Object SLAM with Dual Quadric Parameterization

Conventional SLAM systems lack the ability to create semantically meaningful maps for scene understanding of robots. In this paper, we estimate a quadric surface for each object by detecting objects from different views and propose an object SLAM that uses dual quadric representations as 3D landmarks to overcome this limitation. A dual quadric can represent the position, orientation, size of an object compactly. We devise a geometric ellipse measurement model that addresses the problem of reconstructed object projection, and demonstrate how to integrate it into the SLAM system in order to jointly estimate camera poses and constrained dual quadric parameters. Our method is valuated on the public dataset. Experiments show the validity of creating maps with high-level information.