Point Cloud Wall Projection for Realistic Road Data Augmentation.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• Generating synthetic object points from LiDAR data is vital for advanced driver-assistance systems (ADAS).
• Precisely generating points from distant objects using sparse LiDAR data remains a significant challenge.
• Existing methods often require substantial computational power.

Purpose of the Study:

• To propose a novel framework for generating high-precision synthetic LiDAR points, particularly from distant objects.
• To address the computational intensity limitations of current synthetic data generation techniques.
• To improve the accuracy of 3D object detection models in ADAS applications.

Main Methods:

• A framework with three modules: position determination, object generation, and synthetic annotation.
• Utilizes a spherical point-tracing method for augmenting 3D LiDAR distant objects.
• Employs point cloud object projection and point-wall generation, with pose determination for scenarios like platooning.
• Enhances distant point description using multiple LiDAR systems.

Main Results:

• The framework was evaluated on the KITTI dataset using 3D detection models: PointPillars, PV-RCNN, and Voxel R-CNN.
• Demonstrated an increase in mean average precision (mAP): 1.97% for PointPillars, 1.3% for PV-RCNN, and 0.46% for Voxel R-CNN.
• Achieved improved performance in detecting distant objects and handling complex scenarios.

Conclusions:

• The proposed framework effectively generates precise synthetic LiDAR points from distant objects.
• It offers a computationally efficient alternative to existing methods for ADAS applications.
• The framework shows significant potential for enhancing the performance of 3D object detection models.