FRS-Voxel: A 3D-2D Hybrid Feature Extraction Network for 3D Object Detection

3D object detection is able to determine the precise location and orientation of an object within a 3D space. This capability is of utmost importance for autonomous driving. Voxel-based approaches offer the potential for enhancing 3D object detection precision and real-time efficiency. However, the voxelization procedure inevitably results in the loss of point cloud information, which in turn makes it easier to ignore distant points and reduces the precision of detection. The paper proposes FRS-Voxel, a 3D object identification network that operates based on voxels and uses voxel-based techniques to extract more detailed voxel and bird-eye view (BEV) data, resulting in improved precision in 3D object detection. We construct a FR3D module for sparse point cloud data and improve the ability of 3D CNN to extract additional voxel features. The CSA feature extraction module is designed for BEV features. It can accommodate more context-sensitive data because of a larger receiving field. We assessed the effectiveness of the proposed methodology using the KITTI dataset. Experimental results show that the proposed method is superior to state-of-the-art methods in terms of 3D average precision.