Research And Implementation Of Indoor Positioning Method Based On Monocular Top-View Camera

Monocular visual odometer is a key technology in the field of indoor mobile positioning.It establishes a mathematical model of the camera's trajectory in the real world by tracking the point-line-area features in the video stream image.Considering that the indoor grid-type ceiling has a large number of structured line features and plane features that are favorable for positioning,an indoor positioning system based on a monocular top-view camera tracking ceiling features is designed,and implemented on an embedded platform.The module of line feature initialization,matching,pose optimization and coordinate optimization is improved based on PL-SVO.The ceiling plane features optimization module is added to the system.In this work: 1)A triangulated depth estimation based on matching line segments is proposed.Compared with depth estimation based on matching points in PL-SVO,it got higher initial coordinate accuracy and faster convergence speed.2)Line features are selected with strict matching conditions,and processing for line segment midpoint matching and out of bound is add to improve the matching success rate.3)Three line feature reprojection error models is analyzed,and select the projected line-matching point model combined with feature weighting strategy to improve the positioning accuracy of the system.4)In order to reduce the impact of non-significant endpoint mismatches on the accuracy of line feature coordinates,the GaussNewton iterative method is applied to optimize the position of the spatial line where the line feature is located in the Plücker coordinate system,and the least square method is applied to find the best endpoint position on the straight line.5)The the ceiling plane features is treat as a priori condition to initially estimate the depth of the features and improve the speed of feature initialization convergence.At the same time,the plane features is used as a strong constraint to further optimize the world coordinates of the features on the plane.On an embedded platform,experiment results show that in the actual grid-type ceiling datasets the average drift error per meter of the system proposed in this thesis is 1.48 cm,which is 25.3% lower than that of PL-SVO;The frame rate of the system reaches 78 frames per second,which is 12.6% lower than that of PL-SVO.Compared with other mainstream visual odometers,the system proposed in this thesis has the highest positioning accuracy and excellent running speed in the grid-type ceiling datasets.