Research On The Visual Odometry Method Based On The Combination Of Point And Line Features

Visual odometry is a localization technology for mobile platforms(such as robots,unmanned vehicles,etc.)to estimate egomotion based on image information collected in real time.It is one of the key technologies for mobile robots to achieve autonomous navigation.In recent years,visual odometry have attracted the interest of a large number of researchers,and have great application prospects in the fields of unmanned driving,wearable electronic devices,AR and games.At present,most methods of estimating camera motion by visual odometry are based on point features.However,point features are more dependent on the environment.For example,in scenes with weak texture or illumination changes,point feature based algorithms tend to perform poorly.On the other hand,line segments are usually abundant in artificial structured scenes.But the instability of endpoints of line segments and the lost of connectivity of line segments complicate the matching of line segments.Based on the combination of points and lines,this paper further studies the methods to improve the robustness and accuracy of visual odometry.The main research results of this paper include:(1)A line intersection structure feature and its detection method are proposed.The detection method first detects the line segments in the image.For the case of the breakage of some line segments,the post-processing process of adjusting and merging the line segments along the maximum gradient direction is proposed.The effect of line segment extraction is improved;Then the coplanar line segment pair is selected by using the local spatial proximity strategy;Finally,each coplanar line segment pair is intersected to construct the line intersection structure feature.The experimental results show that the extraction results of line intersection structure features are more abundant and stable in weak texture structured scenes.(2)Based on the intersection structure of the line segment and the gradient information of the local area around the intersection point,a combined descriptor including the structure and the gradient information is proposed.Specifically,a circular area is constructed with the intersection of the line segments as the center.The region is divided into 8 sub-regions,and the orientation histogram of each sub-region is calculated and normalized.At the same time,the shape context is used to describe the structural information between the intersections of the segments,and the two are combined to obtain the combined descriptor.Finally,a matching method using angle information and combined descriptor is proposed.A pair of matching feature not only provides key-point-like matches,but also produces two separate line matches.The experiment shows that the combined descriptor and its matching method proposed in this paper have achieved good results.(3)In the pose optimization,a sampling-based spatial line segment estimation method is proposed.The method samples a plurality of uniformly distributed points in the line segment,and back-projects the points into the 3D space by using the depth information.Then a 3D line segment is fitted to these 3D points using RANSAC and PCA algorithm,which can reduce the impact of depth errors or linearity ambiguity.In the calculation of re-projection error,the cost function of the BA optimization is combined with the line,and the corresponding graph optimization model is constructed.The derivation of the corresponding Jacobian matrix analytical form is given.(4)Based on the proposed line intersection structure feature and optimization method of point-line combination,a complete RGB-D visual odometry system framework is designed and implemented,including feature matching,pose estimation,local map construction and BA optimization.The analysis and comparison of the experimental results show that the system has good robustness and accuracy,especially in the case of weak texture and illumination changes.