Utilizing Robot-Mounted Camera to Improve Model Confidence

Abstract

To use a robot in a close contact operation relies on having an accurate 3D model of the environment to be able to position the robot without colliding. 3D models of process plants may be inaccurate, either because of differences between real equipment and CAD-based models, or changes caused by renewed equipment. This thesis investigates the possibility of using a single camera mounted on a 9 DOF robot manipulator, to identify a known object, and to iteratively improve the confidence of an inaccurate 3D model of the object s surrounding environment to such an extent that the model enables close contact operations. Algorithms that can be used to identify and estimate the pose of a known object is suggested. However, the main focus is on the 3D modeling system, which is evaluated by simulations in a synthetic environment and by robot lab experiments. A 3D model with uncertainty representation is presented. The representation is shown to enable conservative approximations of unmeasured objects, easy collision checking, and a convenient way of fusing overlapping measurements. A shape from stereo technique is introduced based on dense disparity maps using a constant stereo baseline. To handle surfaces with uniform textures as good as possible, the global graph cut algorithm by [Kolmogorov and Zabih, 2002] is used to match the stereo image pairs to calculate the disparity maps. The method is shown to have problems when matching surfaces with uniform color or with high reflectance, which limits the possible objects that can be modeled successfully. An uncertainty measure is introduced to model the limited accuracy of the disparity maps. A discrete Kalman filter approach is presented, to fuse overlapping measurements with different uncertainties into a uniform model. The approach shows successful results when measurements are taken with scattered viewpoints, but gives less accurate results when the viewpoints are clustered. To determine where to take new camera measurements, an algorithm for approximating the next best view is presented. Based on the intermediate modeled 3D model, the algorithm uses a greedy objective function to maximize the uncertainty removed. The overall system presents a framework for 3D modeling with conservative approximation of unmeasured objects to ensures collision avoidance when used by a path planner. The framework may by adapted to use other sensors as well, which may give an increased accuracy.