| This thesis considers a mobile robot equipped with a vision sensor, called the observer, whose task is to track a moving target in an environment with obstacles. Neither the trajectory nor the motion model of the target, nor a map of the environment is known in advance. Instead, at every stage of the tracking process, the observer reads its sensor to acquire a “local” map (i.e., the observer's visibility region) and estimate the target's position in this map. It then uses this information to update the control (commanded velocity) applied to its actuators.; A key difference with tracking problems like missile control and pure visual tracking is the presence of obstacles combined with the observer's ability to move. The observer must exploit its mobility to prevent the target from eventually being occluded by obstacles. This thesis focuses on the question “Where should the observer go next?” and proposes a tracking algorithm that consists of iteratively updating the observer's motion by locally minimizing a function that models the risk that the target escapes the visibility region. This function is made of two components: a reactive component that estimates the lengths of the escape paths of the target and a look-ahead component that assesses the ability of the observer to react. This function is derived from a data structure, called the escape-path tree, for which this thesis presents an efficient computation algorithm. An integrated robotic observer using these techniques was implemented. This observer has limited sensing capability (a laser range sensor providing a 180° scan of the environment). Despite such limitation, it was remarkably successful at target tracking. We attribute much of its success to both the risk minimization approach it uses and the two-component risk function embedded in this approach. This experimental system was successfully tested in an indoor environment with a non-adversarial target performing brisk maneuvers like swift turns behind obstacle corners. |
