|
|
|
|
|
|
|
| IAS-group | Research | Learning | Environment learning | ||
(See ISLA for the new IAS website. Most maintenance activities on this old website have stopped.)
ObjectiveOur goals are:
Research group members
Research AchievementsOur work on reactive navigation focuses on the use of vision. In particular we use the spatiotemporal structure available in the image sequences. We have developed robust methods for optic flowestimation, which outperform other methods. Based on the spatial structure derived from the optic flow a robot is able to perform simple reactive navigation tasks, such as wall following. The "structure from motion" task is usually solved by looking at the spatial structure of the optic flow field. We showed that the temporal structure in the optic flow can be used for a "wall following" task (Dev, Kröse, Groen, 97). Instead of first deriving the spatial structure of the environment it is also possible to use features from the optic flow directly to avoid collisions with obstacles (Dev, Kröse, Groen 98). This is an important step in the integration of reactive and planned behavior. At the moment we are using our expertise of optic flow in a visual-based homing of the robot. Our work on environment modeling concentrates on making a model in the sensor space, instead of making a model in the world space. We worked on this earlier (Kröse and Eecen 94}) using a neural network to model measurements from ultrasonic sensors, and use that model to localize the robot. Recently we started using a vision system to model the environment ("appearance modeling"). Experiments with on a real robot have been carried out. This robot had a standard camera, and by rotating the robot, new views of the environment could be obtained. An active vision strategy was developed (Kröse and Bunschoten 99}. Because the vision system provides high- dimensional images, an important
task is to find an optimal set of features. A schematical set-up of the
Markov robot localization approach can be depicted as follows:
 We investigated Principal Component Analysis for linear feature extraction and tested this on robot localization using range data (Vlassis and Krose 99) and data from an omnidirectional vision system (Kröse, Bunschoten, Vlassis and Motomura 99}. An information-theoretic performance measure was presented (Vlassis and Kröse 99) which can be used to evaluate the individual features. Apart from the features which are selected, the performance also depends
on the method which is used to model the relation between the location
of the robot and the sensor values. We developed methods for probability
density function approximation using Gaussian mixture models, in particular
method to find the correct number of Gaussian kernels (Vlassis and Likas,
99). We also developed methods to model conditional density functions with
mixture models using EM (Vlassis and Kröse 99).
|
