Research On Visual SLAM Algorithm In Point-featureless Environments

SLAM(Simultaneous Localization and Mapping)technology aims to enable robots to simultaneously locate themselves and build maps of the surrounding environment in an unknown environment without knowing their location.With the rapid development of robotics,the demand for visual SLAM technology is also increasing.Especially in recent years,the wide application of RGB-D cameras has greatly facilitated the realization of visual SLAM.The classic visual SLAM is divided into four parts: front-end visual odometer,back-end optimization,loop closure detection and map construction.This paper mainly studies the solution to the problem that SLAM system using artificial point features does not perform well in the point-featureless environments.The environment of point-featureless will have a great impact on the front-end visual odometer and loop closure detection.Therefore,this paper focuses on the two parts in a modular way.In front-end visual odometer,aiming at the problem that it is difficult to estimate camera pose by using point feature visual odometer in the environment of poor point feature,this paper constructs a visual odometer module with line feature as the core to avoid using point feature and improve system accuracy,and introduces depth map inference method to supplement the missing part of depth map.Random sampling consistency(RANSAC)is used to backproject the sampling points of two-dimensional line segments into three-dimensional line segments,and the camera pose is estimated according to the matching relationship.Finally,in the optimization process,this paper is different from the traditional method of using the end points of line segments,but makes full use of the best crossing point in the fitting straight line,and introduces an angle error information,and deduces the Jacobian matrix of the corresponding error with respect to the pose disturbance.In the graph optimization,the reprojection error is used to optimize the pose,which is an extension of the traditional optimization method.In the loop closure detection part,aiming at the problem that the traditional point feature word bag model is difficult to construct in the point-featureless environment,a method based on depth learning is used to avoid using artificial point features.Through Alex Net convolution neural network,image features are directly extracted as image feature vectors,avoiding the use of artificial features.Because feature vectors are often large in dimension and are not easy to directly participate in similarity calculation,this paper combines PCA(Principal component analysis)linear dimension reduction method to reduce thecalculation amount,and formulates corresponding key frame selection strategy,finally determines the loop closure relationship.Finally,the effect of the algorithm is verified by data sets.The experimental results show that the visual odometer method with line features as the core improves the accuracy of the system and enhances the tracking performance of SLAM system due to the introduction of structural information.However,the computation time for 3-D line segment fitting and pose solution still needs to be further improved.For loop closure detection method based on depth learning,loop relation can be found effectively.Because artificial features are not used,it has better effect than traditional word bag model in low texture environment.