Research On Key Technology Of Visual Localization Based On Indoor 3D Dense MAP

In recent years,with the rapid development of wireless communication technology,the demand for location services of users is also growing.Acquiring location information with smart mobile terminals to help life has become an indispensable part of people's lives.Location-based application services have gradually penetrated into various areas of life,and have shown good market prospects.However,the positioning services in the indoor scene are still in the research and development phase,and there are not many indoor positioning systems for large-scale commercial operations.Low positioning accuracy and high positioning cost are the main reasons restricting the development of indoor localization technology.Visual localization in indoor scenes has attracted more and more attention in recent years due to its own technical characteristics,such as high positioning accuracy and low positioning cost in practical applications.By analyzing the domestic and overseas research status of visual localization technology,it can be found that the existing vision-based indoor localization systems have the following problems: First,the visual localization algorithm requires high precision 3D dense maps,but currently there is no specialized 3D dense map construction algorithm for visual localization.Secondly,in the process of visual localization,when existing algorithms are used for database image retrieval,as the retrieval algorithms are not improved based on the characteristics of database images,the image retrieval efficiency is low,and the time cost of image retrieval is large.Finally,although the scale ambiguity problem in monocular visual localization can be solved by different methods,these methods do not consider the influence of camera position relationship on scale estimation in the process of determining the scale ratio.More importantly,there is no effective means to solve the cumulative error problem in visual localization.In response to the above problems,the research contents of this paper mainly focus on the following three aspects:Firstly,for the problem of low accuracy of 3D dense map construction,a 3D dense map construction algorithm is proposed in this paper based on multi-source data constraints.For indoor 3D dense map construction,in order to make the map construction algorithm suitable for different indoor environments,2D point clouds,3D point clouds and visual features are used and the Multi-dimensional Iterative Closest Point method is used to construct the 3D dense map.At the same time,the map optimization problem is converted to multi-objective optimization problem based on the image point optimization function and the space point optimization function,and Nondominated Sorting Genetic Algorithm II is employed to solve the Pareto optimality solution for the optimization problem.Thereby the local map optimization is achieved.In addition,multi-source data are used to detect the loop closure of the camera trajectory in the proposed algorithm,and then the map global optimization is achieved.Compared with the existing map construction algorithms,the 3D dense map constructed by the proposed algorithm has obvious advantages in camera position accuracy and map accuracy.Even in indoor scenes with low visual feature density,the proposed algorithm shows good mapping performance.Secondly,aiming at the problem of excessive image retrieval time overhead in indoor localization systems,a hierarchical clustering-based image retrieval algorithm for visual localization is proposed in this paper.In the proposed algorithm,database images are hierarchically clustered in the offline phase,and then in the online phase,the query image uses clustering results to retrieve database image s.Visual features of the images in the same scene have high correlation.Therefore,according to this characteristic of database images,image clustering methods based on global feature change point detection and local feature tracking are introduced in the proposed algorithm.On the basis of the database image clustering results,a search tree is created for image hierarchical retrieval,and then the query image is able to retrieve database images hierarchically.In addition,the time overhead of the hierarchical image retrieval algorithm is theoretically analyzed in this paper,and simulation results show that when database images are hierarchically clustered by the proposed algorithm,image retrieval efficiency is significantly improved.Compared with single-layer image clustering retrieval algorithms and other multi-layer image clustering retrieval algorithms,the proposed algorithm has obvious advantages in the number of retrieved images.Thirdly,aiming at the scale ambiguity and cumulative errors in monocular camera localization,a monocular camera position estimation algorithm based on drift detection is proposed in this paper.On the basis of the indoor 3D dense map,a scale estimation method is proposed based on the weighted least squares by fully taking into consideration of the influence of camera relative position relationship on scale estimation for visual localization.In addition,a camera position drift detection method based on map interaction is proposed in the algorithm.Based on the information interaction between the camera and the map,cumulative position error s of the camera are calculated by Line Model Maximum Likelihood Estimation Sample Consensus algorithm,and then position drift estimation of the camera is achieved.The possible outliers in the camera absolute position estimation algorithm is analyzed in this paper,and the outlier elimination method based on the Chauvenet criterion is given.In addition,the uncertainty of the camera position estimation algorithm is theoretically analyzed.Simulation results show that the proposed scale estimation algorithm is able to improve the absolute position estimation accuracy of the camera,and cumulative position errors of the camera can be estimated by the drift detection algorithm.By timely switching the localization methods,the error accumulation problem in visual localization can be effectively solved.