Pose Optimization Of Indoor Vision SLAM Via Vanishing Points

Simultaneous localization and mapping(SLAM)is the technique that the robot locates its position and attitude by repeating the observed feature of the map in the course of the movement,and then constructs the map according to its position increment,so as to achieve the purpose of simultaneous positioning and map construction.Instead of using the expensive laser scanners,visual SLAM uses only images(RGB images and Depth images)to achieve real-time positioning in the map.However,the traditional visual SLAM has the trajectory drift problem due to the error cumulating,which often needs loop closure detection to eliminate the cumulative error.In this paper,we discuss the use of the structured line segments in the indoor scene to extract three mutually orthogonal vanishing points,and then use the vanishing points as global observations to optimize the SLAM pose while reducing the cumulative error.First,we propose a real-time vanishing extraction algorithm based on global searching,which can extract the three orthogonal vanishing points in the scene.Based on the technique of polar grid building,the time consuming vanishing points validation procedure is converted into a checking problem on the lookup table,which leads to the real-time performance of our method.Then,we added a vanishing point tracking thread to the traditional visual SLAM algorithm flow.For each frame of the input image,we extract both the feature point and the vanishing point on the current frame.Then we construct the graph model together with the local observations made up of the feature points as the global observations made up of the vanishing points.Then the method of graph optimization can be used to optimize the pose of the current frame.In addition,this paper analyzes the problem of SLAM pose optimization based on graph optimization theory,and gives the construction method of graph model under different observation values.In the experiments,the proposed algorithm is validated on the YUD public data set and compared with the other three commonly used algorithms.Experiments show that the proposed algorithm is the most accurate in not only the original YUD dataset,but also in the line segments automatically detected by the LSD.In the SLAM pose optimization experiment,this paper first tested on the TUM public data set.Experiments show that the proposed pose optimization algorithm based on vanishing point has strong robustness even in complex scenes.Taking into account the fact that the line segments on the traditional public data is not obvious,so we collected a new set of RGBD data via Kinect2.Experiments show that the proposed SLAM pose optimization algorithm based on vanishing point is 10%to 15%higher than the traditional algorithm in the absence of closed loop.