Cooperative Building of Visual Maps by means a Team of Mobile Robots
Dr. Arturo Gil Aparicio
Building a map of the environment is an essential ability that allows a mobile robot to be truly autonomous, since maps are required for a wide range of robotic applica- tions. As a result, map building has generated a great interest and an active research com- munity. This problem has been denoted SLAM (Simultaneous Localization and Mapping) since it considers the situation in which a mobile robot constructs a map and, simultaneo- usly, estimates its pose within this map. This problem is considered inherently difficult, since noise introduced in the estimate of the robot pose leads to noise in the map and viceversa. To date, typical SLAM approaches have been using laser range sensors to build maps in two and three dimensions. Recently, the interest on using cameras as sensors in SLAM has increased and some authors have been concentrating on building three di- mensional maps using visual information obtained from cameras. These approaches are usually denoted as visual SLAM.
In this thesis we consider a feature-based approach to visual SLAM. In this case, a set of distinctive points in the environment is used as landmarks. Mainly, two steps must be distinguished in the observation of visual landmarks. The first step involves the detection of interest points in the images that can be used as reliable landmarks. The points should be detected from different distances and viewing angles, since they will be observed by the robot from separate poses in the environment. At a second step the interest points are described by a feature vector, which is computed using local image information. This descriptor is used in the data association problem, that is, when the robot has to decide whether the current observation corresponds to one of the landmarks in the map or to a new one. When the robot observes a visual landmark in the environment, it obtains a distance measurement and computes a visual descriptor. Next, the descriptor and the measurement are used to recover the landmark in the map that generated the observation. To sum up, the data association is a critical part of the SLAM process, since wrong data associations would produce incorrect maps.
When the robot moves around the environment it will observe the same visual land- marks from different angles and distances. This poses two different problems: First, the same point may not be detected in the images when perceived from different viewpoints. Second, the visual appearance of the point in space will change significatively when seen from various poses in the environment. As a result, it is of great importance the selection of detection and description methods that permit to extract robust landmarks in the en- vironment and describe them invariantly to scale and viewpoint changes. As a result, a chapter in this thesis is devoted to the evaluation of the detectors and descriptors typically used in visual SLAM.
An important subfield within mobile robotics that requires accurate maps is the per- formance of collaborative tasks by multiple vehicles. Multiple vehicles can frequently accomplish any task faster than a single one. However, little effort has been done until now in the field of multi-robot visual SLAM, which considers the case where several robots move along the environment and build a map. In this thesis we concentrate on this pro- blem and propose a solution that allows to build a map using a set of visual observations obtained by a team of mobile robots. We propose an approach to the multi-robot SLAM problem using a Rao-Blackwellized Particle Filter (RBPF). To the best of our knowledge, this is the first work that uses visual measurements provided by several robots to build a common 3D map of the environment.
The validity of the approach is showed by means of a series of experiments both using simulated data and real data captured with a team of real mobile robots. The results presented demonstrate that the approach is suitable to build visual maps using a robot team in a wide range of situations.