Research On Visual-inertial Odometry Based On Point-line Features

With the rapid industrial development in the field of industrial robots and service robots,mobile robot positioning technology has attracted a lot of attention in a wide range of industrial applications,especially in the field of simultaneous localization and mapping research which is associated to Indoor Photogrammetry is the most hot.At present,most mainstream visual SLAM algorithms are based on feature points for data association and back-end optimization.However,in some scenes with low illumination and weak texture,the localization accuracy and robustness of SLAM system are reduced due to the insufficient number of feature extraction.In order to solve this problem,we introduced line features which is improved in the front-end data association stage of SLAM,and built a visual inertial odometer based on the fusion of point-line features to improve the positioning accuracy and robustness of the localization algorithm in complex scenes.The results show that the ORB feature is robust to the environment and the speed is the fastest through the experimental test of multiple feature extraction and matching algorithms.In the feature point matching mode,the matching speed based on LK optical flow is about eight times that of Brief descriptor.Therefore,ORB feature and LK optical flow matching methods were selected to extract and match feature points.LSD line features and LBD descriptors were selected for data correlation with feature points,and the segmentation problem of LSD line features extraction was improved.The comparison experiments before and after the improvement showed that the number of line feature matches could be increased by 3.8%.The method of Pluck coordinate and orthogonal representation is used to describe the coordinate transformation of line features and the optimization of subsequent reprojection error of point-line features.The initial value of camera pose transformation was obtained by integrating the IMU,which was applied to the back-end optimization.The IMU error was constructed by pre-integrating the IMU.In the back-end optimization,using a sliding window model and setting the key frame selection strategy,through the marginalized update camera pose and key frames constraint information preservation,and the constraints as a priori information and dotted line projection error and IMU measurement through the method of beam adjustment methods to optimize together,build a dotted line feature fusion based visual inertia odometer.In order to fully verifythe performance of the point-line VIO system proposed in this paper,we built an experimental platform in the real environment,calibrated the camera,andselected three aspects of system robustness,positioning accuracy and real-time performance,compared and analyzed the system performance through data sets and experiments in real environment scenes.In terms of robustness,line features have better stability in complex environments,.after the introduction of line features,point-line VIO system runs stably in complex scenarios and has good robustness.In terms of positioning accuracy,we compared the dataset and the real environment and found that the improve of line feature can reduce the positioning error of the system.In the dataset scene,the average trajectory error is 0.155 m,and the rotation error is reduced by 18.3% on average.The average reduction of translation error is 10.7%,which improves the positioning accuracy of the system.Through the comparative experiments in complex environment and non-complex environment,we found that in the non-complex environment,the translation error is reduced by 13.1%,and the RPE is reduced by 16.2%,and the more complex the surrounding environment is,the more obvious the positioning accuracy improvement effect is.In terms of real-time performance,point-Line VIO takes the longer time and has the lower real-time performance.