PhD Thesis

3D pose estimation in complex environments

Work default illustration

Information

  • Started: 19/04/2010
  • Finished: 07/04/2017

Description

Although there has been notable progress in the pose estimation literature, there are still a number of limitations when existing algorithms should be applied in everyday applications, especially in uncontrolled environments. This thesis has addressed some of these limitations, namely, pose computation for uncalibrated cameras, pose computation without knowing the correspondence between 2D and 3D points, the computation of the pose when the points of interest are not reliable and the computation of the pose using exclusively depth data.

The problems addressed, and consequently the contributions, have been analyzed in increasing order of complexity. In each new stage of the doctoral thesis the existing restrictions to obtain the 3D pose of the camera were increased.

The first contribution focused on providing a technique for obtaining the pose of an uncalibrated perspective camera with better robustness and precision than existing methods. By means of the re-formulation of the perspective equations used in calibrated methods and the study of their numerical stability, an extended formulation of the perspective equations is obtained that offers a closed solution to the problem and a greater stability in the presence of noise.

The second contribution of the thesis has focused on the fact that most of the algorithms are based on having a set of 2D-3D correspondences. This task usually involves the extraction and pairing of points of interest. In this doctoral thesis we developed an algorithm that addresses the estimation of the correspondences between points and estimation of the pose of the camera jointly. By solving both problems together you can optimize the steps to take in a better way than solving them separately. In the works published as a result of this work the advantages inherent in this approach to the problem have been shown.

The third contribution of the thesis has been to provide a solution for the estimation of the camera pose in extreme situations in which the image quality is very deteriorated. This is possible through the use of learning techniques from high quality data and 3D models of the environment and the objects present. This approach is based on the notion that, using high quality data, we can train detectors that are able to recognize the objects in the worst circumstances.

The fourth contribution of the thesis is the creation of a pose estimation method that does not require color information, but only depth. By means of a definition of local volumetric appearance and the dense extraction of characteristics in the depth image a method is obtained that is comparable in precision to the state of the art but an order of magnitude faster.

The sum of these contributions in 3D pose estimation tasks has made it possible to improve 3D reconstruction, robotic vision and 3D map relocation tools. All contributions have been published in journals and international conferences and of reputed scientific prestige in the area.

The work is under the scope of the following projects:

  • RobInstruct: Instructing robots using natural communication skills (web)