Research On Intelligent Vehicle Location Technology Based On Vision Inertial Navigation Fusion

Navigation and positioning technology is an important technical basis for unmanned vehicles,and SLAM technology based on multi-sensor fusion is one of the mainstream technologies to achieve high-precision positioning of intelligent vehicles.Due to the complexity of intelligent vehicle driving environment,even the vision inertial navigation SLAM with the best fusion positioning performance still has shortcomings in some driving environments,for example,weak texture environment will lead to poor positioning accuracy,and the environment with a large number of similar objects will lead to low accuracy of loop detection.In order to solve the above problems,this paper proposes the VINS localization algorithm based on point line feature adaptive fusion and the loop detection algorithm based on multi-source information,in order to improve the comprehensive positioning performance of the system.The key contents of this thesis are as follows:Aiming at the poor positioning accuracy of vision inertial navigation SLAM in the weak texture environment,and the real-time performance of the system will be reduced if the point and line features are fused directly,and the positioning accuracy is not significantly improved in the rich texture environment,this thesis proposes a VINS positioning method based on the adaptive fusion of point and line features.The grid method is introduced in the front end to evaluate the quality of the point features of the current environment,and the external environment is judged to be weak texture environment mainly from the point feature extraction proportion,point feature tracking proportion and the distribution uniformity of point features.If the texture environment is weak,the visual constraints are constructed by fusing the point and line features.If the texture environment is rich,the visual constraints are constructed only by using the point features.That is to say,in different texture environments,the back-end can construct the loss function and estimate the vehicle pose based on different constraints.For the environment with a large number of similar objects will reduce the accuracy of loop detection,this thesis proposes a loop detection method based on multi-source information.Firstly,the grid map is established based on the GPS positioning results of the current key frame,and the historical grid of the vehicle is recorded based on the GPS positioning results.Then,the GPS positioning results of the current key frame can be used to determine whether the vehicle is in the historical grid that has been reached,Finally,bag of words model is used to calculate the similarity between "current key frame" and "historical key frame in grid",so as to judge whether the loop is looped.Finally,the effectiveness of the proposed method is verified based on Kitti and Euroc data sets.The experimental results show that by fusing the point line features in the weak texture environment,the proposed VINS localization method can improve the positioning accuracy by about 6.89%,and avoid the introduction of line features in the rich texture environment,which can ensure the real-time positioning.Based on the multi-source information loop detection method,the loop detection accuracy is improved by 16.74%,and the recall rate is improved by20.09%.The results are in line with expectations.