A Study Of Key Technologies On Indoor Simultaneous Localization And Mapping Based On RGB-D Camera

In recent years,with the rapid development of robotics technology,robots have appeared more and more widely in people's field of vision and have played a unique role in many fields.Simultaneous Localization and Mapping is a hot issue in the field of robotics.It is also very challenging and is considered to be the key to truly autonomous movement of robots.A camera named Kinect introduced by Microsoft in 2010 is a new type of visual sensor that can simultaneously acquire color images and depth images.This unique feature makes it attractive to researchers and developers and a new SLAM method called RGB-D SLAM has gradually become a hot topic in the research field.SLAM algorithms are generally divided into front end and back end.The front end obtains the pairing points of adjacent images through feature extraction and feature matching,and then uses this to perform inter-frame motion estimation.The back end mainly uses the closed-loop detection to increase the inter-frame constraint,and then optimizes the global pose to eliminate the accumulated errors in the front end matching process.At present,the RGB-D SLAM system still has the following problems:1 The feature matching accuracy is not high;2 The depth image has large noise which affects the motion estimation;3 The loop closure detection efficiency is low so it is difficult to apply it to large scale scenarios.In response to the above issues,this paper mainly did the following work:(1)The influence of different feature points on the accuracy and running speed of the algorithm is studied,including SIFT features,SURF features,and ORB features.Experiments have shown that SIFT features and SURF features can achieve better accuracy than ORB features.The price paid is a significant reduction in operating speed.If they are to be run into an actual system,they need to be accelerated by the GPU.The ORB feature's running speed is several times than that of SITF and SURF.At the same time,the accuracy loss is within an acceptable range.Therefore,the ORB feature is selected for subsequent experiments.(2)Aiming at the problem that the accuracy of feature matching is not high,a multi-level false match culling strategy is proposed.After using FLANN algorithm to obtain the rough matching result,ratio filtering and cross filtering method are used to perform the first elimination,and then RANSAC algorithm is used to perform further elimination to obtain more robust matching points.(3)In view of the noise existing in the depth image,the motion estimation algorithm is improved.The perspective N point algorithm is combined with the PROSAC algorithm and a PROSAC-PnP algorithm is proposed.The PROSAC algorithm first sorts the matching point pairs according to the matching distance and the ratio of the nearest distance and the second nearest distance,and then preferentially selects the sample from the sorted matching pair to calculate,which can speed up the calculation and improve the accuracy.(4)A loop closure detection algorithm based on visual dictionary is proposed for the problem that the loop closure detection is inefficient and difficult to be applied to large-scale scenes.The visual dictionary constructs a vocabulary tree of feature points in an offline manner,making it possible to quickly calculate similarity of images.The loop closure detection is used to increase the inter-frame constraint,and then the global graph optimization tool g2o is used to optimize the pose graph,resulting in a globally consistent camera pose and point cloud map.In this paper,a detailed experiment on the published data set was performed and the experimental results are given.The experimental results are analyzed from two aspects:robustness and real-time performance.At the same time,the paper also compares the experimental results with other literatures.The comparison results show that the algorithm is superior to other algorithms in terms of real-time performance and robustness.