Object-level dynamic SLAM

Visual Simultaneous Localisation and Mapping (SLAM) can estimate a camera's pose in an unknown environment and reconstruct an online map of it. Despite the advances in many real-time dense SLAM systems, most still assume a static environment, which is not a valid assumption in many real-world scenarios. This thesis aims to enable dense visual SLAM to run robustly in a dynamic environment, knowing where the sensor is in the environment, and, also importantly, what and where objects are in the surrounding environment for better scene understanding. The contributions in this thesis are threefold. The first one presents one of the first object-level dynamic SLAM systems that robustly track camera pose while detecting, tracking, and reconstructing all the objects in dynamic scenes. It can continuously fuse geometric, semantic, and motion information for each object into an octree-based volumetric representation. One of the challenges in tracking moving objects is that the object motion can easily break the illumination constancy assumption. In our second contribution, we address this issue by proposing a dense feature-metric alignment to robustly estimate camera and object poses. We will show how to learn dense feature maps and feature-metric uncertainties in a self-supervised way. They formulate a probabilistic feature-metric residual, which can be efficiently solved using Gauss-Newton optimisation and easily coupled with other residuals. So far, we can only reconstruct objects' geometry from the sensor data. Our third contribution further incorporates category-level shape prior to the object mapping. Conditioning on the depth measurement, the learned implicit function completes the unseen part while reconstructing the observed part accurately. It can yield better reconstruction completeness and more accurate object pose estimation. These three contributions in this thesis have advanced the state of the art in visual SLAM. We hope such object-level dynamic SLAM systems will help robots intelligently interact with the human-existing world.Open Acces