Multiscale Region-Based Convolutional Neural Networks for 3D Object Detection with LiDAR Sensors

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• LiDAR-based 3D object detection is crucial for autonomous driving, especially in low-light conditions.
• Advancements in sparse convolutional networks are needed due to point cloud data sparsity.
• Multiscale feature fusion can enhance detection accuracy by leveraging information across different scales.

Purpose of the Study:

• To improve 3D object detection performance in autonomous vehicles using LiDAR data.
• To address the challenges posed by sparse point cloud data in poor lighting conditions.
• To enhance existing 3D voxel-based detection networks with a novel fusion mechanism.

Main Methods:

• Implemented a refinement fusion network with cross-attention modules integrated into 3D voxel-based detection networks.
• Utilized a refined, realistic strategy for point cloud data augmentation.
• Evaluated the system on the KITTI dataset across three object categories.

Main Results:

• The proposed system demonstrated substantially improved results in 3D object detection.
• Experimental results confirmed the effectiveness of the refinement fusion network and data augmentation techniques.
• The system achieved superior performance compared to existing approaches.

Conclusions:

• The developed LiDAR-based 3D object detection system effectively addresses limitations of current methods.
• The integration of cross-attention modules and refined data augmentation significantly boosts detection performance.
• The proposed system shows strong potential for enhancing the safety and reliability of autonomous driving systems.