Research Of The Visual-inertial Odometry Based On Point,line And Plane Under Dynamic Environment

Visual-inertial odometry is widely used in service robots,UAVs and underwater autonomous vehicles because of its advantages of low cost and convenient deployment.Although it has been widely studied by scholars,the relevant algorithms need to be further improved in harsh environments such as no texture,fast illumination change,Manhattan world hypothesis and dynamic environments.Aiming at the above problems,this paper improves the system,including initialization problem,geometric relationship between planes and line feature robustness in dynamic environment.Aiming at the problem of visual inertia initialization,this paper introduces the line feature into the structure from motion to construct the point-line structure.Then,the loose-coupled method is used to solve the gyro zero bias,pose,scale factor and gravity acceleration in turn,and the key frame selection mechanism and the re-projection error model of line feature are improved.The experimental results show that the proposed algorithm can reduce the initialization error and improve the positioning accuracy of the system without affecting the initialization speed.Aiming at the problem of plane feature geometric relationship,this paper proposes an algorithm based on key plane to impose horizontal and vertical constraints between planes in the nonlinear optimization function.This method filters the marginalized features,defines the key plane,and applies the geometric residual between the planes in the factor graph.In addition,the building normal vector is customized,which is used to detect the horizontal plane and vertical plane,and the building normal vector is updated with the extraction of the key plane.The experimental results show that the proposed algorithm can increase the positioning accuracy of the system and make the plane map more in line with the expectation on the premise of maintaining the realtime performance of the system.Aiming at the robustness of line features in dynamic environment,a dynamic line feature screening method based on self-motion constraint,optical flow method,semantic segmentation method and re-projection error method is proposed in this paper.The above four methods originally applied to point features are extended to line features to accurately screen line features that may be located on dynamic objects.Then,in the optimization function,the dynamic feature is given a smaller weight than the static feature.Experimental results show that the proposed algorithm can accurately judge whether the extracted line features are located on the dynamic object,and can effectively reduce the impact of the dynamic object on the whole navigation system.