| With the rapid technological and social development,people have more requirements for the robot intelligence.As a basic function,grasping is playing a crucial role on the road of robot intelligence.Currently,the technique of grasping in a certain environment is mature,but it still lacks of autonomy and adaptability to the environment,which seriously affects the popularization and application of robot.Due to the fast development of artificial intelligence,robot s based on machine vision are showing an outstanding performance in environmental perception,dynamic decision and action control,so that more and more scholars show their interests in this field.However,the deep detection model with good training performance needs plenty of labeled training data,and its performance reduces sharply in cross-domain detection tasks.As to resolve these problems,this thesis carries out the research work on multi-object grasping detection based on domain adaptation,where multi-object grasping detection algorithm with hierarchical feature fusion,robust deep softmax regression against label noise and image style transfer using generative adversarial network is proposed to improve detection ability of cross-domain grasping detection.With the algorithm,robots can grasp objects steadily,faster and precisely in complex environment.The thesis consists of the following parts:(1)Aiming to solve the problems of traditional grasping detection algorithms,such as low detection accuracy in unstructured environment and weak ability of generalization,the thesis studies a multi-object grasping detection algorithm which is characterized with hierarchical feature fusion based on attention mechanism.This algorithm firstly extracts network features with different scales from multiple convolution layers of deep neural networks,then detects grasping poses for all target objects in complex environment by candidate area boxes.To decrease the number of parameters for detection model and increase the detection speed,a multi-function detection neural network model is built combined with object detection and grasping pose estimation modules,which realizes object grasping pose detection via an end-to-end way.Moreover,for taking full advantage of useful information and reducing noise interference of multilevel network features,attention mechanism is adopted while hierarchical feature fusing.The experimental results show that the accuracies of object detection and grasping pose estimation of the proposed algorithm are 95.31% and 87.10%,respectively.The grasping detection speed is up to 31.25 fps.The results demonstrate that the proposed algorithm can simultaneously detect objects and estimate their grasping poses fastly and exactly in complex environment.(2)In order to make full use of unlabeled samples of target domain to improve the cross-domain recognition capability for deep recognition model and against label noise,this thesis proposes a novel unsupervised domain adaptation by robust deep softmax regression.The proposed method sets up pseudo labels on unlabeled samples of target domain firstly,and then learns the deep recognition model using labeled source domain data and pseudo-labeled target domain data at the same time.To decrease the negative transfer of the recognition model caused by error labels of training data,robust deep softmax regression is built for target domain data according to the confidence of their pseudo labels.Theoretical derivation for the robust deep softmax regression is made as well.On experimental evaluations,image recognition accuracies on two benchmark databases,Office-Caltech and Office-31,improve more than 13% and 7% respectively,demonstrating the effectiveness of the robust deep softmax regression in unsupervised domain adaptation.(3)To improve robustness of machine vision for robot grasping system,a domain adaptation algorithm for multi-object grasping detection is studied.On one hand,cross-domain generative adversarial networks are constructed to generate pseudo-target images utilizing images of source domain,so as to reduce the discrepancy between source domain and target domain on the data level.Although generated images are much similar to real target images,especially their background,the target object characteristics and labeled information are consistent with those of source domain images.On the other hand,a domain anti-classifier is embedded into multi-object grasping detection module,which reduces the difference of latent network features of two different domains in the optimizing process of detection model.Furthermore,the domain adaptation algorithm mentioned above with robust deep softmax regression against label noise is also adopted to train an outstanding multi-object grasping detection model by using both source and target domain data simultaneously.Finally,three multi-object grasping dataset with different background are collected for evaluating multi-object grasping detection models based on cross-domain adaptation.And here comes the result that the accuracy of the proposed domain adaptation algorithm for multi-object grasping detection improves more than 40% on cross-domain detection tasks compared with non-domain adaptation algorithm,which verifies the effectiveness of the proposed algorithm.(4)In order to test the effectiveness of algorithms proposed in this paper,a series of experiments for robot grasping are carried out in real environment.Firstly,an intelligent robot grasp operating system based on machine vision is set up,and then works,such as camera calibration,hand-eye calibration of robot and motion control for object grasping,are accomplished.Finally,many robot grasping experiments,such as multi-object classification by robot grasping,novel object grasping and stacked bolt grasping,are carried out in unstructured environments.Experiment results show that multi-object grasping detection algorithm based on domain adaptation can guide robots to grasp kinds of objects intelligently in complex environment and shows a strong robustness in across-domain detection tasks.Meanwhile,it also has been proven that the grasping detection algorithm can be extended into unknown object grasping tasks.Hence,the study of this thesis lays a theoretical and practical foundation for further study o f robot intelligence grasping. |
