Research On 3D Scene Reconstruction Of SLAM And Object Recognition Under Stereo Vision

In the field of visual SLAM,3D reconstruction is always a promising research topic whose goal is to reconstruct 3D model for spatial scenes and built infrastructure.Owing to 3D model is one of the most important basis for VR and AR,thus researchers must be faced to the problem that how to reconstruct the 3D model with high precision as soon as possible.Furthermore,3D visual perception nowadays has undergone a dramatic evolution,enabling various innovative applications such as self-driving cars,automatic scenes mapping and high-quality object scanning using commodity sensors,for instance,Microsoft Kinect,Leap Motion,ZED Stereo Camera,Asus Xtion,Intel RealSense,and so on.Admittedly,the ultimate goal of SLAM and vision technique is to make robots see and understand the real world,which will enable robots make decision and action planning automatically.In this work we have implemented a subset of the overall project.We have used ZED stereo camera to capture binocular images as the input.Then we have used various stereo image matching algorithms to generate the disparity map so thus to create 3D point clouds.Here,to enhance the global consistency and reduce the overall error of the reconstructed point cloud model,we have trained a data-driven matching local geometry feature to adapt to the noisy and low-resolution generated point cloud.As for the goal of helping robots to learn to see and understand the real world,the only task of 3D point cloud representation is not adequate.Furthermore,we try to make the point cloud represented in semantic form.Because semantic information allows robots to perceive the surrounding environment and to recognize objects.To achieve this purpose,we have implemented semantic segmentation of point cloud.And we trained the customized neural network model by using the S3 DIS Dataset,a famous indoor large-scale semantic segmentation dataset from Stanford University.The deep learning network architecture fully considers the characteristics of 3D point cloud:(1)Unordered,the sequence variations between input data have no effect on the results;(2)There is structural information between adjacent points;(3)The neural network should have invariance to any geometric transformations(translation,stretching,flipping,etc).So there is a general symmetry function constructed through the deep learning network.And as the result,the model finally realizes the semantic segmentation of the indoor 3D scene from the original point cloud reconstructed by the ZED stereo camera.Finally,a relatively complete semantic map of indoor environment is established.