Related Experiment Video
Updated: Aug 7, 2026
07:14
Tracking Infiltration Front Depth Using Time-lapse Multi-offset Gathers Collected with Array Antenna Ground Penetrating Radar
Published on: May 1, 2018
A 4D radar and LiDAR fusion framework for weather-robust 3D object detection
Huaijin Liu1, Jixiang Du2, Hongbo Zhang2
1School of Artificial Intelligence, Guangzhou Maritime University, Guangzhou, 510000, China.
Summary
This study introduces DDMDGF, a new LiDAR-4D Radar fusion framework for robust 3D object detection in autonomous driving. It effectively reduces radar noise and adapts to weather changes, improving detection accuracy.
Area of Science:
- Computer Vision
- Robotics
- Sensor Fusion
Background:
- 3D object detection is crucial for autonomous driving.
- LiDAR struggles in adverse weather; 4D Radar is noisy and low-resolution.
- Existing fusion methods lack adaptability to weather and noise.
Purpose of the Study:
- To develop a weather-robust 3D object detection framework fusing LiDAR and 4D Radar.
- To address sensor degradation and noise in adverse conditions.
- To enhance the reliability of autonomous vehicle perception.
Main Methods:
- Proposed DDMDGF framework with Semantic-guided Foreground-aware Denoising (SFD) module.
- SFD module suppresses radar noise using semantic features and dynamic thresholding.
- Multi-scale Dual-attention Gated Fusion (MDGF) module adaptively fuses sensor data.
Main Results:
- DDMDGF significantly improves 3D object detection performance on K-Radar and VoD datasets.
- Achieved +7.3% AP3D and +4.9% APBEV improvements over L4DR on K-Radar.
- Outperformed L4DR by up to 1.8% mAP in severe fog conditions on VoD-Fog.
Conclusions:
- The DDMDGF framework offers enhanced robustness for 3D object detection in adverse weather.
- The SFD and MDGF modules effectively handle sensor noise and adapt to changing environmental conditions.
- This approach advances the reliability of sensor fusion for autonomous driving systems.
