Grasp Detection Based on Faster Region CNN

Abstract

The robot grasp technology has recently aroused increasing research interests thanks to its foundation and importance in the field of robotics. Based on the deep learning method, this paper introduces a grasp detection model with the improved model Faster Region Convolutional Neural Network (Faster-RCNN). The orientation of the ground truth box in the grasp detection is random, so the orientation issue is one of the key points in grasp detection and differs from the other object detection researches. To tackle with this problem, this paper applies the five-dimensional parameters to represent the grasp rectangle. The method puts forward the improved Region Proposal Network (RPN) to export the tilted graspable region, including the size, the location, the orientation and the score belongs to the grasp class or non-grasp class. The RPN extracts the candidate proposals via using a more efficient CNN, instead of the inefficient selective search method. In the classification branch, the softmax function works to determine whether the anchor box is foreground or background. The regression of the angle is performed in the regression branch. In addition, the improved Non-Maximum Suppression (NMS) is used to generate the optimal inclined predicted grasp rectangle. To cope with the insufficient data size in the Cornell Grasp Dataset, the data augmentation and transfer learning methods are applied in the training phase. During the test, the results reveal that the detection accuracy of the model proposed in this paper on the dataset is 92.3% in terms of the image-wise splitting and 92.5% with respect to the objective-wise splitting on the Cornel Grasp Dataset, respectively.