LiDAR-camera-system-based unsupervised and weakly supervised 3D object detection.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• LiDAR camera systems are crucial for autonomous driving 3D object detection.
• Limited annotated data hinders the full utilization of acquired sensor data.
• Existing methods struggle with large volumes of unannotated data.

Purpose of the Study:

• To propose a novel LiDAR-camera-system-based unsupervised and weakly supervised (LCUW) network for 3D object detection.
• To develop methods for effectively leveraging both annotated and unannotated data.
• To enhance the accuracy and efficiency of 3D object detection in autonomous systems.

Main Methods:

• An independent learning mode for unsupervised data preprocessing of unannotated data.
• An Accompany Construction mode for weakly supervised preprocessing with minimal annotated data.
• A full aggregation bridge block for stepwise feature fusion and deepening representation.

Main Results:

• The proposed LCUW network effectively generates high-quality training data from unlabeled data.
• The Accompany Construction mode enables accurate detection with limited annotations.
• The full aggregation bridge block significantly improves feature representation and detection accuracy.

Conclusions:

• The LCUW network offers a robust solution for 3D object detection using LiDAR camera systems.
• The method advances the practical application of autonomous driving systems by addressing data annotation limitations.
• The approach demonstrates strong performance through comparative, ablation, and runtime experiments.