Research On Vehicle Integrated Navigation Method Based On Monocular-Visual/Inertial Sensors

With rapid development of the Chinese economy,the road area increases fast,and the private vehicle ownership rate of continues to rise.Subsequently,people's demand for location service is also increasing,which puts forward higher requirements for accuracy,stability and reliability of vehicle navigation system.The mainstream vehicle navigation system is based on the Strapdown Inertial Navigation System(SINS)and Global Navigation Satellite System(GNSS).However,in modern cities environment,where are many tall buildings and lush trees,the signal of navigation satellite is easy to be blocked or be influenced by multipath effect,which seriously affects the positioning accuracy of satellite navigation system.It will decrease the positioning accuracy and reliability of vehicle navigation system based GNSS.In view of the situation that GNSS signals is easy affected in the urban environment,and there are some structured architectures with line features in cities,it is proposed that the vehicle integrated navigation method based on monocular vision/inertia sensors to ensure the stability and accuracy when GNSS is unavailable.Studies in this thesis start from the monocular vision attitude assisted strapdown inertial navigation method.The relative attitude information which is calculated by extractied line features from monocular vision images.The sequence analysis and modeling of line feature attitude have been carried out.The description,matching and tracking of dynamic line features are explored to make the line feature can be tracked In dynamic case.Furthermore,the visual odometer based on point features,which are used in dynamic line feature description,is fused into the monocular vision/inertial navigation method.Then,the monocular vision attitude and velocity information assisted inertial navigation method is realized.And the images are pre-processed by Mask R-CNN.It is explored that using Long Short-Term Memory neural network to optimize inertial navigation information.A new information fusion method is explored that monocular vision attitude/velocity and inertial navigation information fusion based on Long Short-Term Memory neural network.The main research contents of this thesis are as follows:(1)In order to solve the problem that GNSS signal is easy affected in urban environment,it is proposed that monocular vision attitude assisted inertial navigation method,based on line features.Firstly,introduce the related coordinate systems and their transformation relationships.Secondly,the image distortion correction is carried out,the line feature is extract,and the attitude characteristic is calculated and translated into the navigation coordinate.Thirdly,the attitude information sequence is analyzed and modeled.Then,the attitude information is fused with strapdown inertial navigation system by improved Kalman filter under static condition.Finally,the monocular vision attitude assisted inertial navigation method feasibility is verified through static experiment.(2)In order to solve the problem that the line feature is difficultly to matching and trace in dynamic environment,a dynamic line feature description based on point feature and line feature is proposed.Firstly,the common line features and point features are compared and analyzed,and the suitable line feature and point feature are selected to describe the dynamic line feature.Secondly,based on the point feature matching method and the relationship between point feature and line feature,a dynamic line feature matching method is designed,which includes coarse matching and fine matching.Then the attitude information is obtained by the dynamic line feature tracing.Thirdly,the dynamic attitude information is fused with the strapdown inertial navigation system based on the improved untraced Kalman filter in dynamic condition.Finally,the feasibility of the proposed dynamic line feature description and tracking is verified by the dynamic experiment.(3)In order to make full use of dynamic line feature,a monocular vision attitude/velocity assisted inertial navigation method is proposed.Firstly,images are preprocessed by Mask RCNN to avoid the unstructured examples influence the dynamic line feature matching and tracking.Secondly,the monocular vision odometer,which based on the point feature in dynamic line feature description,is introduced into the monocular vision attitude assisted strapdown inertial navigation system.Therefore,the monocular vision attitude and velocity information assisted strapdown inertial navigation system is realized.Thirdly,the observability of the system is analyzed.Finally,the effectiveness of the proposed scheme is verified by experiments.(4)To explore a new information fusion method of visual information and inertial navigation information.An optimization method of inertial navigation information based on Long Short-Term Memory,and an information fusion technology of monocular vision attitude/velocity assisted strapdown inertial navigation system based on Long Short-Term Memory are proposed.Firstly,the structure and hyperparameters of Long Short-Term Memory networks are introduced.The structure determines that Long Short-Term Memory can be used for navigation information optimization and information fusion.Secondly,the optimization method of strapdown inertial navigation information based on Long Short-Term Memory is proposed.The distribution of the hyperparameters of the Long Short-Term Memory networks is obtained from the simulation data training and testing.And The distribution is used to evaluate the applicability of the proposed navigation information optimization method.The method is evaluated and verified by experiments.Thirdly,it is explored that monocular vision attitude/monocular velocity/inertial information fusion based on Long Short-Term Memory method.The hyperparameter distribution is obtained by simulation and used to evaluate the applicability of the proposed method.Finally,the positioning accuracy of the monocular vision attitude/velocity assisted strapdown inertial navigation system based on untraceable Kalman filter and Long Short-Term Memory are compared through experiment,to verify the effectiveness of the proposed method.In this thesis,both computer vision part and the information fusion part are explored by traditional methods and deep learning methods.For the image processing part,the point feature and line feature extraction,description and matching are all based on traditional computer vision technology.Meanwhile,the image preprocessing based on Mask R-CNN involves the image processing technology based on Neural Network.For the fusion methods of visual information and inertial information,the traditional method,Kalman filter,and the deep learning method,Long Short-Term Memory are discussed.