FG-Pillar: enhancing LiDAR-based small object detection via fine-grained pillar features.

Area of Science:

• Computer Vision
• Robotics
• LiDAR Technology

Background:

• LiDAR sensors provide 3D coordinates but generate sparse point clouds.
• Sparse data leads to information loss and difficulty in extracting fine-grained object details.
• Existing 3D object detection methods struggle with low-density point clouds.

Purpose of the Study:

• To propose a novel pillar-based network for enhanced 3D object detection.
• To improve the network's ability to handle sparse LiDAR data and detect small objects.
• To enhance feature extraction and fusion for better accuracy in challenging conditions.

Main Methods:

• Introduced a dual-stream adaptive feature enhancement module with residual-gated refinement and channel grouping.
• Designed a context-sensitive feature fusion module incorporating context aggregation and semantic refinement branches.
• Utilized a pillar-based network architecture for processing 3D point cloud data.

Main Results:

• The proposed network outperformed the baseline PillarNet on the KITTI dataset.
• Achieved significant improvements in Average Precision (AP) for pedestrians and cyclists in 3D and Bird's-Eye View (BEV).
• Demonstrated enhanced accuracy in detecting small objects and improved multi-scale feature integration.

Conclusions:

• The novel network effectively enhances 3D object detection accuracy from sparse LiDAR point clouds.
• The proposed feature enhancement and fusion modules are crucial for improving sensitivity to fine-grained and global information.
• The method shows promise for broader applications of LiDAR in small-object detection tasks.