Research On The Integrated Positioning Method Of Inertial/Visual Aided By Convolutional Neural Network

The current mainstream navigation system is the Strapdown Inertial Navigation System(SINS)and Global Navigation Satellite System(GNSS).However,in high-rise urban areas or indoor environments,satellite signals are denied(interruption,interference,multipath,etc.),the accuracy and continuity of positioning will be greatly affected.Therefore,only inertial navigation is applied in challenging environments such as GNSS denied environment.However,the simple inertial navigation accumulates errors as time increases,so other navigation elements need to be introduced to increase the accuracy and stability of the navigation.In recent years,visual navigation has been continuously developed and applied due to its advantages such as strong intuitiveness,high stability,and relatively low price.Among them,monocular cameras have become popular in research due to their low cost,small size,and convenient application.But the disadvantage of visual positioning is that it occupies a lot of system resources,and the processing speed is slow.In order to improve the processing speed,it can be used with offline training neural network for visual information extraction and classification.As an important part of image processing in machine learning,neural network has the function of quickly classifying and recognizing images.Therefore,the present paper has studied the vision/inertial integrated system aided by neural network,which realizes the function of suppressing the accumulation of inertial device errors in the GNSS denied environment.Aiming at the problem that it is difficult to accurately locate the latitude and longitude of the vision,the visual positioning method of this topic is the landmark positioning method using template matching.First,landmarks were set up and landmark information was obtained.The information includes size and location,and then save these landmark information as a location database.The selection of landmarks is random,not necessarily a single object,but can be a combination or a landmark with background.This topic has changed from simple to complex.The first subject studied is single-point positioning,i.e.,each landmark only matches one frame of picture.The advantage of this method is that the positioning accuracy is high and the size information of the landmark is not required.Once the matching is successful,the position of the carrier will be very accurate.On the other hand,the disadvantage of this method is that there is the possibility of missing the landmarks,and the carrier’s route requirements are high;and the way to make up is to increase the number of landmarks.The second method of visual positioning is the continuous matching positioning method,which is an improvement of the first method.Through the introduction of camera characteristic parameters and landmark size information,the landmark location library information is made more comprehensive.On the picture matched to the template information,the relative position of the carrier and the template is calculated by computing the size and position of the template;thereby the position of the carrier itself is obtained.For the problem of wasting resources on the processing of pictures without landmarks in the use of landmark positioning,the solution proposed in this paper is to introduce neural network to recognize landmarks in the picture,that is,quickly determine whether the picture contains landmarks and recongnize them.This study improved VGGNet network structure based on the Inception model to recognize landmarks.The recognition accuracy after training is improved compared with the traditional model.The recurrent network GRU model was improved.The specific method was to use the output of the convolutional neural network Squeeze model as the input of the GRU model,and then optimize the structure to form the Squeeze & GRU network structure.After training with this structure,the recognition accuracy was further improved,and the training time was greatly shortened.In order to reduce the complexity of the system this study uses a simplified integration of inertial sensors,3D Reduced Inertial Sensor System(3DRISS).Compared with the Inertial Measurement Unit(IMU),the system reduces 2 gyroscopes and 1 accelerometer,and adds an odometer that uses an accelerometer instead of a gyroscope to calculate the pitch and roll angle of the carrier.In this study,with the introduction of vision,visual odometry(VO)is used to participate in the RISS system,forming the VO-3DRISS system,which provides new ideas for further research.Meanwhile,the error that the traditional odometer accumulates continuously over time is eliminated because the visual odometer can perform scale correction with the help of the short-term distance traveled of the carrier provided by landmarks.Finally,through the Extended Kalman Filter(EKF),visual information and inertial navigation system are fused to form a neural network-aided vision/VO-3DRISS system,realizing positioning in a GNSS denied environment.This paper uses three sets of experiments to verify each algorithm: indoor robot experiment,indoor cart experiment and outdoor car experiment.The reference benchmark for the indoor robot experiment is the indoor map.The robot vision-aided INS solution trajectory and the single INS solution trajectory are simultaneously drawn on the indoor map to verify whether the positioning error of the single INS system can be improved;the indoor cart experiment uses two sets of inertial navigation systems: a set of low-precision mechanical gyroscopes and a set of high-precision fiber optic gyroscopes,compares the visual-aided low-precision INS solution trajectory and the single high-precision INS solution trajectory on the map to verify the performance of the system under the conditions of lighting and use of low-precision inertial devices;the reference datum for outdoor vehicle,mounted experiments is Global Positioning System(GPS)trajectory,in order to increase the accuracy of the datum,the experiment adopts Real-time kinematic(RTK)carrier phase difference technology,taking GPS trajectory as the real trajectory to verify the accuracy of this system in outdoor environment and locomotive speed.The experimental results show that the navigation system of the INS aided by neural network and monocular vision can effectively improve the error divergence of the inertial sensor in the GNSS denied environment.