VEM-SLAM - Virtual environment modelling through SLAM

Abstract

The problem of mapping a real environment and recognizing the objects contained in this environment is a problem in the Computer Vision area and has received attention with the advances of SLAM solutions and robust object recognition solutions. The problem with Simultaneous Localization and Mapping (SLAM) robotics is to create a (generally geometric) map of the scene while estimating the viewer’s pose. The solutions to this problem are used in several areas where a map of the environment is desirable and extract geometric information from it. Object recognition allows us to identify the object in the scene according to the object classes of the reference database. For recognition in 2D images, the best solutions are based on convolutional neural networks. However, to obtain the 3D geometric information of the objects in the scene, other techniques are necessary that vary according to the model of the reference 3D object. In this work, we present a new approach to deal with the pose estimation of 3D objects from images of static scenes of indoor environments. We also propose a new integration between an object detector and a monocular SLAM solution based on keyframes. As results we demonstrate an improvement in the estimation of the camera’s trajectory in relation to the original method and a use of the system implemented in the creation of virtual environments.