3D object detection plays an important role in autonomous driving, which provides 3D location information of targets for subsequent decision-making modules. Existing 3D object detection algorithms(eg, Stereo R-CNN) mostly use two-stage networks which causes large computation cost which limitating their use scenarios. Some scholars proposed CenterNet in an anchor-free way for conducting 3D object detection and achieved impressive performance. Therefore, this paper proposes a cascade framework for 3D object detection based on key points. The framework is divided into two stages. The first stage processes the stereo image input using CenterNet as the backbone with additional branch to obtain the orientation, dimension and 3D location. In the second stage, cascade geometric constraints are utilized to refine the 3D box output by filling out the false positive in previous stage. Experiments show that the proposed method can achieve better performance on the kitti 3D benchmark and improve the network efficiency.

CICTP