Stable Pose Estimation Using Ransac with Triple Point Feature Hash Maps and Symmetry Exploration

In this paper, a scene analyser is introduced which is based on Ransac (Random Sampling Consensus). This scene analysis approach is developed for robotic applications in particular, where poses of objects need to be estimated accurately that robots can grasp objects reliably. For assembly or manipulation purposes even an approximate pose estimation is not sufficient. For many objects appearance based similar poses exist, which influence the assembly strategy strongly when objects are gripped. Thus, robust pose estimation is required which is achieved by using triple point feature hash maps. This new feature vector is compared to two other feature vectors obtained from point pairs. It is shown that object poses can be estimated more precisely with roughly equal computation times with the new feature vector. Furthermore, in order to increase stability, symmetries are exploited and included into the entire scene analysis pipeline. The pipeline of the introduced scene analysis approach is illustrated and evaluated with various scenarios. The method presented here is successfully used for assembly applications.