Fitting superquadrics in noisy, partial views from a low-cost RGBD sensor for recognition and localization of sacks in autonomous unloading of shipping containers

A significant amount of cargo is transported in sacks, e.g., coffee and cacao, which are predominantly handled through manual labor, e.g., when being unloaded from shipping containers. There is hence a huge potential for automation. We present here a perception pipeline to recognize and localize sacks with a low-cost sensor. The pipeline is embedded in an industrial demonstration system for container unloading. In addition to the application background, there are two main contributions presented in this paper. First, we introduce a new numerically stable form of superquadric fitting. This is of interest for the application of superquadrics in general far beyond the concrete application scenario in this paper. Second, we introduce a fast convexity test between two neighboring patches that leads to a robust segmentation for sack/bag-recognition.