Research Of Real-time 3D Reconstruction Based On Stereo Vision

3D reconstruction technology for the robot has provided a reliable environment map,which laid the foundation for its navigation,path planning and obstacle avoidance.Based on the existing research,this paper focus on stereo matching and visual simultaneous localization and mapping,and we realized the process of 3D reconstruction based on binocular in real-time.This is mainly including the following contents:(1)Research on the camera imaging principle and theory of calibrationAs the most important sensor for 3D reconstruction,we need to study the camera imaging principle and the relationship between camera coordinate frame and other reference coordinate frames clearly.This paper summarizes the general idea of understanding the transformation between coordinate frames from the expression of coordinate frame.(2)Research on the theory of visual simultaneous localization and mappingVisual simultaneous localization and mapping is the key function of 3D reconstruction.Based on the theory of probability,this paper expounds and demonstrates the general ideas and methods to solve the problem of state estimation,which combined with the theory of Lie Group and Lie Algebra.(3)Research and implementation of stereo matching algorithmThe stereo matching algorithm can generate dense disparity graphs by rectified stereo pairs.In this paper,the local stereo matching algorithm based on convolution neural networks is selected after comparing the performance of various stereo matching algorithms.Based on the deep learning framework Caffe,a convolution neural networks for extracting image features is constructed and trained by KITTI database.Finally,the stereo matching algorithm based on convolution neural network is realized by combining the semi-global smoothing algorithm.(4)Realization of visual simultaneous localization and mappingIn this paper,the process of visual simultaneous localization and mapping combined the advantages of feature point method and direct method.ORB-SLAM2 is a new method of simultaneous localization and mapping based on feature point method.On the basis of it,the integration method can greatly save the development cycle.This paper rectifies its tracking part on the ORB-SLAM2 framework,which improves the efficiency of simultaneous localization and mapping.(5)Realization of 3D reconstructionIn order to further improve the availability of the map,this paper combines the 3D occupancy map based on octree,and transforms the point cloud map into 3D occupancy map.It is laid the foundation for the robot path planning and navigation in the future.