Up-Sampling Method for Low-Resolution LiDAR Point Cloud to Enhance 3D Object Detection in an Autonomous Driving...

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• High-resolution LiDAR is costly, limiting 3D object detection research budgets.
• Low-resolution LiDAR (32-Ch or less) results in reduced detection accuracy due to sparse point clouds.

Purpose of the Study:

• To propose a simple and effective method for up-sampling low-resolution point cloud data.
• To enhance 3D object detection performance by reconstructing sparse data into denser point clouds.

Main Methods:

• Converted 3D point cloud data into a 4-channel 2D range image (x, y, z, intensity).
• Employed interpolation based on pixel distance and range values of six neighbors to reconstruct empty spaces.
• Addressed over-smoothing issues common in conventional interpolation methods.

Main Results:

• Improved operational speed and object detection performance compared to deep-learning super-resolution methods.
• Applied to a 32-Ch point cloud input for a point-pillar detection model.
• Achieved significant mAP increases on the KITTI dataset: 9.2%p for pedestrians, 6.3%p for cyclists, and 5.9%p for cars.

Conclusions:

• The proposed up-sampling method effectively enhances 3D object detection using low-resolution LiDAR data.
• The technique offers a cost-effective solution for improving autonomous driving perception systems.
• Future work will explore applications in diverse dataset environments beyond autonomous driving.