R-CNN Based 3D Object Detection for Autonomous Driving

3D object detection plays an important role in autonomous driving, which provides 3D location information of objects for subsequent decision-making modules. Existing 3D object detection algorithms can be divided into three ways: lidar-based, stereo image-based, and monocular image-based methods. Lidar-based methods depends on large and expensive lidar sensors to provide depth information which largely increases expense. Stereo image-based method mostly uses stereo images input with multistage networks which causes large computation cost thus limiting their using scenarios. Meanwhile some scholars proposed methods like Deep3DBox which only utilize monocular image input and can obtain competitive precision. However, their lack of depth property brings about unstable performance. To deal with that, we propose a novel method which uses both monocular image and cascade geometric constraints to obtain robust detection. The framework is divided into two stages. The first stage processes the monocular image input using key points-based detection network CenterNet with additional branch to regress the orientation, dimension, and center projection of bottom face. In the second stage, increasing IOU threshold can filter out unprecise 2D bounding boxes which cause performance degradation. After that cascade geometric constraints are utilized to obtain the final 3D box output. Our framework doesn’t depend on any external sources or subnetworks and can be trained end to end. We tested the proposed method on the KITTI-3D benchmark to test its ability and efficiency.