Research On The Key Technologies Of SLAM Based On RGB_D

SLAM(Simultaneous Localization and Mapping)is recognized as the core of autonomous navigation for intelligent mobile robots.Since the launch of Microsoft's first depth camera Kinect in 2010,it has attracted a large number of visual SLAM researchers because of its low cost,abundant information collection and direct access to pixel depth information.SLAM research based on depth camera is called RGB-D SLAM.It is not only a hot topic in the field of robot autonomous navigation,but also one of the most challenging topics.Classical visual SLAM is based on feature points.It is mainly divided into four parts: front-end visual odometer,back-end optimization,loop detection and mapping.The frontend preliminarily estimates the position and posture of the camera and the landmark by processing a series of images collected by the camera.At present,the mainstream algorithm of the back-end optimization is non-linear optimization.It receives the information of visual odometer and loop detection to optimize the initial estimated camera position and posture and landmark,and obtains the globally consistent map and posture information.Loop detection is used to detect whether the camera has a loop,and if a loop occurs,the cumulative error can be eliminated.The main contents of this paper are as followsFirstly,by studying the imaging principle of the camera,we can get the distortion parameters which need to be calibrated.Zheng You Zhang calibration method is used to calibrate the two cameras of Kinetic camera and realize the angle alignment between the depth camera and the color camera.Secondly,the front-end of SLAM algorithm is mainly studied and improved.Three feature point algorithms commonly used in visual SLAM are studied.ORB is selected as the feature point method of this paper through real-time comparative experiments.The combination of threshold method and RANSAC method is used to eliminate mismatches.The experimental results show that the accuracy and efficiency of this method are greatly improved.Aiming at the problem that the depth error of Kinect generation camera affects the accuracy of ICP,the non-linear optimization method is used to solve PnP instead of ICP to solve the pose transformation of adjacent frames.The method of relocation is introduced into the visual odometer.The uniform velocity model is introduced as the initial value of PnP when the accuracy of RANSAC is insufficient when the matching points are few.Thirdly,mainly studied the back-end algorithm of SLAM.Back-end algorithms are mainly divided into two aspects: non-linear optimization and loop detection.Aiming at the non-linear optimization,the main BA algorithm and pose map optimization algorithm are studied at first.Then,by analyzing the advantages and disadvantages of BA and pose map optimization,we choose pose map optimization as the non-linear optimization algorithm of our SLAM scheme.Aiming at loop detection,the principle of eliminating accumulated errors by loop detection and the scheme of loop detection based on BOW are introduced.Specifically,it can be divided into: how to use K-means++ algorithm to train the dictionary expressed by K-fork tree offline,how to use TF-IDF to calculate the weight of each word in the dictionary,how to use the words in the dictionary to represent an image,how to quickly screen out the key frames containing common words from a large number of key frames,and how to calculate the similarity between images.The advantage and disadvantage of optimization can satisfy the optimization of pose maps in large-scale scenes.The process of loop detection based on BOW is introduced.Finally,TUM data set is used to test the algorithm,and the obtained camera pose is used to stitch point clouds.The test results are compared with those of other papers in terms of positioning error and operation efficiency.The results show that the method used in this paper can achieve better results.