Turning drone video into orthomosaic maps

What an orthomosaic actually is

An ordinary aerial photo is not a map. The camera sees the world in perspective, so tall objects lean away from the center of the frame and hills appear larger than valleys. Surveyors call this relief displacement, and it grows the farther a feature sits from the point directly below the lens.

Orthorectification removes it. Using the camera's geometry, its position and attitude for each shot, and a model of the ground surface, every pixel is moved to where it would sit if viewed from straight above. The result has uniform scale, so distances and areas can be measured directly. Stitch many of those corrected images into one seamless picture and you have an orthomosaic.

Why video is a workable input

Photogrammetry software such as OpenDroneMap is built around still images, but a video is just a dense stream of them. Pull frames at a steady interval and you get the same raw material a mapping flight would capture, often with more overlap than you need.

OpenDroneMap can extract frames from video itself, capped at 500 frames by default through its video-limit setting. Doing the extraction yourself gives more control: you choose the interval, drop takeoff, landing and transit footage, and keep only frames from the part of the flight that actually covers the site. That pre-processing is exactly the job of tools like DJI Frame Prep in the Forge.

Overlap, angle and sharpness

The software finds the same features in neighbouring frames and triangulates from them, so overlap is the single most important number. The OpenDroneMap flight guidance recommends roughly 70 to 80 percent overlap for orthophotos, rising to 80 to 83 percent over complex buildings and vegetation. With video, you set overlap after the fact by picking the frame interval to suit your speed and altitude; too sparse and the reconstruction breaks into pieces, far beyond the target and you mostly add processing time.

Camera angle matters too. For 2D maps the same guidance suggests pointing the camera nearly straight down, about 5 to 10 degrees off nadir. Motion blur is the other quiet killer: a smeared frame has few matchable features, so a steady, moderate flight speed pays off more than extra footage.

Putting the frames on the map

Still photos from a drone usually carry GPS in their EXIF data. Frames pulled from video do not, but DJI aircraft can record a subtitle (.SRT) sidecar with position for each moment of the clip. Matching each extracted frame to its SRT position lets you write an image geolocation file, geo.txt, which OpenDroneMap reads automatically from the project folder. Its first line names the coordinate system; each following line gives an image filename and its X, Y and optional Z.

The drone's own GPS puts the map approximately in the right place, which is enough for many uses. When you need survey-grade placement, add ground control points: targets or sharp features whose positions are measured on the ground with precise GPS. OpenDroneMap's documentation suggests at least five for most jobs and eight to ten for larger ones, spread evenly across the site, including corners and centre, and each visible in several images.

A practical sequence

Fly a steady grid or lawnmower pattern with the camera near nadir. Extract frames at an interval that gives about 75 percent overlap, discard frames outside the survey area, and generate geo.txt from the SRT telemetry. Load the frames and geo.txt into WebODM or OpenDroneMap, start with default settings, and only then raise orthophoto resolution for a final run. Check the result against a known distance on the ground before trusting any measurement taken from it.