Research On Robot Positioning Technology For Indoor Environment Based On Panoramic Vision

Simultaneous Localization and Mapping(SLAM)technology,as the excellent direction of space analyzing and processing technology,helps a lot for mobile robots to independently plan their movements.In recent years,the visual SLAM technology,as the branch of SLAM,has become a hotspot in this research field due to its relatively low cost of sensors and convenient installation.At present,the development of the visual SLAM algorithm still has defects,such as the limited field of view of ordinary cameras and the high computational costs.A catadioptric camera with a 360° field of view was selected as the visual sensor and its internal parameters were obtained.In the monocular conditions,the amount of information that the camera can obtain is maximized by using it.And a semi-direct monocular visual odometry(SVO)algorithm with low computational costs is selected as the basic framework of the system.The open source version of SVO is not perfect.This paper analyzes its algorithm flow and proposes three improvement methods:1)Aiming at the problem of rejecting outliers in the acquired data,the graph-Cut Random Sample Consensus(RANSAC)algorithm was applied.This algorithm improves the effect by introducing a local optimization(LO)step,and treating standard RANSAC as an energy minimization problem;introducing maximum likelihood estimation;and adding spatial coherence to the energy term.The graph cut algorithm is used to optimize the internal and external points' division,and the outliers' elimination is based on it.2)Aiming at the scale uncertainty of the monocular camera,a scale restoration method based on geometric constraints was applied.Firstly,feature points are matched on the input image,and then Delaunay triangulation is performed on the matched feature points.The geometric model of the triangle area is compared with the road geometry model to determine whether the triangle area is a road area or road geometry model is updated.The scale recovery of the SLAM system is completed according to the final road geometry model.3)Considering the problem of the depth uncertainty of map points,an unsupervised monocular depth estimation method based on the Convolutional Neural Network(CNN)was applied.Firstly,the two sets of input images are used to simulate the left and right sequences of the binocular images to reconstruct each other.Then,the reconstructed disparity map is compared with the corresponding images to obtain the loss function,and subsequently the original image is reversely reconstructed.By using this way,the convolutional neural network can be trained.Finally,it is possible to estimate the depth value only by inputting a single image.By using the estimated depth value as the prior depth information instead of the average depth value of the key frames of the scene used by the original algorithm,the depth uncertainty of map points can be effectively reduced.This paper proposes a visual SLAM system based on panoramic vision,which improves the accuracy of real-time camera pose estimation and the quality of localization maps by taking advantage of the large angle of view of the panoramic camera and three improved algorithms.Through experiments on indoor data sets and indoor actual environment,an internal point rate of 82.68% was achieved,the average translation error was as low as 1.18%,and the average rotation error was as low as 0.0027deg/m.Compared with other algorithms,the positioning accuracy was improved 63.23% averagely,verifying the effectiveness of this algorithm and its advantages over other algorithms.