Real-time 3D Reconstruction Based On Monocular SLAM

Reconstruction of three dimensional scene is widely used in mobile robot,unmanned aerial vehicle,unmanned vehicle,augmented reality and virtual reality and so on.The research has an important significance and application value.Reconstruction of three dimensional scene by visual Simultaneous Localization and Mapping mainly has three aspects: probability estimation model based on bayesian,keyframe-based and bundle adjustment method,graphic optimization method.There will have three categories of difficulty when we use monocular vision SLAM.First,feature-based methods usually incude the extraction and matching of feature points,which will cause a large amount of calculation,lead to low accuracy and could not accurately describe the cuerrent scene and so on.Second,inherent scale blur of monocular camera and the depth information of the target object could not be obtained directly.Large scale track drift will arise after a long time.Third,the reality scene always with a large number of similar objects,shelter,moving objects,dramatic change in the depth and multi-closed loop and so on.Starting from the above three categories of difficulty,we will have a depth research about the system and algorithm of the reconstruction of three dimensional scene based on monocular SLAM.In this paper,an improved feature descriptor and optimization method were used in the traditional method which was based on feature points,the result showed that the robust and sustained tracking capability had been improved significantly.In the improved system framework: image pyramid and weighted Gauss-Newton optimization algorithm were introduced into the system for real-time and robust.The local semi-dense depth map was created by the depth map of keyframes,and then non-essential keyframes were removed and the local map was merged into the global map,which made the accuracy and real-time of the system had been improved significantly.We made Chow-Liu tree algorithm and bag of words model together to build vocabulary tree and in order to obtain the probability distribution of exterior features,then the similarity of scene was calculated by calculating the current scene and all features both in probability and similarity with the previous scene in order to form a cumulative matching sequence.Finally,an additional degree of freedom which was used to measure the scale and a graphic optimization framework were introduced into the system for posture graphic optimization.The system contains three threads: camera tracking,mapping and closed-loop detection.At last,the result showed that the system we designed could achieve more accuracy,robustness,real-time and flexibility.