Collision risk assessment for autonomous robots by offline traversability learning
2012 (English)In: Robotics and Autonomous Systems, ISSN 0921-8890, E-ISSN 0921-8830, Vol. 60, no 11, 1367-1376 p.Article in journal (Refereed) Published
Autonomous robots should be able to move freely in unknown environments and avoid impacts with obstacles. The overall traversability estimation of the terrain and the subsequent selection of an obstacle-free route are prerequisites of a successful autonomous operation. This work proposes a computationally efficient technique for the traversability estimation of the terrain, based on a machine learning classification method. Additionally, a new method for collision risk assessment is introduced. The proposed system uses stereo vision as a first step in order to obtain information about the depth of the scene. Then, a v-disparity image calculation processing step extracts information-rich features about the characteristics of the scene, which are used to train a support vector machine (SVM) separating the traversable and non-traversable scenes. The ones classified as traversable are further processed exploiting the polar transformation of the depth map. The result is a distribution of obstacle existence likelihoods for each direction, parametrized by the robot's embodiment.
Place, publisher, year, edition, pages
2012. Vol. 60, no 11, 1367-1376 p.
Autonomous robots, Collision risk assessment, Embodiment, Robot navigation, Traversability learning, v-disparity
Engineering and Technology
IdentifiersURN: urn:nbn:se:kth:diva-107244DOI: 10.1016/j.robot.2012.03.004ISI: 000311654000003ScopusID: 2-s2.0-84867875036OAI: oai:DiVA.org:kth-107244DiVA: diva2:576849
FunderSwedish Foundation for Strategic Research EU, FP7, Seventh Framework Programme, FP7-ICT-2009-6-270212
QC 201212142012-12-142012-12-102013-01-08Bibliographically approved