Context-Aware Fusion for 3D Object Detection in LiDAR-Camera Systems

LIDAR and camera are essential sensors in autonomous driving systems, but each has shortcomings in handling various complex environments. Therefore, many studies use fusion technology to bridge the gap between sensors to achieve better perception performance. This paper proposes a novel context-aware decorator (CAD) to densify and decorate the point clouds for 3D object detection. The purpose of CAD is to provide the point cloud with prior knowledge of object position and context features. It would make the point cloud texture richer in the object region. The decorated point cloud can be fed to the LiDAR-based 3D object detectors for classification and 3D bounding boxes regression. The LiDAR-based detectors combined with the CAD experiment showed that our method could improve by 4.24% mAP for PointPillars, and at least provide +2% mAP widespread improvements for other methods in the KITTI benchmark.