The Research And Implementation Of Robotic Scene Understanding Supported By Cloud Computing

By using the cloud computing,big data infrastructure to enhance the robot's ability to carry out tasks in complex environments,cloud robotics is regarded as the "key to raise the next generation of robots".The cloud can amplify the robot capability,however,it also brings uncertainty,such as the high network transmission delay,the controllable performance and so on.This is unacceptable for many robotic applications,especially when they interact directly with the physical world.How to get a proper balance between capacity expansion and quality of service is a challenge that cloud robotics needs to deal with.In this paper,we take the robotic scene understanding as the research task.Based on the cloud robotic architecture and mechanism,we study to address these challenges.Robotic scene understanding refers to the process of recognizing the scene on the basis of the perceptual data.Robot should not only obtain the scene geometry information,but also the semantic information(such as object label and location).Cloud services,such as object recognition that already exist on the Internet,can provide support for the latter,but as mentioned above,the delay is long and unstable,so the quality of service cannot be guaranteed and the public service cannot be directly applied to robots.Therefore,this paper proposes a hybrid cloud architecture for robotic scene understanding in open environment,and we study the key mechanisms under this architecture.Specifically,the main work of the project is as follows:(1)Propose a hybrid cloud architecture for robotic scene understanding in open environment.To address the problem that the Internet services cannot provide the required quality of service for robotic applications,we propose a hybrid architecture consisting of two layers: mission cloud and public cloud.The former has controllable resource and knowledge related to the task,and thus can recognize the familiar objects in real time.The latter is based on Internet big data and can recognize the unfamiliar objects in open environment.We deploy the object detection engine in the mission cloud for the SVM(Support Vector Machine)and the deep learning classification algorithm separately.(2)Design the collaborative mechanism in the hybrid cloud architecture.The core of hybrid cloud architecture is the collaborative mechanism of the mission cloud and public cloud,so that the they can work together seamlessly.The familiar objects can be quickly recognized in the mission cloud to obtain the final results,and the unfamiliar objects can be successfully filtered and uploaded to the public cloud for rerecognition.Based on the characteristics of two kinds of mission clouds which are oriented to SVM and deep learning classification algorithm,this paper designs the corresponding cooperative working mechanisms and the applicable conditions are given through theoretical analysis of performance gains.(3)Design a label-location mapping mechanism based on data fusion.Scene understanding requires that the semantic information returned by the cloud be properly annotated on the environment map.In this paper,a label-location mapping mechanism based on data fusion is proposed,which converts robot position information,object depth and angle information into object position information,and then completes the robot scene comprehension process by adding semantic tags.Based on the above architecture and mechanism,the prototype system of cloud robot scene understanding is designed and implemented.Based on CORE and Faster R-CNN,the experiments are carried out and the data are verified.