Multi-robot formation and cooperation using visual tracking

Abstract

Camera-based multi-robot formation is a core and challenging field of research in distributed robotics. The goal is to manage multiple robots maintaining them in a group, capable of extending for various robotic missions. There has been a substantial body of work mostly based on the following assumptions: (i) The availability of inter-communication capability among the robots, (ii) the use of special sensors such as omni-directional cameras, laser scanners, etc., and (iii) the use of special (or fiduciary) markers on some objects of interest (e.g. the leader robot). Those who have pursued successful multi-robot formations have always used at least one of these assumptions. However, these assumptions limit the scope of applicability of the solutions that depend on them. These systems have the following representative limitations: (i) The size of robots should be large enough to be able to handle the large payload of the special sensors, (ii) the use of inter-communication requires that communication systems be either already installed or carried by robots, and (iii) robots are expected to have predictable motions. Many researchers have tried to address the vision-based formation problem using some or all of these assumptions, but none has attempted to go beyond these assumptions yet. In this thesis we approach the problem described above using a monocular camera to address the size limitation. This, in turn, allows us to use much smaller robots than in previous research. We also allow unpredictable motions of the robots, so that our solution can have broader applicability to several missions, such as, search, safety, surveillance, exploration, etc. In addition, because we impose fewer restrictions in the assumptions than in prior work, we are faced with various challenging issues when investigating object tracking and robot control. We present a new robotic formation methodology along with a unique technique for a group of robots based on visual tracking only. Instead of relying on prediction models about a target robot, our robots execute target tracking from consecutive images. For real-time processing, we propose a new object tracking method, limiting the pool of candidates and enhancing the matching performance. Our new multi-robot formation technique has great potential for many robotic applications. Accordingly, this technique motivates us to come up with a new solution to the multi-robot coverage problem. Specifically we deal with the challenging problem of determining the optimum set of robots and their optimized paths, given a set of locations of interest. In contrast, existing approaches to the coverage problem deal with a given set of robots and attempt to find optimized paths. Therefore, in this thesis, we have come up with a new approach, which addresses the coverage problem for a variety of real world applications.