An Intelligent Composite Pose Estimation Algorithm Based on 3D Multi-View Templates

Abstract

For service robots, intelligent grasping is a core step to accomplish lots of household tasks. The spatial pose estimation of target object is the prerequisite to calculate the grasping pose of manipulator and perform the intelligent grasping. This paper proposes a composite algorithm to estimate the pose of target whose templates obtained from multiple views. With the premise of successful grasping, we divide the household items into two categories based on the difference of the demanded pose accuracy, and use different algorithms to estimate the pose of two categories. For the object with high demanded pose accuracy, an improved pose estimation algorithm is proposed, which combines template-selected method based on VFH and point cloud registration algorithm of key points. Finally, the whole pose estimation algorithm is evaluated by grasping experiments. The result indicates that: when the template is extracted from only 12 views, the success rate of grasping is over 90%., and the average estimation time of the two kinds of objects are 254.9ms and 984.2ms respectively. In conclusion, the algorithm takes into account of the requirement of both accuracy and calculation speed for intelligent grasping based on sparse multi-view templates.