Pre-Segmented Down-Sampling Accelerates Graph Neural Network-Based 3D Object Detection in Autonomous Driving.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Graph Neural Networks (GNNs) excel with irregular point cloud data.
• Large-scale LiDAR processing with GNNs faces computational challenges due to neighbor searching.
• Current down-sampling methods in GNNs reduce computation but can harm detection accuracy by not differentiating point categories.

Purpose of the Study:

• To develop an efficient GNN-based 3D object detection method for large-scale LiDAR point clouds.
• To address the computational burden of GNNs in 3D detection.
• To improve computational efficiency without compromising detection accuracy.

Main Methods:

• Proposed a LiDAR point cloud pre-segmented down-sampling (PSD) method to selectively remove background points while retaining foreground object points.
• Developed a lightweight GNN-based 3D detector capable of processing raw, down-sampled LiDAR data directly.
• Evaluated the model on the KITTI 3D Object Detection Benchmark.

Main Results:

• The PSD method significantly improves computational efficiency by reducing background points.
• The proposed lightweight GNN detector achieves effective 3D object detection from down-sampled point clouds.
• The model demonstrates effectiveness and efficiency on the KITTI benchmark, validating its performance for autonomous driving.

Conclusions:

• The proposed PSD method enhances GNN efficiency in LiDAR point cloud processing for 3D detection.
• The lightweight GNN detector offers an efficient solution for autonomous driving applications.
• Selective down-sampling is crucial for balancing computational efficiency and detection accuracy in GNN-based LiDAR processing.