Object Grasp Detecting Based on Three-level Convolution Neural Network

Referring to the characteristics of object grasp of human, a three-level serial convolution neural network (CNN) for object grasp detecting is proposed to realize high-accuracy grasp of unknown objects. In the proposed three-level serial CNN, the first level network can locate the object position roughly to determine the position for the searching of the grasping rectangles in the next level CNN. The second level network is used to obtain the preselected grasping rectangles and get very less features with a quite small network, so as to quickly find out the available object grasping rectangles and to eliminate unavailable object grasping rectangles. The third level network is applied to reevaluating the preselected object grasping rectangles and get more features with a bigger network to exactly evaluate every preselected object grasping rectangle and obtain the best preselected object grasping rectangle. The experimental results validate that the grasping accuracy of the three-level serial CNN increases by 6.1% compared with the single CNN. Finally, the three-level CNN realizes high-accuracy grasping with an actual Youbot.