Showing papers by "Olav Egeland published in 2015"
12 Oct 2015
TL;DR: The use of a dynamic model to compute the particle filter prior gives improved tracking accuracy and reduces the required noise level in the model, which makes it possible to predict the motion of the object for use in robotic applications.
Abstract: In this paper we present a particle filter for visual tracking using a physical model of the object to be tracked. Moreover, a parameter estimation scheme is implemented to identify the physical parameters of the dynamic model. This is based on recently developed methods for online estimation of static parameters using stochastic gradient search methods. The use of a dynamic model to compute the particle filter prior gives improved tracking accuracy and reduces the required noise level in the model. This makes it possible to predict the motion of the object for use in robotic applications. The performance of the method is validated in experiments with visual tracking of a free swinging pendulum of the type used in robotic loading of objects for automatic paint lines.
6 citations
11 May 2015
TL;DR: The first contribution of this paper is that the Fisher information matrix is used to quantify the information content from each image feature, and this improves the robustness of tracking algorithm significantly compared to the standard approach based on an unweighted likelihood function.
Abstract: Handling moving objects with robot manipulators is a challenging task as it involves tracking of objects with high accuracy. An industrial application of this type is the loading and unloading of objects on an overhead conveyor. A robotic solution to this problem is presented in this paper, where we describe a method for the interaction of an industrial robot and a free swinging object. Our approach is based on visual tracking using particle filtering where the equations of motion of the object are included in the filtering algorithm. The first contribution of this paper is that the Fisher information matrix is used to quantify the information content from each image feature. In particular, the Fisher information matrix is used to construct a weighted likelihood function. This improves the robustness of tracking algorithm significantly compared to the standard approach based on an unweighted likelihood function. The second contribution of this paper is that we detect occluded image features, and avoid the use of these features in the calculation of the likelihood function. This further improves the quality of the likelihood function. We demonstrate the improved performance of the proposed method in experiments involving the automatic loading of trolleys hanging from a moving overhead conveyor.
1 citations