Corner Detection Based Real-Time Workpiece Recognition for Robot Manipulation

Workpiece recognition is vital and essential for robot manipulation which is one of the most important skills for robot. In this paper, we present a corner detection based strategy to recognize the type of workpiece so as to offer important visual cues for robot manipulation. Our framework works by three steps. Firstly-, the bounding-box of workpiece is detected using a multi-scale convolutional neural network according to the closed regions (enclosed by edges). Secondly, three types of corners (Y-type, A-type and L-type) are detected by employing a simple neural network and the results are further refined according to both the detection probability and geometric relationship between corners. Finally, the type of workpiece is recognized on the basis of the relative position of corners. Due to that the workpiece detection step greatly reduces the searching space for the corner detection, a real-time process is achieved. Another important characteristic of our method lies in that the detected corners can be further used as basic modelling element in reconstructing the 3D structure of the workpiece, which is beneficial for the robot to decide the grasping position and pose. With the PKII-HR6.4) platform, a manipulation controller is established based on the proposed approach. Experimental results show that our approach is comparable with some state-of-the-art work in precision of recognition, and our PKII-HR6.0 robot is able to precisely recognize and locate the Tetris-shaped building blocks so that to well accomplish the manipulation tasks.