Recovering Scale in Monocular DSO Using Multi-sensor Data

摘要

Monocular visual odometry like DSO (Direct Sparse Odomctry) or visual SLAM can be used for UAV navigation. But the scale of DSO is set more or less arbitrarily during its initialization, because the absolute scale is unob-servable by using a single-camera. Therefore, the map reconstructed by DSO and the pose estimated by DSO can't be used directly. In order to recover the scale, firstly, we use IMU, GPS and ultrasonic data and Kalman filter algorithm to calculate the position and pose of UAV. Particularly, we improve the Kalman filtering algorithm instead of using the original algorithm. After the calculation, we get the position and pose of UAV with real scale, and then estimate the proportion between DSO's scale and real scale by least square method. Finally, using this proportion to correct the scale of point cloud, we can get the point cloud map with real scale which can be used for UAV navigation in the future work.