A target positioning method based on monocular depth estimation and scale recovery

Abstract

The application discloses a target positioning method based on monocular depth estimation and scale recovery, and belongs to the field of aerial image processing, and specifically comprises the following steps: taking an image sequence actually shot by a UAV as the basis, using existing commercial software to produce a depth estimation data set under the UAV perspective; obtaining a relative depth map of a scene through an unsupervised monocular depth estimation network; estimating the height of a camera based on scene segmentation and a two-step ground point optimization algorithm, and then comparing the height with a prior real camera height to obtain a scale factor of the scene; combining the scale factor and the relative depth to obtain the absolute depth of the scene; detecting common targets through a target detection algorithm; and finally positioning the targets using the principle of pinhole imaging. The application combines the absolute depth estimation method and the target detection method, realizes positioning of common targets in a low-altitude UAV shooting scene, and has high accuracy.

Images