Mathieu Gonzalez Ph.D. Defense

The goal of SLAM (Simultaneous Localization and Mapping) is to estimate the trajectory of a moving camera while building a map of its environment. Classical algorithms usually build a purely geometrical and homogeneous map, thus there is a semantic gap between the internal representation and the real world in which the system is evolving. Our goal during this thesis was to build a SLAM system that can harness semantic information to contrain the SLAM and improve its performance. To this end, we first propose a light neural network to estimate the pose of objects in the scene. Objects can serve as high level landmarks for a SLAM system, improving camera pose and adding information into the map. This network however has to be trained for specific objects. We then propose a SLAM system that can create clusters of 3D points corresponding to generic objects in the scene. With some a priori knowledge about object classes we can estimate their geometry in real time to improve both the map and camera pose estimation. Finally we propose new SLAM able to robustly estimate camera pose in dynamic scenes and to estimate the trajectories of all moving objects in the scene. A priori knowledge allows us to constrain the movement of objects to be plausible with respect to the structure of the world. We also propose to improve object tracking by injecting LiDAR data into our SLAM system.