Dynamic 3D-Vision

Measuring and recognising the surfaces of the surrounding world forms a ubiquitous problem in automation and robotics. The knowledge of the environment allows a flexible and autonomous behaviour in different situations. Stereo vision belongs to the most popular techniques for gathering this information because it provides the dense depth information necessary for complex grasping tasks. Compared to laser scanners stereo cameras also have the advantage of higher framerates, so they are widely used for mobile robots. Porta (2005) e.g. uses the Small Vision stereo system (Konolige, 1997) to enhance localisation of a mobile robot: Features from depth maps are used additionally to appearance based intensity features. Other examples include Zhu et al. (2004) or Kang et al. (1995). However, the main problems of stereo vision remain speed and robustness. In order to accelerate the time-consuming registration of the stereo images and avoid specialised hardware, Sun (2002) employs an intelligent subregioning mechanism which reduces the search space of the correspondence analysis. Another approach builds upon the usage of modern SIMD processor instructions as documented by Sunyoto et al. (2004). Kim et al. (2005) on the other hand achieve real-time behaviour by segmenting foreground objects from the background. Depth information is then only updated for moving objects. Of course this approach is problematic for mobile robots. The lack of robustness of stereo analysis for particular scenes mainly arises from depth discontinuities and ambiguous surface texture. Kang et al. (1995) avoid these ambiguities by projecting textured light on the scene, but this is no general solution. Kim et al. (2005) made experiments with an adaptive matching window to increase the accuracy near edges. Zhao and Katupitiya (2006) examined the effect of occlusion and developed a method that detects occlusion areas and adapts a matching window appropriately. To evaluate and compare the robustness of different stereo algorithms, Scharstein et al. (2001) propose a taxonomy for different stereo algorithms and create a testbed including stereo images with groundtruth. Using this testbed we will document the results of the software system for the computation of dense disparity maps presented here. Our stereo system unites some of the speed optimisations mentioned above and hence achieves real-time behaviour. The calibration procedure and some comments on the brightness change constraint will be given. We will also present results for the distance measurements with a PMD camera (“Photonic Mixer Device”, Schwarte (2001), Kraft et al. (2004)) which is a technique for measuring the distance of an object by the time of flight of an active infrared illumination. The calibration procedure and the specifics of the measurements will be described, especially for scenes with surfaces almost in parallel to the