Research And Implementation Of Indoor Positioning System Based On Monocular And IMU Fusion

In some indoor intelligent equipment applications,positioning problem has been difficult to solve.How to obtain location information with low cost and high precision has become the key to be solved in indoor positioning.With the advantages of small size,low cost and abundant information,visual SLAM(Simultaneous Localization and Mapping)has been widely used in indoor positioning research.Monocular vision has simple structure,small amount of calculation.The visual SLAM of direct method is generally faster than that of feature point method.So monocular direct vision SLAM has high application value.However,the visual SLAM of the direct method is proposed based on the condition that the gray level is constant,which has high requirements for the image and is vulnerable to factors such as illumination and camera exposure.It is also a common problem that monocular vision is difficult to obtain real scale.The visual inertial measurement unit(IMU)fusion scheme effectively solves the problem of monocular visual scale loss and poor robustness.This paper mainly studies the positioning scheme of visual-IMU fusion.Based on the direct visual inertial odometer VI-DSO,a modification scheme is proposed.The main work is as follows :In the initialization phase,the VI-DSO algorithm is initialized at any scale,and the IMU error estimation is calculated in the back-end optimization.Although the speed of system initialization is improved to a certain extent,the initialization ignores the IMU error,which will have a certain impact on the positioning accuracy.Based on the original algorithm,this paper improves the initialization scheme.In this paper,the maximum A Posteriori(MAP)estimation method is used for initialization.Firstly,the camera pose is estimated by pure vision,and then the vision and IMU are jointly optimized to estimate the parameters of IMU.In the depth filtering stage,the VI-DSO algorithm uses a depth filter based on Gaussian model.The model judges whether the measured data is calculated by comparing the variance of the measured data.Due to the artificially determined threshold,the judgment of mismatch points can only be limited to a certain range,and the effect of external points is limited.In this paper,a more complex Gaussian-uniform model is used to determine the outliers.The model determines whether the measured data is calculated by the mismatch point by estimating the probability of whether the measured value is an external point.This judgment method has a wider range of effects on external points.At the same time,in order to improve the accuracy and efficiency of matching calculation,the NCC(Normalized Cross Correlation)matching algorithm is used to calculate the matching point,and the matching pixel block is redesigned to calculate.In the back-end optimization process,the VI-DSO algorithm selects the Gauss-Newton method when optimizing the parameters.Although the structure is simple and the solving speed is fast,the solution is unstable.In marginalization,only feature constraints,motion distance and other factors are considered to design the optimization window.The influence of special motion states such as uniform velocity and static state on sliding window is ignored.This paper improves the back-end optimization scheme based on VI-DSO algorithm.The Levinburg-Marquart method is used to replace the Gauss-Newton method to solve the parameters.The judgment of the motion state is increased in the marginalization,and the marginalization strategy is optimized.The finiteness of the algorithm is verified by data set test on each improvement scheme.The test results show that the initialization speed of the proposed algorithm is about 33 % higher than that of the VI-DSO algorithm,the mapping effect is also significantly improved,and the positioning accuracy is improved by about 34.5 %.Compared with other mainstream VIO schemes,it also has certain advantages.Finally,a simple mobile test platform is constructed,and the improved algorithm is transplanted to the platform,and the positioning accuracy is tested in the real scene.The results show that the average error accuracy of the proposed algorithm is about25 % higher than that of the VI-DSO algorithm,and the indoor error is less than 5 cm,which has good positioning ability.