LiDAR-camera-system-based 3D object detection with proposal selection and grid attention pooling.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• 3D object detection is crucial for autonomous driving.
• LiDAR-camera systems offer complementary 3D structure and 2D texture data.
• Fusing data from LiDAR and cameras presents challenges due to differing sensor properties.

Purpose of the Study:

• To propose a novel LiDAR-camera based 3D object detection method.
• To improve the quality and recall rate of 3D proposals by integrating 2D image information.
• To develop an effective feature fusion strategy for enhanced detection performance.

Main Methods:

• A proposal selection module, including proposal addition (back-projecting 2D proposals onto LiDAR depth) and proposal filtering (using IoU matching).
• Grid attention pooling to weigh LiDAR and image features, generating salient fused features.
• Utilizing accurate 2D proposals from RGB images to refine 3D proposals.

Main Results:

• The proposed method demonstrates superior performance compared to existing approaches.
• Ablation studies confirm the effectiveness of individual components, particularly proposal selection and grid attention pooling.
• The method successfully generates effective fusion features from LiDAR and camera data.

Conclusions:

• The novel LiDAR-camera 3D object detection method significantly enhances detection accuracy and recall.
• The proposed techniques, including proposal selection and grid attention pooling, are effective for sensor fusion.
• This work advances the application of LiDAR-camera systems in autonomous driving.