Precise estimation of the geoposition and orientation of ground-level video cameras from multiple sensors.

摘要

This dissertation addresses the problem of precisely determining the geodetic position (geoposition) and orientation of multiple ground-level video cameras. Each video camera is calibrated and equipped with a Global Positioning System (GPS) receiver and compass-magnetometer. The GPS receiver measures the latitude, longitude, and height above mean sea level of the video camera and the orientation of the video camera is derived from data acquired by a compass-magnetometer, which measures the pitch, roll, and yaw of the camera. Additionally, features are tracked throughout the video acquired by the calibrated camera in order to measure the relative camera motion between successive video frames. Each of the measurements from this disparate set of sensors is first mapped such that they are all relative to a common Earth-centered, Earth-fixed Cartesian coordinate frame. The uncertainty of each measurement is also propagated through this mapping. The geoposition and orientation of each camera is independently estimated from all of its associated sensor measurements. The measurements and their associated uncertainties are input at different frequencies and are sometimes incomplete due to GPS dropouts, corrupt video frames, etc., yet the recursive estimation process uses these multiple measurements to reliably calculate the most probable geoposition and orientation with quantified uncertainty at each video frame.; Further, if multiple cameras are imaging the same region of a scene, cross-camera feature correspondences are established using a combination of guided matching and robust feature comparison. The resulting independent observations of corresponding features contained in the scene are used to jointly estimate the maximum likelihood of the geoposition and orientation of all cameras imaging the same region of a scene for which feature correspondences have been established. This yields decreased relative errors between the cameras, resulting in more precise estimates of the geoposition and orientation of the cameras. This approach scales well and allows the video cameras to be located anywhere in the proximity of the Earth.