Semi-direct Method Of Stereo Vision SLAM Based On Feature Point Optimization

Simultaneous Localization and Mapping(SLAM)is a hot topic in the field of intelligent mobile robot.With the development of computer vision technology in recent years,SLAM method based on vision has become a research hotspot.There are two main SLAM methods based on vision: feature point method and direct method.Among them,feature point method solves the problem by extracting feature points in the environment and establishing matching.When the environment texture is rich and the matching is correct,feature point method has high robustness and accuracy.It is the mainstream method of visual SLAM,and the image matching and inter-frame pose estimation involved are relatively mature.But this kind of method also requires the texture feature of the image to be obvious,and the effect is not good in weak texture or image blur.In addition to the high requirement for image features,feature extraction and descriptor calculation are time-consuming.The direct method is based on the assumption of image gray consistency,and the photometric information of the image is used to directly model and solve the problem.Since the time of feature extraction and descriptor calculation is saved,the computational efficiency is high.However,the direct method is sensitive to changes in illumination and may fail to track when illumination conditions are poor.In order to combine the advantages of the two methods,this paper proposes a fusion SLAM method with the direct method as the main body and the feature point method as the auxiliary optimization under the stereo camera platform.The main research results are as follows:(1)A semi-direct stereo vision SLAM method based on feature point optimization is proposed.Start with from coarse to fine strategy of image alignment between frames,when optimizing fails,photometric error and projection error are joint to optimization,estimating to get the initial position,and gauss-newton method is used to optimize the projection error,with outer point out mechanism to increase the dynamic and robustness under shade condition,reliable front-end tracking results are obtained.The back end adopts the nonlinear optimization based on key frame to obtain more accurate pose calculation results.(2)The closed loop detection based on DBo W3 is studied and introduced into the fusion SLAM method designed to improve the consistency of SLAM mapping.First,using ORB features to build DBo W3 based visual word bag library,the closed-loop detection is realized.On this basis,global optimization is carried out by using ceres.(3)In order to verify accuracy and real-time performance of this method,the standard and open datasets EUROC,KITTI and urban complex dataset are used.The positioning accuracy is compared with relevant methods,and the real-time performance is verified.Experiments show that the proposed method has good robustness,can adapt to a variety of environments,and can meet the real-time requirements.This thesis contains 46 figures,13 tables and 59 references.