A user-oriented, comprehensive system for the 6 DoF recognition of arbitrary rigid household objects

En savoir plus sur le livre

The objective of this thesis is to develop a model-based object recognition system for the 6 degrees of freedom localization of typical rigid household objects, enabling intuitive teaching of new objects. The perceptual process of object recognition is divided into three main areas: data acquisition, object modeling, and object localization, each examined individually with distinct contributions. The recognition process begins with one-shot images of range and color data. While data acquisition often relies on single sensors, this thesis presents a novel sensor fusion technique that combines 2.5D input from stereo and range imaging systems to enhance data quality. For object modeling, a method for dense object modeling directly on the robot is introduced, alongside two stand-alone setups: a turntable and a chessboard for manual camera movement. Initial work includes a fastSLAM-based in-gripper object modeling approach, further developed with a bundle adjustment algorithm for faster registration. Two novel binary descriptors for textured and texture-less object modeling are proposed, enhancing descriptor computation speed and recognition rates. A scale-invariant extension of the binary descriptor ORB is introduced, alongside a global histogram-based descriptor for texture-less objects. To improve robustness in texture-less recognition, data association is constrained spatially, and an adaptive sliding window approach is propose