Research On Indoor Positioning Method Based On Visual-Inertial Navigation Combination

People’s demand for indoor positioning is expanding,which is principally reflected in the hope that people can stably carry out high-precision positioning for a long time,and there is no need to set up a base station and the cost is low.Among many indoor positioning technologies,vision-based positioning methods rely on low-cost cameras to provide rich visual information and do not require additional auxiliary equipment,and are gradually becoming the mainstream indoor positioning solutions.However,the positioning accuracy of vision based positioning method decreases due to the problems of too fast rotation,obstacle occlusion and sensitivity to ambient light,which sorely affects the positioning accuracy and soundness of the whole system.The inertial measurement unit is able to allotment the acceleration and angular velocity of the platform in actual time at an extraordinarily high sampling rate,hence it can precisely reckon the motion state in a short time,which is complementary to the camera at low frame rate.At the present time,the methods based on visual-inertial data fusion are primarily separated into filtering method and graph optimization method.For the most part,with the same amount of computation,the fusion method based on graph optimization can frequently achieve more desirable accuracy and durability.Consequently,this paper will use graph optimization method to fuse visual and inertial data,and thrive it to some extent.The main content of the paper mainly includes the following parts:In the first place,the review with reference to current situation of visual and inertial navigation information fusion are summarized,and the vision and inertial navigation are introduced.The vision part mainly includes the camera model,coordinate system transformation and attitude description.The inertial navigation part substantially includes the pre-integration calculation of inertial measurement unit and the derivation of error propagation model.Secondly,aiming at the information redundancy caused by adjacent image frames,a key frame screening strategy based on cooperative game theory is proposed to reduce the information redundancy,and the algorithm also takes into account the data association between key frames.It winds down the penalty of computer resources,warrants the permanent operation of the system,and helps the positioning accuracy by stamping out some poor key frames.The RMSE value of ATE on Eu Ro C datasets reaches 0.051 m.Then,in order to make the system in the thesis have globally consistent trajectory estimation,a deep learning-based loop closure detection algorithm is studied.The polymerization of unsupervised learning and convolutional neural network have the ability to pluck out the philosophical features that are not susceptible to luminosity.For loop closure detection,the vector similarity search method is used to ensure the efficiency of loop closure detection.The final performance and time tests on datasets that are available publicly elaborate the validity and robustness of the algorithm.In the end,three experiments are designed according to the different environment and motion state.The integrity of the system is proved by two rounds of experiments in the laboratory;The trajectory estimation results of complex motion in the laboratory prove the stability of the system;Experiments on a corridor with similar scenes and sparse feature points further prove the robustness of the system.