Quadric-based Traffic Sign Landmarks Initialization for Object-oriented EKF-SLAM

Abstract

Visual Simultaneous localization and mapping (vS-LAM) are estimation problems in which a vehicle is localizing itself while reconstructing a map of the environment. Such approaches are usually solved using handcrafted 2D features resulting as simple geometrical shape landmarks. The main problems of such features are: (1) they are not robust enough and (2) are often lost or incorrectly matched over consecutive frames and (3) they do not carry any semantic information. In this paper, we propose to consider higher level semantic objects as landmarks. We focus on Traffic Signs that are excellent candidates for this role: they have a simple geometrical shape, can be easily detected and carry a strong semantic information. We propose to include such new high level landmarks into an EKF-SLAM framework. For this purpose, we developed an efficient way to estimate the 7 degrees of freedom of polygonal and circular traffic signs from a single stereo-vision observation to initialize such landmarks in a very accurate way so it can be easily used in the EKF based estimation process.