Research On Multi-camera Panoramic SLAM Algorithm Based On The Combination Of Points And Lines

While the robot is visiting a new environment,it uses sensors to collect data,extract information and compute from it to obtain its real-time position in the environment as well as to build a map of the surroundings.This process is the so-called Simultaneous Localization And Mapping(SLAM).Aimed to solve this problem,researchers have proposed many algorithms.Among them,the systems that use image data collected from cameras are called visual SLAM(v SLAM),which has become a research hotspot due to its inexpensiveness and convenience.Visual SLAM based on features is currently the mainstream method.Feature-based v SLAM extracts abstract geometric features from image sequences and performs data association to estimate relative pose between different frames,and then recover the camera trajectory and construct a sparse feature map.Common v SLAM systems use cameras with small field of view,hence the information acquired from each frame is very limited.Many visual SLAM systems make use of only point features in the environment to estimate trajectory and construct feature map.Point features have only one dimension and contains less scene information than other high-dimensional features.Point features in those systems are easy to lose track when the quality of image data is poor or the camera moves too fast,resulting in unstable tracking results,and feature map that is not intuitive enough to reflect the real environment.In view of the above-mentioned defects in the current feature-based visual SLAMs,this thesis uses the multi-angle data collected by the Ladybug 5+ multi-camera panoramic rig,and uses the method of joint calculation of point and line features to estimate the camera pose.Firstly,we use the Ladybug SDK tool to convert the collected original stream data into sequence image data,perform distortion correction on the image obtained by the original wide-angle lens,and obtain the internal and external parameters of the camera.We use the ORB algorithm to extract and match point features;Then we use the LSD algorithm to extract line segment features,and match them through the LBD descriptor.In order to simplify the calculation and to improve the stability of the pose calculation,an orthogonal parameter method derived from the Plücker coordinate system is used to describe the straight line features in threedimensional space.Then we use the graph structure to establish an optimization model of information obtained in the above steps,also to estimate the camera's posture and mutual relationship at each frame,and further optimize the feature position.At the same time,we use DBo W to detect closures of sequence images,and at last we visualize the acquired trajectory and features in the map of scene.The experimental results of above methods on multiple sets of public data sets and Ladybug 5+ multi-angle image data show that:(1)The line feature extraction and matching method used above can effectively extract and make use of line segment features with strong structural significance in the scene.(2)The tracking accuracy and stability of the tracking method combined with the point and line features are improved,compared with the method using only the point features.Three sets of monocular image experiments show that the accuracy of the point and line features combined with the solution method is improved by 2%,0.4%,and 0.2%,respectively.(3)Multi-angle data further improve the trajectory computation accuracy by 0.04%-0.11%.(4)The sparse map constructed using line features can reflect scene information more intuitively.At the end of the thesis,the method of multi-angle panoramic image data based on the point and line combined feature SLAM algorithm is summarized,and the future research directions are prospected and imagined.