Research On Visual Odometry With IMU Fusion

Visual odometry is one of the important technologies for mobile robot localization and navigation.Monocular visual odometry can adapt to environments with different scales,however,it is a challenge for monocular odometry to restore the absolute scale,and the accuracy is easily affected by the environment.Therefore,this paper presents a monocular odometry framework that fuses IMU with semi-direct method,which can solve the problem about scale uncertainty,and improve the localization accuracy and the running speed of the odometry.The sparse direct method is chosen as the front-end pose tracking process in this paper,minimizing the photometric error by means of direct method.It enables the algorithm to avoids the calculation of features and descriptors,which can speed up the algorithm compared with the feature-based method,and is beneficial for the real-time performance as well.However,due to the strongly non-convexity of the image,the solving process is easy to fall into the minimum which results to the failure when the motion is too large,as the direct method relies entirely on the gradient search in the pose solving.This paper associates the IMU data with the tracking process closely.The IMU is used to provide accurate short-term motion constraints and a good initial gradient direction for the direct method and to correct tracking failures,which improves the accuracy and robustness of visual odometry.Initialization is required before visual odometry to ensure that the correct initial value and high-quality 3D points are restored during the initialization process.For the visual-inertial odometry,the initialization of the bias of IMU and the gravity is needed,as well as the calibration of the transformation between the IMU and camera coordinate systems,which is used to avoid the cumulative errors.In this paper the visual-inertial coupled initialization is adopted,which utilizes IMU to restore the scale and takes the image information as constraints to initialize the bias and the gravity direction.The transformation between the two sensor is calibrated at the same time,while the scale and the poses acquired by IMU pre-integration are updated based on the gravity.In that case,the initialization process can provide a more accurate initialization result.The data fusion between vision and IMU is implemented by the tightly-coupled sliding window optimization algorithm.The poses,velocities,bias and scale are chosen as the optimization states when minimize the re-projection error,the IMU pre-integration measurement error and the prior error to acquire the fused pose estimation result.Marginalization algorithm can eliminate the historical state in the sliding window and update the prior estimate,and the states inserted into the windows are chosen by motion.Dynamic marginalization used in this paper divides the prior into three parts based on the current scale estimation and chooses the historical scale information retained in the prior dynamically.This ensures that the scales in prior are consistent with those in window,and decrease the scale error decrease as well.Experiments demonstrate that the algorithm which fuses direct method visual odometry and IMU information proposed in this paper can improve the localization accuracy and the tracking speed of monocular visual odometry,and show a reduction of the scale error while solving the scale uncertainty problem.Finally,the research work of this paper is summarized and the problems remain to be solved are analysed.