Work

Image-based Novelty Detection for Cognitive Mobile Robots

Year: 2012
Type: dissertation
Abstract: Robots need to be able to perceive their environment when they execute object manipulation tasks or when they interact with humans.To this end, robots are equipped with sensors to acquire visual and geometric information.Geometric representations of the environment are important for tasks where the robot has to measure the distances to the objects in the surroundings, e.g. during navigation or an object grasping process.Image-based representations acquired from camera images, in turn, provide information for object recognition and a reference model of the environment for the detection of novel objects.In the first part of this thesis a novel probabilistic appearance representation is investigated.This environment model is inspired by image-based scene representations and thus represents the colors of the robot's 3D environment at a densely spaced series of viewpoints.Instead of storing the raw color value captured by a camera, however, a Gaussian model is used for the luminance and chrominance at each pixel of a view.Using depth information and camera pose data stored at each viewpoint, the probabilistic priors can be interpolated at intermediate viewpoints, as virtual images are synthesized from image-based representations.The Gaussian distributions model the uncertainty of the appearance which can arise from erroneous pose or depth estimates or from moving objects in the scene.The expectation and uncertainty of the appearance of the scene are used for assessing the level of surprise of novel visual stimuli.As the performance analysis shows, the surprise measure is a detector for novelty which is superior to measures based on image differencing.Umgebung wird sie verwendet, um das Hinzufügen oder Entfernen von Objekten festzustellen, wobei Beleuchtungsschwankungen unterdrückt werden.Dies ermöglicht es dem Roboter zu unterscheiden zwischen neuartigen Ereignissen, wie neuen oder verschwundenen Objekten, die relevant für seine Aufgaben sind, oder nicht-relevanten neuartigen Ereignissen wie Beleuchtungsänderungen.In diesem Zusammenhang erlaubt es eine spezielle Methode, Spiegelungen auf Oberflächen zu identifizieren und zu unterdrücken.Daraus entsteht ein Ansatz, der in der realen Welt mit Oberflächen, die aus beliebigen Materialien bestehen, vielseitig eingesetzt werden kann.
Author: Werner Maier
Language: English
Open Access status: green
OpenAlex ID: W2749019279
Cites: 111
Cited by:
Related to: 9