Object Grasping By Robots Using Fast Single Shot MultiBox Detector

The basic research and industrialization of intelligent robots have been developing rapidly in recent years for wide applications ranging from smart home,industrial,agricultural picking to medical rehabilitation.Motivated by the fact that the intelligent level of robot in target recognition and target location is far from human intelligence,this thesis proposes to use deep learning for object grasping.It has been shown that deep learning is powerful in feature description which result in better features than the conventional methods.Consequently,it is of great significance to carry out the theory and system research of robot grasping based on deep learning.On survey of the state-of-the-art deep learning recognition algorithm,including YOLO(You Only Look Once),Faster R-CNN(Region-based Convolutional Neural Network),and SSD(Single Shot MultiBox Detector),The performances of the three algorithms are evaluated on VOC2007 dataset and our collected fruit dataset.Experimental results show that the performance of SSD method is the best in terms of speed and accuracy.Hence,the SSD method is employed for object grasping in this dissertation due to its unique feature in using multi-scale convolutional bounding boxes to improve detection performance.Although the recognition performance of SSD is generally good,there is space to further improve its speed,because large convolution kernels yield more parameters,and more delicate details of the feature map are needed for small targets.To this end,this dissertation focuses on optimizing SSD method to improve the speed and accuracy of object recognition.Also,the learning strategy of deep learning is deeply analyzed in this dissertation based on collected fruit database.The experimental results show that the improved SSD has better performance.In order to explore the mapping between twodimensional images and three-dimensional space,the three-dimensional localization using Kinect sensor is studied.Moreover,the kinematics model of the NAO robot is built.Based on aforementioned techniques,the experimental results for object grasping using NAO robot demonstrate that the system can effectively learn human's intelligence in target detection and location,and the system is robust in object grasping under different environments.