Research On Monocular Visual-inertial SLAM Based On Nonlinear Optimization

Mobile robot has become the focus of current research.Recent augmented reality,virtual reality and unmanned vehicle promote development of the technology.Simultaneous localization and mapping(SLAM)is core of it.However,the SLAM of single sensor can't meet the demand of some applications either in accuracy or robustness.The paper presents real-time monocular visual-inertial system based on features method.The system can be applied to indoor,outdoor,and large scale environments.The system don't require any prior information to initialization.it consists of three section,including initialization,tracking,local map and loop-closure.The paper adopts a loosely-coupled sensor fusion method to get initial parameters,which including transformationfrom body frame to world frame,gravitythe real scale,velocity of IMU,accelerometer biases and gyroscope biases.The paper propose two ways of tracking.If the initialization is successful,we use visual inertial to tracking,otherwise only visual tracking is used.In visual inertial tracking,we also adopt two different cost function according to whether the map is updated.Our system constructs the local map to reduce the cumulative error and enhance robustness.Wealso use marginalization and sparsificationto ensure that information is not lost and computationally bounded.We fixed the linearization point to compute Jacobian matrix in marginalization,which avoid misinformation.Finally,we also take advantage of loop-closure to eliminate long time cumulative error.We advocate the 4-DOF pose graph optimization in loop-closure owe to the pitch and roll are observable in system.The system is tested in a recent public EuRoC dataset,and it get very high precision of initialization.It achieves error of less than 5% in scale factor,the gravity magnitude converge to about 9.8 m/s2,accelerometer biases converge to about 0 and gyroscope biases converge to about310 /-?t.Our algorithm also is competitive in positioning accuracy.It achievesaverage error less than 8 cm and the biggest error less than 10 cm.The result is the state-of-the-art in monocular visual-inertial odometry.In the physical test,our algorithm is also excellent.In the small space range,the accuracy of location is about 10 cm.