PhD Thesis

Estimació del moviment de robots mitjançant contorns actius

Work default illustration


  • Started: 27/10/2003
  • Finished: 27/10/2007



This thesis deals with the motion estimation of a mobile robot from changes in the images acquired by a camera mounted on the robot itself. The motion is deduced with an algorithm previously proposed in the framework of qualitative navigation. In order to employ this algorithm in real situations, a study of its accuracy has been performed. Moreover, relationships with the active vision paradigm have been analyzed, leading to an increase in its applicability.

When perspective effects are not significant, two views of a scene are related by an affine transformation (or affinity), that it is usually computed from point correspondences. In this thesis we explore an alternative and at the same time complementary approach, using the contour of an object modeled by means of an active contour. The framework is the following: when the robot moves, the projection of the object in the image changes and the active contour adapts conveniently to it; from the deformation of this contour, expressed in shape space, the robot egomotion can be extracted up to a scale factor. Active contours are characterized by the speed of their extraction and their robustness to partial occlusions. Moreover, a contour is easy to find even in poorly textured scenes, where often it is difficult to find point features and their correspondences.

The goal of the first part of this work is to characterize the accuracy and the uncertainty in the motion estimation. Some practical experiences are carried out to evaluate the accuracy, showing the potentiality of the algorithm in real environments and with different robots. We have studied also the epipolar geometry relating two views of a planar object. We prove that the affine epipolar direction between two images can be recovered from a shape vector when the camera motion is free of cyclorotation. With a battery of simulated as well as real experiments, the epipolar direction allows us to analyze the global accuracy of the affinity in a variety of situations: different contour shapes, extreme visualization conditions and presence of noise.

Regarding uncertainty, since the implementation is based on a Kalman filter, for each motion estimate we have also its covariance matrix expressed in shape space. In order to propagate the uncertainty from shape space to 3D motion space, two different approaches have been followed: an analytical and a statistical one. This study has allowed us to determine which degrees of freedom are recovered with more accuracy, and what correlations exist between the different motion components. Finally, an algorithm to propagate the motion uncertainty at video rate has been proposed.

One of the most important limitations of this methodology is that the object must project onto the image under weak-perspective visualization conditions all along the sequence. In the second part of this work, active contour tracking is studied within the framework of active vision to overcome this limitation. Both relate naturally, as active contour tracking can be seen as a focus-of-attention strategy.

First, the properties of zooming cameras are studied and a new algorithm is proposed to estimate the depth of the camera with respect to an object. The algorithm includes a simple geometric calibration that does not require any knowledge about the camera internal parameters.

Finally, in order to orientate the camera so as to suitably compensate for robot motion when possible, a new algorithm has been proposed for the control of zoom, pan and tilt mechanisms, and the motion estimation algorithm has been updated conveniently to incorporate the active camera state information.

For more information Go to thesis page

The work is under the scope of the following projects:

  • SGR ROBÒTICA: Grup de recerca consolidat - ROBÒTICA (web)