Filter Fusion: Camera-LiDAR Filter Fusion for 3-D Object Detection With a Robust Fused Head

The different representations of images and point clouds make fusion difficult, resulting in the suboptimal performance of 3-D object detection methods. We propose a camera-light detection and ranging (LiDAR) filter fusion framework for 3-D object detection based on feature secondary filtering. This framework uses two uncoupled object detection structures to extract images and point features and a robust camera-LiDAR fused head to fuse features from multisource heterogeneous sensors. Unlike previous work, we propose a novel four-stage fusion strategy to fully use unique features extracted from two uncoupled 3-D object detectors. Our network fully extracts heterostructural features through dedicated detectors, which makes the extracted information more sufficient, especially for smaller objects. In addition, we propose a difference function for more efficient fusion of independent features from uncoupled object extractors. We mathematically prove the validity of the robust fused head and verify the effectiveness of our filter fusion framework in a test scene and on the KITTI dataset, particularly in KITTI pedestrian detection. The code is available at: https://github.com/xuminglei-hit/FilterFusion