Pix4DMapper was used to process the UAS photographs into an orthomosaic image and a digital surface model for use in GIS software. The University acquired a research license to work with the imagery. A few parameters had to be collected and converted to match the software’s requirements. Once included the software finished the process, creating a seamless high quality image of the island.
Before running in Pix4D a customized visual basic tool processed the GPX files from the quadcopter, linked the photographs by time stamp, and output the results into text CSV files. The output column values were then formatted for use by Pix4D Mapper with the following data fields: image filename, latitude, longitude, altitude, pitch, roll, horizontal accuracy, and vertical accuracy.
In this project the pitch and roll values were set to zero for processing because the camera was gimbaled, maintaining a relative nadir value to offset pitch and roll by the quadcopter. Even without a gimbal the Pix4DMapper adjust for minor differences off nadir. Our output report indicates the majority of images were within 3.0 and 3.36 degrees of vertical (Omega and Phi).
Horizontal position accuracy was assumed to match the accuracy of the on board GPS unit of about 3 to 5 meters. Vertical accuracy would normally be very high because we used the barometric parameters not the GPS elevation, however this had to modified to adjust for the starting elevations. Unknown to us at the time our UAS set and recorded altitudes relative to takeoff location, which in some cases were 10 meters higher than others. This meant while our waypoint flight altitude was set to a consistent 40 meters for each flight those launched from higher locations were affectively up to 50 meters. This was visibly revealed by initial runs of the Pix4D software by showing the predicted altitude against the computed altitude from the software.
For the final run of the software all flight altitudes were adjusted by adding LIDAR derived elevation values for the take-off sites. With these adjustments the altitude still varied between flights, but were now linked to a common base. After correcting for the LIDAR derived ground elevation the Pix4D output showed a close tie between predicted and final altitude values.
The left hand image below shows how Pix4D corrected our original altitudes (blue spheres), automatically adjusting the values to match what the software calculated from the photo overlap (green spheres). The right hand image shows how using adjusted takeoff altitudes provided a better match to the final values determined by the software.
Although Pix4D does its own corrections for our mistakes, when planning future flights the take-off locations altitude will be added or subtracted from the desired waypoint altitude before flying to create a consistent flight altitude above the surface. This should improve the results for a given set of flights and assure proper ground coverage.
Heading was maintained to a single direction throughout each flight within a range of error well handled by Pix4DMapper. The goal was a heading that aligned the images to True North. When converting images the quality of the pixel resolution is easier to maintain if less rotation is required during processing and this orientation makes it easier for the user to quality check the photographs. The April flight was set approximately 10 to 14 degrees off magnetic North because it was thought an adjustment was required to account for declination. It turned out the Mikrokopter waypoint manager already takes declination into account, so the second May flight was set to a heading of 360 degrees. Pix4DMapper adjusted the images where necessary, showing the readings were within 8.25 degrees of intended direction (Kappa).
After the flight the photographs were downloaded by USB to a PC computer. The photographs were organized into two folders, one for each flight day. Each day has specific to a time correction on the camera to account for time drift in the camera settings. Zero (0) seconds on April 26th, minus two (-2) seconds for May 5th. To get this time correction an on screen photograph was taken of the official NIST US time from http://www.time.gov. Taken the same day as the flight, the adjustment is the time difference between the JPEG file EXIF time and what is visible as captured on screen.
Over 500 photographs were taken. Each photograph was reviewed in an image viewer. Some of the images were deliberately taken off nadir after waypoint flights to capture other views of the island, these and other images not usable for mapping were removed from the mapping project. Images at extreme orientations and those well below the flight altitude were also removed. Images out of focus or blurred by motion were removed, however, if there was insufficient overlap by neighboring images to maintain 3-D vectors at a particular location the photograph was retained. A final count of 445 images were selected for the mapping portion.
The image file names were run through a Visual Basic program to link the photographs to GPX data from the quadcopter using the corrected time stamps. Using the output CSV file and a link to the photographs, the Pix4Dmapper software then evaluated the images and determined 439 images had enough coverage of inland features to allow for calibration and geolocation.
Ground control points (GCPs) can help Pix4D adjust the final output results to a true location. If a CGP includes elevation data the software can use the values to lock the approximate altitude to a proper frame of reference, adjusting for errors in the camera parameters and UAS altitudes. To do this the software requires at least three GCPs for a 3-D lock. Using online web services eleven (11) NGS survey points were identified in the records as markers on the island. Of those, three were identified as lost in the survey report descriptions. Of the eight markers describe as valid five were visible in the aerial images, providing a good distribution for control across the island.
No elevation data was included with the NGS ground control descriptions other than relative height of the objects the markers were place upon. For example in one location a cement pillar has a clear survey marker embedded in its surface, the report description only provided the height of the pillar relative to the ground, in this case four feet high. To add elevation records to these points bare earth LIDAR values were averaged surrounding the marker locations, then the feature elevations from the reports were added to the result. This technique provided a relatively accurate 3D ground control for the software. An accuracy range was assigned within the software based on assumptions made for each point.
After the software was run using available control points the digital surface model showed the West end of the island “dipped” below the East end by approximately 2 meters. In an attempt to help the program correct for true elevation at these locations an artificial control point was created on a western most feature, the goal was to bring the West end “up” to a level plane. Because there were no controls on the far West end of the island we had to create a GCP for Pix4D to use. From the first run of the Pix4D software a clearly visible portion of a raised armament was selected for use.
While the height of this feature relative to real world coordinates was unknown, the X, Y coordinate and the height of the feature relative to the surrounding ground was considered accurate within centimeters. With this assumption these coordinates and relative height difference were taken from the software and the height adjusted to real world values by adding an averaged value from the surrounding LIDAR pixels. The model was then run again with this adjusted point included. The result brought the island height variation within half a meter from West to East, better matching the water levels surrounding the island.
The output may not be perfect, but even without true surveyed control points the results proved very effective, more so when you consider the project data was not originally collected with Pix4D in mind. In the future 3D survey points could be collected across visible features in the dataset, if the software was then run again using these points the quality of the output could be improved.
Exported snapshots of the final orthomosaic:
Return to main page