Deep segmentation of 3+1D radar point cloud for real-time roadside traffic user detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Smart Infrastructure

Background:

• Smart cities require robust perception systems for road safety and traffic management.
• Automotive radar offers reliable performance in adverse conditions but faces challenges in object detection resolution.
• Traditional clustering methods struggle with vulnerable road users and separating nearby objects.

Purpose of the Study:

• To develop a deep learning-based 3D radar point cloud clustering methodology for smart infrastructure.
• To enhance the accuracy and efficiency of traffic participant detection using radar data.
• To enable real-time perception for intelligent transportation systems.

Main Methods:

• A deep learning approach combining semantic and instance segmentation of 3D radar point clouds.
• Utilizing a deep neural network for point cloud processing and object clustering.
• Developing a system tailored for smart infrastructure-based sensor setups.

Main Results:

• Achieved 95.35% F1-macro score for semantic segmentation.
• Obtained 91.03% mean average precision (mAP) at 0.5 IoU for instance segmentation.
• Real-time pipeline operates at 43.61 FPS with low memory footprint (<0.7 MB) on edge devices.

Conclusions:

• The proposed deep learning method significantly improves 3D radar point cloud clustering for smart city applications.
• The system demonstrates high accuracy in semantic and instance segmentation, outperforming traditional methods.
• The real-time performance and low resource requirements make it suitable for edge deployment in intelligent infrastructure.