RaSS: 4D mm-Wave Radar Point Cloud Semantic Segmentation with Cross-Modal Knowledge Distillation.

Area of Science:

• Robotics and Artificial Intelligence
• Sensor Fusion for Autonomous Systems
• Environmental Perception Technologies

Background:

• Autonomous driving relies on sensors like LiDAR and cameras for environmental perception.
• 4D mm-Wave radar offers robust performance in adverse weather, capturing 3D point clouds and Doppler velocities.
• Radar's inherent data sparsity and noise limit its application in point-level tasks like semantic segmentation.

Purpose of the Study:

• To investigate the feasibility of utilizing 4D mm-Wave radar for semantic segmentation tasks.
• To develop a novel framework for radar-based semantic segmentation.
• To address the challenges posed by sparse and noisy radar data.

Main Methods:

• Introduction of the ZJUSSet dataset, providing point-wise class labels for radar and LiDAR data.
• Proposal of RaSS, a cross-modal distillation framework designed for radar semantic segmentation.
• Development of an adaptive Doppler compensation module to enhance segmentation accuracy.

Main Results:

• The RaSS model demonstrated significant performance improvements over existing baselines and competitors on the ZJUSSet and VoD datasets.
• The framework effectively handles sparse and noisy 4D radar data for semantic segmentation.
• Validation of 4D radar's potential for detailed environmental perception tasks.

Conclusions:

• 4D mm-Wave radar is a viable sensor for semantic segmentation in autonomous driving, overcoming previous limitations.
• The RaSS framework and adaptive Doppler compensation module represent a significant advancement in radar perception.
• Future work will involve releasing the code and dataset to facilitate further research in this area.