Real-Time Grasp Detection Using Improved FMM and Cascaded Neural Networks

The successful grasping task for the robotic arm requires precise grasping posture. In this paper, we use the cascading depth network to predict the optimal object grasping pose. The model is mainly divided into two steps: i) generating a set of candidates that contain the regions of objects; ii) getting the optimal capture position by detecting the candidate region, and combining the depth image to obtain the three-dimensional coordinates of the capture position for objects. Due to flaws and edge noise in the depth image of Kinect, an improved FMM (Fast Marching Method) algorithm is used to repair the depth image hole, and then the joint bilateral filtering algorithm is employed to recover the edge noise of the depth image. Experimental results in public dataset and real scenes have demonstrated the effectiveness of the proposed method.