| With the rapid development and application of information technology,locationbased services has gradually become a link among technologies in various fields,among which indoor positioning technology is a research hotspot and one of the difficulties in addressing indoor location services.Visual positioning uses intelligent terminal cameras to perceive scene visual information so as to obtain camera pose,which can realize accurate positioning of indoor pedestrians;with the advantages of no facility dependence,strong heterogeneity,strong universality,and high positioning accuracy,visual positioning has become an important method for sensing pedestrian location based on terminal cameras.Due to the complexity and variability of large-scale dynamic indoor scenes,the following problems exist for pedestrian-oriented indoor visual positioning: as for large-scale indoor 3D maps,visual localization requires to obtain the initial poses in the global scene,which has the problems of long time consumption and visual deviation;in the indoor complex environment,there is less visual texture information or redundant features,resulting in low accuracy of feature matching and low efficiency of continuous pose solving for continuous visual positioning;in dynamic indoor scenes,visual features are occluded,changed or,missing,resulting in visual positioning failure and positioning jumping;and in the study of complex cross-floor and multi-scene visual positioning interior models,the design of visual positioning schemes and the exploration of model reliability are rarely carried out.In view of the above problems,this dissertation focuses on the smartphone-based indoor visual positioning technology with the collaborative information sensing of a variety of sensor as the main body,explores the indoor visual positioning methods based on geomagnetic and inertial information enhanced sensing,prioritizing on the research of geomagnetic and visual,inertial and visual information multiple sensingenhanced strategies,constructs a high-precision,continuous and reliable visual positioning model for the public,and conducts experimental testing and reliability verification in large-scale complex indoor cross-floor environments,with the main contributions focusing on the following four aspects:(1)For the problem of low initialization efficiency of visual positioning in largescale indoor scenes,this dissertation proposes an efficient initialization method for visual localization assisted by indoor geomagnetic features.The method uses indoor geomagnetic information to achieve global rough position recognition under large 3D maps,and a rapid estimation model of local 3D map for visual positioning initialization is constructed;the indoor non-visual geomagnetic features can effectively avoid the problem of visual deviation in similar indoor scenes,effectively alleviate the problem of error position recognition based on direct visual feature matching or image retrieval auxiliary matching,and improve the speed and effectiveness of visual localization initialization in large-scale scenes.At the same time,to address the lack of distinguishability of geomagnetic features in indoor complex topological space,a geomagnetic feature extraction and localization method based on multi-scale recurrent neural network is proposed,which uses coarse and fine-grained features of long sequence geomagnetic for spatial and temporal correlation feature extraction to improve geomagnetic representation variability,achieve stability and accuracy of indoor position recognition based on geomagnetic features under heterogeneous equipment,multi-users,and multi-scenarios,and ensure a priori location recognition accuracy and reliability of visual localization initialization.(2)For the problem of low feature-matching accuracy of visual positioning and low efficiency of pose solution in complex dynamic indoor scenes with similar visual features or few textures,this dissertation proposes a visual feature-matching and positioning optimization method based on built-in MEMS enhanced sensing of smartphones.This method takes advantages of high-frequency and continuity of the relative positioning mode of inertial sensors to construct a visual feature matching mechanism guided by the prior positional information of MEMS inertial sensors.The pose estimation algorithm considering the pedestrian motion constraint is studied,and a carrier stability prior pose solution and error uncertainty estimation model based on MEMS inertial sensor is constructed.This dissertation proposes to combine prior pose,error range,and camera model to reconstruct the local 3D point cloud of the scene,constructs an efficient solution model of a visual feature matching pool,and establishes a robust matching of visual features in complex indoor scenes,realizing the 2D-3D data correlation for pose solving.The internal point rate of visual feature matching in various complex scenes is increased to 63%,and the average time for continuous localization is reduced to 0.22 seconds,which significantly improves the visual feature matching accuracy and camera pose solving efficiency..(3)For the problem of visual feature matching failure due to light change or occlusionin dynamic indoor scenes and the existence of visual localization jumpings and discontinuities,this dissertation proposes an inertial and visual fusion indoor localization method based on factor graph optimization.Based on the inertial sensing of pedestrian motion state,this method constructs a factor map loosely coupled positioning framework for inertial and visual information joint sensing,explores the advantages of the complementary positioning method based on the relative positioning method of MEMS inertial sensors and the absolute positioning method based on visual sensing,effectively fills the positioning gap of visual feature matching failure in dynamic and texture missing scenes,and improves the effectiveness of position output when visual positioning fails by about 48% in various scenarios.Meanwhile,The global position constraint based on high-precision visual positioning method is established to reduce the cumulative error of PDR positioning based on inertial sensors,and realizes that the maximum positioning error in various scenarios is better than 1.15 meters,among which the average positioning errors in the hall,walking stairs and narrow corridor scenes are 0.43 meters,0.41 meters,and 0.46 meters,respectively,and the positioning accuracy is better than that based on single sensor positioning method..(4)For the design comprehensive performance test and analysis of continuous high-precision sensing model in large-scale cross-floor indoor scenes,this dissertation proposes an indoor localization model MIVI based on geomagnetic/inertial/visual multi-source fusion,and selects a complex multi-floor indoor environment for testing,including seven typical indoor scenes such as parking lot,walking stairs,lobby,corridor,etc.,and compares the proposed positioning method with four positioning models of PDR,3D-LBMS,VINS,and VBL.The experimental results show that MIVI significantly outperforms the inertial-based PDR and 3D-LBMS models;compared with the visual positioning model the average positioning accuracy in multiple test routes is about 85% better than the VINS positioning system,and the positioning effectiveness is about 42%better than the VBL model.Moreover,considering the influence of image resolution difference,handheld posture change,and device heterogeneity factors of visual positioning,a comprehensive test of model robustness is conducted,and the results show that the MIVI positioning model in this dissertation has the feasibility and advanced long-distance,cross-floor and multi-scene t tracking solution,which can realize stable and reliable high-precision continuous positioning in complex indoor scenes.This dissertation has 93 figures,23 tables,and 291 references in total. |
PDF Full Text Request