Research On Multi-state And Multi-view Constraint Algorithms For Vision/inertial Integrated Navigation

With the development of automatic driving,robot and other fields,the demand for autonomous navigation technology is more and more urgent.As two typical technologies of autonomous navigation,visual navigation and inertial navigation are complementary in performance.The inertial navigation system has the characteristics of high update rate and good dynamic performance,which can overcome the disadvantage of the poor robustness of the visual navigation in the more severe environment.But the pure inertial navigation system also has the characteristics of error accumulation over time,and the problem is more serious when the low cost inertial measurement unit is adopted.Based on the observation of external environment,visual navigation information can reduce the error accumulation of inertial navigation to a certain extent in the positioning and navigation scene of relative environment.At the same time,the aid of inertial measurement unit makes the scale factor of monocular vision become significant,which can overcome the problems caused by the lack of depth information in visual navigation.The research of vision/inertial integrated navigation algorithm has become a hot direction of autonomous navigation.In the algorithms for vision/inertial integrated navigation,a large class are based on Kalman filter.The comprehensive balancing on complexity of the algorithm,real-time performance and the adaptability of scene expansion are needed.In this paper,the typical multi state constraint Kalman filter(MSCKF)algorithm and multi-view geometry constraint Kalman filter algorithm are compared and studied.The innovative work of this paper is as follows:?The corresponding system models for different precision of inertial navigation system are given.The observation models of MSCKF algorithm and multi-view geometry constraint algorithm are derived in detail.Through the analysis of the system model based on the actual IMU and scene,it provides the basis for the actual problem of the algorithm and the improvement of the algorithm.?The applicability of the algorithm is obtained by the detailed comparative analysis of the two algorithms.MSCKF is suitable for high precision and low real-time scene,and multi-view geometry constraint algorithm has an advantage in resource constrained platform,thanks to its complete following of visual odometry concept--only to estimate its motion without interaction with the environment,reducing the complexity of the algorithm.In contrast,we also found the defects that multiple-view geometry constraint algorithm diverges in the state of low speed.That problem was verified by the open source data set with high precision and low precision data with Android platform of sensor,and results and theoretical analysis coincide.In view of the problem that the multi-view constraint model does not apply to the scene of low speed,it is modified,and the new constraint model has certain effect in low speed by experimental verification.?The influence of different accuracy of IMU on visual/inertial navigation algorithm is analyzed.The results of the analysis can be used as a reference for the use of the algorithm in different navigation needs and different precision of the sensor platforms.For example,in the mobile phone platform the corresponding algorithm for noise processing is needed.