Localization of 3D objects using model-constrained SLAM

Angelique Loesch, Steve Bourgeois, Vincent Gay-Bellile, Olivier Gomez, Michel Dhome

Abstract: Accurate and real-time camera localization relative to an object is needed for high-quality Augmented Reality applications. However, static object tracking is not an easy task in an industrial context where objects may be textured or not, have sharp edges or occluding contours, be relatively small or too large to be entirely observable from one point of view. This paper presents a localization solution built on a keyframe-based SLAM algorithm. This solution only requires a video stream of a 2D camera (color or grayscale) and the prior knowledge of a 3D mesh model of the object to localize (also named object of interest in this document). The 3D model provides an absolute constraint that drastically reduces the SLAM drift. It is based on 3D-oriented contour points called edgelets, dynamically extracted from the model using Analysis-by-Synthesis on the graphics hardware. This model constraint is then expressed through two different formalisms in the SLAM optimization process. The dynamic edgelet generation ensures the genericity of our tracking method, since it allows to localize polyhedral and curved objects. The proposed solution is easy to deploy, requiring no manual intervention on the model, and runs in real time on HD video streams. It is thus perfectly adapted for high-quality Augmented Reality experiences. Videos are available as supplementary material.