Research On Key Technologies Of Position And Attitude Measurement Based On Fusion Of Vision And Inertial Information Under Complex Indoor Environment

Position and attitude measurement of moving bodies in indoor environments is widely used in robot assembly workshops,building structure surveys,indoor rescue and disaster relief,and exploration in unknown environments.Pose estimation has become the key to ensuring accurate position and attitude of motion systems.Currently,the fusion of visual sensors and inertial sensors is one of the important methods for pose estimation in indoor environments.However,when people are faced with an unknown,low-light indoor and complex texture environment and irregular movement,the traditional way of visual and inertial information fusion will be severely limited.This paper focuses on the research of pose estimation in the harsh indoor environment and irregular movement.We start with the system framework,and separately analyze the definition of image data,the accuracy and speed of feature points and line extraction,and the accuracy of multipose estimation based on multi-sensor fusion.This thesis proposes three key technologies of inertial-assisted motion image deblurring,feature point and line extraction method,and visual and inertial data fusion method.The main research contents and research results are as follows:(1)Aiming at the problem of the loss of image detail information obtained by traditional imaging models in low light,long-exposure imaging methods can be used to reproduce textures,but the image will be blurred due to motion in the longexposure.An inertial sensor-assisted blind restoration method of motion blurred images is proposed.According to different motion states,the causes of image blur are analyzed in detail,and several methods of objective evaluation without reference are introduced to provide basis for image quality evaluation of different restoration algorithms.The estimation of Point Spread Function and image deconvolution algorithm are two key technologies for image restoration.A linear,discrete,and mixed-motion Point Spread Function model is established during the exposure time,and then the Reinforcement Learning(RL)iterative model is constructed in reverse order in combination with the blurred image.In the iterative process,image evaluation is carried out by means of objective image evaluation without reference,and finally the optimal image quality is achieved through layerby-layer iteration.(2)Aiming at the universality of indoor natural landmarks,a combination of deep learning-based feature vector line segment group detection and Line Segment Detector(ILSD)based feature vector line segment extraction is proposed to achieve feature line segment retrieval.Through in-depth study of the characteristics of indoor environment images,narrow semantic images are introduced.A method of semantic segmentation of indoor environment based on image gradient is proposed to realize the segmentation of ceiling and wall semantic images.In the semantic image,the characteristics of the vanishing line and the plumb line vector group are summarized.The region-based Fully Convolutional Networks(R-FCN)based detection of vanishing line vector group and plumb line vector group is proposed,which can effectively reduce the detection error rate.Then,ILSD is used to extract vector line segments,and the speed and reliability of extraction are effectively improved through screening and clustering methods.(3)Aiming at the limitation of the fusion of visual and inertial information in special indoor environment,a vision and inertial data fusion algorithm combining tight coupling and loose coupling is proposed.The time domain and spatial domain consistency of vision and inertial sensors are analyzed in depth,and the visual and inertial initialization model is established.Based on the feature vector line segments extracted from high quality images,the feature point model is constructed,and then the visual pose model is established.In view of the zero bias and random error of the inertial sensor,a multi-rate Kalman filter error compensation method combining online and offline is proposed.The online uses of the fusion of visual and inertial information of adjacent frame calibration method for zero bias estimation,and the offline uses of Allan variance method for random error estimation,and then establish the inertial pose model.Aiming at the two states of indoor corner area and high speed transient state and linear normal state,an extended Kalman filter combining loose coupling and tight coupling is proposed to realize data fusion.The indoor imaging motion platform is built,and the data obtained is used to analyze and compare with traditional vision,inertia,and fusion models,which demonstrates the feasibility and effectiveness of the algorithm in practical applications.From the moving objects’ position and attitude estimation under complex indoor environment,the thesis explores the blind restoration of moving images,the detection and extraction of feature points and lines of indoor environment images,and the organic fusion of visual and inertial data.It provides new solutions and technical support for fast and accurate pose estimation in challenging indoor environments.