Research On The Visual Aided GNSS/SINS Integration For Vehicle Precision Positioning

High-precision position and attitude is a fundamental module for mobile measurement and intelligent carriers.In recent years,with the prosperity of autonomous drivind,smart robot,unmanned carrier,etc,the requirement for high availability,high reliability,high accuracy location services is expanding gradually.Especially in the complex environment of cities,the problem of positioning has become a bottleneck which everyone pays close attention to.The integration of multi-frequency global navigation satellite system and high performance inertial measurement units(IMUs)is still the most reliable strategy in open-sky conditions,but the positioning performance will be degraded severely when the satellite signals suffer from frequent blockages.Camera,as a low-cost,passive and portable sensor,can provide addition information to assist the positioning of GNSS/INS.GNSS can establish a unified global space and time reference,providing all-weather location services;SINS is independent,high sampling rete,stable and continuous;camera not only can perform relative positioning,but also achieve absolute positioning supported by prior map.Its performance is between GNSS/SINS,which is an effective complement to the GNSS/SINS combination system.GNSS,SINS and vision sensors are also three major sensors deployed in driverless vehicles.It can be seen that Vision/GNSS/SINS has strong complementary advantages and multi-source fusion positioning is an important direction of PNT in the future.As a result,this thesis focuses on the Visual aided GNSS/SINS integration for vehicle precision positioning and aiming to improve the positioning ability of vehicles in the complex urban environment.The main works ara as follow:This paper makes full use of the complementarity of GNSS,inertia and vision,and proposes a tight coupling positioning method of GNSS/inertial/visual based on extended Kalman filter with multi-state constraints.The state model and observation model of the scheme are derived in detail.The vehicle-borne measured data is used to evaluate the performance of the positioning method.The results indicate that in the case of GNSS signals lost for 3 minutes,the visual/inertial tight coupling positioning can greatly suppress the IMU’s error drift,providing high-precision speed and attitude.The maximum position error of the position is less than 2 m.And the error accumulates with distance is less than 1‰.When the satellite is partially occluded for a long time,the tight combination method can obtain positioning accuracy of decimater to sub-meter level.This paper implements and improves three front-end algorithms(SIFT/ORB/LK).The algorithms are tested by measured data.The results show that the SIFT feature method takes an average time of 0.258 s,which is about twice that of ORB and LK.Its feature points are distributed unevenly,but the tracking effect is stable.The tracking effect of LK is greatly affected by the lighting.And as for test data,the overall positioning accuracy of LK front-end method is slightly higher than SIFT and ORB.Based on YOLOv3,the dynamic obkect detection remove the dynamic features on pedestrians and vehicles as outliers and only leaving static feature points for visual odometry.Tests show that the accuracy of dynamic object detection can reach more that 90%,and the average detection time is only 0.05 s.After removind the dynamic interference,the lateral and forward positioning accracy of the visual odometer is improved by about 0.5m.The high-precision pose information obtained by post-processing of the POS system is used to obtain the centimeter-level visual pointclouds map through left and right image matching,inter-frame feature tracking and optimization.And the points are screened using multiple typed of quality evaluation indicators.Meanwhile,the lane map is obtained through traditional lane extraction and Kalman filtering tracking algorithm.After achieving the prior map,the fusion positioning method based on visual matching/lane constraints are proposed.The experiments prove that with the aid of prior map,the applicability of the positioning system in complex environments can be greatly improved.When the GNSS signals are completely interrupted,assisted with visual point cloud map,the position accuracy can be maintained in 10 cm even if only one landmark point is matched;When the GNSS signals are severely blocked for 300 s,the lateral constraint of the lane line and the forward constraint of the odometer/ road-direction sataellite can make the algorithm maintain the position accuracy of about 20 cm.