3d Shape Reconstruction In Traffic Scenarios Using Monocular Camera And Lidar

In the near future, a self-driving car will be able to perceive and understand its surroundings by composing a 3D environment map at object level. In this map, the 3D shapes of surrounding objects will be precisely reconstructed. The technique to reconstructing 3D object shapes using a monocular camera and a Lidar is presented in this paper. The proposed approach combines deep neural networks with an optimization process called 3D Shaping in which object pose and shape are jointly optimized. A significant performance improvement by the proposed approach in estimating object 3D orientation and the occupancy bounding box is proven through quantitative evaluation.