Robust 6d Pose Determination in Complex Environments for One Hundred Classes

Abstract

For many robotic applications including service robotics robust object classification and 6d object pose determination are of substantial importance. This paper presents an object recognition methodology which is capable of complex multi-object scenes. It handles partial occlusions and deals with large sets of different and alike objects. The object recognition process uses local interest points from the SIFT algorithm as features for object classification. From stereo images spatial information is gained and 6d poses are calculated. All reference data is extracted in an off-line model generation process from large training data sets of a total of 100 different household items. In the recognition phase these objects are robustly identified in sensor measurements. The proposed work is integrated into an autonomous service robot. In various experiments the recognition quality is evaluated and the position accuracy is determined by comparison to ground truth data.