Cross-Domain Generalization for LiDAR-Based 3D Object Detection in Infrastructure and Vehicle Environments

Area of Science:

• Intelligent Transportation Systems
• Computer Vision
• Internet of Things (IoT)

Background:

• 3D object detection is vital for Vehicle-to-Everything (V2X) cooperative perception in intelligent transportation.
• Heterogeneous LiDAR point clouds from diverse sensor configurations pose challenges for 3D object detection model generalization.
• Variations in scale and data heterogeneity degrade model performance.

Purpose of the Study:

• To address the generalization challenges of 3D object detection models with heterogeneous LiDAR point clouds.
• To propose a novel neural network architecture that improves performance across different sensor configurations.
• To enhance feature fusion and robustness in V2X cooperative perception.

Main Methods:

• Introduction of the Dual-Channel Generalization Neural Network (DCGNN).
• Incorporation of a novel data-level downsampling and calibration module.
• Utilization of a cross-perspective Squeeze-and-Excitation attention mechanism for feature fusion.

Main Results:

• DCGNN demonstrates superior performance compared to detectors trained on single datasets.
• Significant improvements observed over selected baseline models on the DAIR-V2X dataset.
• The proposed methods effectively handle variations in point cloud scale and heterogeneity.

Conclusions:

• DCGNN effectively overcomes the generalization problem in 3D object detection for heterogeneous LiDAR data.
• The model enhances the reliability and accuracy of V2X cooperative perception.
• The approach offers a promising solution for robust intelligent transportation systems.