Visual-inertial Odometer For UAV Combined With Active Exposure Control

The premise for an Unmanned Aerial Vehicle（UAV）to perform tasks accurately is to obtain its own position,which is usually called UAV self-localization technology.According to the type of sensors used,UAV self-localization can be divided into Passive Localization requiring interaction with preset equipment and Active Localization using sensors such as lidar,inertial measurement unit and camera,etc.In a GPS-denied environment,the active localization contributed more robustness for our requirement.In active localization,the bias of IMU can be estimated and the trajectory drift can be corrected by using the pose calculated by visual information;When the image is blurred or features are lost,the pose estimation of IMU can ensure the continuity of UAV localization in a short term.At the same time,the lost scale of monocular camera can be recover based IMU data.Therefore,VIO is an UAV self-localization scheme with cost performance,accuracy and robustness.VIO can be extended from odometry to SLAM system,which means simultaneously estimate its pose and landmark in surrounding environment structure.In the process,different levels of maps can be established for pose optimization and subsequent navigation.Generally,the sensor based pose estimation is called the front-end of SLAM,and the further optimization combined map and pose is named the back-end.The pose error,meanwhile,gradually converged.A complete SLAM system should have features like global pose optimization and loop closure in order to achieve required accuracy.In the front-end part,an active exposure control algorithm based on predictive model is proposed,which uses the photometric response function of the camera to predict the best exposure.At the same time,a comprehensive evaluation index of image gradient information,image entropy and noise impact is proposed.A concise way is used to balance the exposure parameters,improve the image quality and speed up the iteration speed.Subsequently,based on the front-end improvement,a tightly coupled optimization scheme of visual-inertial combination was adopted,a SLAM back-end with loopback detection was added,and the IMU pre-integration technology was used to unify the marginalized prior,IMU,and camera residuals in the Jacobian matrix,and a complete initialization process is proposed.Finally,the improved semi-direct SLAM scheme is systematically benchmarked on the Eu Roc dataset and experiments are carried out on a UAV equipped with Real Sense T265 to evaluate the speed,robustness and accuracy of the algorithm,as well as its performance in embedded suitability in the system.