar.
• To develop methods for mitigating ISRJ interference and improving detection of sparse targets.

Main Methods:

• Constructed a continuous Implicit 4D Complex Field using polar decoupling and Von Mises distribution modeling to recover phase gradients.
• Introduced a Complex Latent Diffusion Module (CLDM) for microscopic phase correction to address ISRJ interference.
• Utilized Mamba-2's Structured State Space Duality (SSD) for an O(N) linear complexity bi-directional state-modulation backbone, entangling visual and radar data streams.
• Incorporated a generative detection head with Consistency Models and a physics-consistent self-supervised loss.

Main Results:

• Achieved state-of-the-art (SOTA) performance on the View-of-Delft (VoD) and K-Radar datasets, demonstrated by high mean Average Precision (mAP) and NuScenes Detection Score (NDS).
• Demonstrated a jamming suppression gain of 0.88 under extreme electronic warfare conditions.
• Successfully addressed the 'hollowing' effect in far-field detection.

Conclusions:

• Complex-Flow Mamba effectively preserves radar phase coherence, significantly enhancing autonomous perception capabilities.
• The framework demonstrates robust performance against ISRJ and improves detection of challenging targets, setting a new SOTA.
• Deep entanglement of heterogeneous data flows and advanced signal processing techniques are key to overcoming current limitations in radar perception.