Dynamic 3D-Vision

Abstract

Common state of the art mechanisms for the measurement of the surrounding environment in real time usually pose a trade off between high speed, robustness and accuracy. With applications for mobile robots in mind, this work focuses on the faster methods stereo analysis and PMD camera. Our research aims at the computation of robust and dense depth maps in real time. First, the performance of three standard stereo algorithms is examined with regard to two different measurements of similarity. The subsequent optimisation of the standard methods by using modern SIMD instructions and programming techniques like e.g. recursive subdivision leads to an increase of speed by a factor of four. As a result, for the WinnerTakes-It-All algorithm we achieve a computation time of 250ms (plus 230ms for building the cost matrix), which can be considered real time. The accuracy of our stereo setup is determined experimentally and a scene reconstructed from stereo data is shown. To improve robustness and speed the image derivative is evaluated. In order to deal with poorly structured environments experiments with an adaptive block size are conducted. The resulting disparity maps are more dense but the resulting block sizes for homogeneous regions slow down the correspondence analysis. Hence, this approach is not suitable for real time. A fast post processing step dealing with a disparity histogram is introduced to discard wrong matches. The subpixel disparity is computed as a measure of plausibility. As a comparatively new technique the PMD camera is used for distance measurement. The PMD camera provides directly the depth information for every pixel without the intensive computation that characterises stereo analysis. The camera was calibrated with an accuracy of 5 percent for distances over 1.5m. For smaller distances a more complex model than a pinhole camera with a point light source is needed. We observe that the measurement error increases quadratically with the distance, which is an effect of the reduced amount of light received from distant surfaces. Other inaccuracies result from the reflectance properties of the recorded surfaces or extreme geometric arrangements of the scene. A big advantage of the PMD camera is that it does not rely on the texture of a surface or the visibility of objects in a second camera like a stereo camera. With an (adjustable) integration time of 80ms per image it is also much faster. On the other hand, the stereo camera has a higher image resolution as well as higher depth accuracy. Especially the behaviour on and near edges is better. We thus made experiments combining both methods and they turn out to compensate the disadvantages of each other very well. As a result, we obtain robust and dense depth information in real time.