Monocular 3D Detection for Autonomous Vehicles by Cascaded Geometric Constraints and Depurated Using 3D Results

3D object detection is a key task in 3D vision perception of autonomous vehicles. In this paper, we present a novel two-stage 3D object detection method aimed to get a more accurate 3D location of an object. We modify existing 3D properties regressing network by adding two additional components, viewpoints classification and the center projection regression of a 3D box's bottom face (CBF). The center projection is associated with a similar triangle constraint to acquire an initial 3D location of a closed-form solution. For no truncated objects, the previous predicted location is involved in the initial value of over-determined equations constructed by the 2D-3D boxes fitting constraint with the configuration determined by the classified viewpoint. Then the recovered 3D information is utilized to purify the detection results. Results of comparison with state-of-the-art methods on the KITTI dataset show that although conceptually simple, our method outperforms more complex and computationally expensive methods. Furthermore, our method can filter out false alarms and false detection in both 2D and 3D results.