Dynamic Feature Fusion for Sparse Radar Detection: Motion-Centric BEV Learning with Adaptive Task Balancing.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Sensor Fusion

Background:

• 4D millimeter-wave radar is crucial for autonomous driving perception.
• Existing methods face challenges with sparse point clouds and dynamic object characterization.
• Robust motion modeling is essential for reliable detection in complex environments.

Purpose of the Study:

• To propose a novel motion-aware framework for enhanced 4D millimeter-wave radar detection.
• To address limitations in existing methods regarding sparse data and dynamic object representation.
• To improve the accuracy and efficiency of object detection for autonomous vehicles.

Main Methods:

• A Bird's Eye View (BEV) fusion network with velocity vector decomposition and dynamic gating.
• A gradient-aware multi-task balancing scheme to resolve optimization conflicts.
• A two-phase progressive training strategy involving multi-frame pre-training and single-frame refinement.

Main Results:

• Achieved 33.25% mean Average Precision (mAP)_3D on the TJ4D benchmark.
• Demonstrated superiority in pedestrian detection (+4.16% AP).
• Maintained real-time performance at 24.4 FPS on embedded platforms with minimal parameter overhead (1.73 M).

Conclusions:

• The proposed motion-aware framework significantly advances 4D radar perception for autonomous driving.
• The method effectively encodes motion patterns and balances multi-task learning.
• This work provides a robust and efficient solution for perception under challenging conditions.