Template-based state estimation of dynamic objects

In order to plan their missions and to carry them out successfully, mobile robots operating in changing environments need to keep track of the state of objects. The perception of changes in the environment and the integration of changes into the robot's world model is therefore an important problem in mobile robotics. Most of today's systems plan their missions based on static models, thus limiting their applicability. We introduce a method to maintain environment models by estimating the state of changing objects, e.g. their current position and configuration, from sensor data. Unlike other methods, which acquire and maintain sub-symbolic environment models, our method automatically maintains a symbolic CAD model. The method proposed is a Bayesian state estimator which computes the maximum likelihood estimate of the state of a dynamic object by matching templates of the object against proximity information obtained by the robot. The algorithm employs Monte Carlo Markov localization to determine the robot's position in its environment. The localization provides a probability density of the robot's position, and matching takes this density into account, to achieve robust state estimates even while the robot is moving. Experiments carried out on a mobile robot in our office environment illustrate the capabilities of our approach with respect to the robustness of the state estimates.