MUFFIN-HGCN: Multi-Feature Fusion Hierarchical-GCN for 3D Object Detection

3D object detection play a pivotal role in various applications, such as autonomous driving and environmental perception. However, the challenging task of detecting targets (e.g., vehicles and pedestrians) in 3D point clouds obtained from LiDAR sensors is hindered by two primary factors. Firstly, point clouds are non-Euclidean data and lack rigid data properties. Secondly, these point clouds are sparse and deficient in semantic information. In this paper, we propose a novel deep learning framework for 3D object detection. The proposed approach involves two stages. Specifically, we first present an innovative fusion technique that combines images and point clouds using transformers. This technique enhances the integration of semantic and geometric features while accurately mapping points to pixels without relying on camera extrinsic parameters. Additionally, a multi-level graph convolutional network (GCN) architecture is further introduced for object detection. This architecture effectively handles the non-Euclidean characteristics of point clouds while achieving precise object identification within them. A comprehensive series of experiments on the KITTl and nuScenes datasets validate the effectiveness of the proposed framework.