Topological scene recognition and visual navigation for mobile robots using omnidirectional camera

Abstract

Mobile robots are intelligent agents that are designed to serve people in various circumstances. With time, more and more robots will share the surrounding environments with humans, for example, service robots in families and autonomous cars on the road etc. In all these cases above, it is important that the robots are able to provide an efficient representation of the environment, which on one hand can be used as a common understanding among all involved subjects in the scenario; on the other hand, can support the robots to fulfill assigned missions. Psychological works demonstrated that humans can understand their surroundings by mostly topological meanings. I adopt this concept all across this thesis. An example can be observed as for service robots. In most situations, human may like to control the house-hold robot by naming go to “kitchen” or “grandpa’s bed”. Topological modeling of the environment would greatly help these applications. Moreover, navigation among topological nodes in the environment is essential to carry out practical services. Based on these observations, with this thesis, I present several contributions in both the fields of robotic scene recognition and visual navigation. In the first part, I consider the environment modeling using an omnidirectional camera as the only sensor. I present a topological scene recognition algorithm, which is modeled by a Dirichlet Process Mixture Model (DPMM). Topological segments with a similar appearance in video sequences can be automatically clustered and recognized in real-time. In the second part, regarding the navigation among topological nodes, waypoint-based topological modeling of the environment is presented, centered on a visual homing framework using Image-Based Visual Servoing (IBVS). A finite state-machine is then adopted, which enables topological navigation between waypoints by fusing Visual Homing and odometry motion. Experiments on datasets and real environments show the competence of the proposed algorithms against state-of-the-art.