Automatic Complex Scene Acquisition And Reconstruction Based On Mobile Robot

With the rapid development of robot and 3D sensing technology,how to understand scene especially the complex indoor scene becomes more and more important to many regions like human-computer interaction,simultaneous localization and mapping,virtual reality and so on.This paper aims to solve the problem of active scene reconstruction and recognition based on mobile robot.Firstly,the robot needs to construct the space and get the 3D model of the scene.Then,the understanding of the scene model can be started with the recognition of the single object and then extend to scene recognition.Therefore,this paper describes from the following three aspects: the reconstruction of three-dimensional scene,recognition of single object,scene recognition.3D reconstruction is wel studied,so we choose Kinect Fusion technology to reconstruct the scene.Our main point is the complexity of the indoor scene explore and reconstruct.In this paper,a method of robot path planning based on tensor field is proposed.Tensor filed on ground is constrained by scene and update which guide the robot.At the same time,we also optimized the trajectory of the hand-held scanner which making it smoother.For a given model,we divide the recognition problem into two parts: the fine-grained object recognition and the next best view(NBV)prediction.We focus on fine-grained object recognition.Due to the large amount of model data,a single classifier can not complete the work,so we propose a tree-like hierarchical classifier to realize the recognition from coarse to fine objects;Objects are often occluded because relationships between objects are complex which one single view is not enough to identify.So we propose the problems of next-best-view.Inspired by handwriting recognition,we introduce the reinforce learning to predict the next-best-view.The ultimate purpose of identifying a single object is to understand the scene.For a picture of scene,it's hard for CNN to recognize it due to high cost and the complex relationship of objects,so we introduce reinforce model.Reinforce model takes one patch to identity and decide the next patch location,and then integrate all the information to understand the relationship between part and overall,leading to understanding of scene recognition.At last,the above methods are intergrated into the robot system to achieve active scene exploration and identification system.