Photogrammetry and Structure from Motion Concepts

GIS 584: UAS Mapping & Analytics

Justyna Jeziorska, Helena Mitasova & Corey White

Photogrammetry & Structure from Motion (SfM) Concepts
GIS 584 · UAS Mapping & Analytics

Welcome everyone. Today we’ll connect the dots between images you capture with a drone and the metrics you can trust, orthophotos, point clouds, and elevation models.
By the end, you’ll understand why raw photos can’t be measured directly, what orthorectification does, and how SfM reconstructs 3D from overlapping 2D images. We’ll keep this practical and tie each idea to UAS field work and data QA.

Objectives

• Understand how remote sensing & photogrammetry support geospatial data acquisition
• Recognize aerial photo types and the measurement challenges they introduce
• Explain why image processing (orthorectification) is needed for measurements
• Grasp SfM concepts for creating 3D models from 2D images

These are the takeaways. We’ll move from fundamentals to workflow design.
Keep thinking about two questions: What do you want to measure? and What accuracy do you need?
Those answers determine flight planning, ground control, and processing parameters.

What is Remote Sensing?

• Sensing without contact: acquire data from a distance
  
• Passive sensors: measure reflected/emitted energy (RGB, multispectral, thermal)
  
• Active sensors: emit energy and measure returns (LiDAR, radar, sonar)

Remote sensing gives us information without touching the surface.
Drones typically carry passive RGB or multispectral cameras; LiDAR UAS are active sensors and directly measure range.
Keep this distinction in mind, SfM uses passive imagery but infers depth from parallax.

Photogrammetry (Why it matters)

• 3D coordinate measurement using photographs
• Two broad modes: aerial vs terrestrial (close-range)
• Historically crucial for mapping; now democratized via UAS

Photogrammetry is the science of turning photos into measurements.
With UAS, we’ve moved from expensive manned flights to frequent, targeted surveys, faster iteration, but also new pitfalls like rolling shutter and lens distortions we must correct.

Aerial Image Types

• Vertical: camera nadir-pointing; lower perspective distortion; better for mapping
• Oblique: camera tilted; richer façades/context; higher distortion; good for 3D

Vertical images are standard for map accuracy and uniform scale.
Obliques capture façades and undercuts but bring strong perspective effects, useful for 3D context and texture, less for planimetric measurements unless carefully modeled.

Why Raw Photos Mislead

• Perspective projection (not orthographic)
• Relief displacement: elevated features “lean” away from nadir
• Lens and platform distortions (rolling shutter, radial/tangential lens errors)

If you try to measure distances on a raw photo, elevation and tilt cause scale to vary across the image.
Orthorectification corrects these effects so you can trust the geometry.

Orthorectification (What it does)

• Removes perspective & relief distortion
  
• Compensates optical distortions (camera model)
  
• Produces a uniform-scale orthophoto suitable for measurement

Think of orthorectification as “unwarping” the image onto a known surface using camera parameters and elevation.
The output is metrically reliable, if the inputs and models are correct.

What We Need (Inputs)

1. Digital imagery with sufficient overlap (≥70% forward, ≥60% side as a baseline)
2. DEM/DTM to remove relief distortion (or derive via SfM)
3. Exterior orientation (IMU/PPK/RTK or estimated in bundle adjustment)
4. Camera calibration (intrinsics: focal length, principal point, distortion)
5. Ground Control Points (GCPs) for georeferencing & accuracy assessment
6. Software implementing collinearity equations / bundle adjustment (SfM)

Overlap enables feature matching; more texture yields stronger geometry.
IMU/RTK help, but even with them, you’ll want checkpoints to validate accuracy.
Bundle adjustment simultaneously refines camera poses and 3D feature locations.

Geotagging (Limits & Use)

• Embeds camera position in EXIF (approximate location/altitude)
• Useful for initialization, not precision georeferencing
• Expect meter-level errors without RTK/PPK or GCPs

Geotags get you in the right neighborhood.
For survey-grade outputs, you still need either high-quality GNSS (RTK/PPK) or well-placed GCPs and checkpoints to quantify accuracy.

Ground Control Points (GCPs)

• Marked or photo-identifiable, stable, and well-distributed
• Known X, Y, Z with survey-grade accuracy
• Essential to improve precision and to validate results (use checkpoints)

Separate GCPs (to constrain the solution) from checkpoints (to evaluate it).
Distribute both across the project area, edges and varying elevations, to avoid biased solutions and inflated accuracy claims.

From 2D Images to 3D (SfM)

• Automated feature matching across overlapping images
• Bundle adjustment estimates camera intrinsics/extrinsics + sparse 3D
• Densification (stereo matching) → point cloud → mesh/DSM/orthophoto

SfM recovers relative 3D structure from parallax; scale and orientation come from geotags, RTK/PPK, or GCPs.
Quality hinges on image texture, overlap, camera model, and outlier handling.

Multiple-View Geometry (Intuition)

• Correspondence: find the same feature in different views
• Triangulation: intersect rays from multiple cameras → 3D point
• Camera geometry: solve for poses that best explain all matches

Every matched pixel defines a ray in 3D; with two or more rays we triangulate a point.
The optimization balances all matches, minimizing reprojection error while refining camera parameters, this is the heart of SfM accuracy.

UAS Photogrammetric Workflow

1. Mission planning (GSD target, overlap, altitude, speed, lighting, safety)
2. Field ops (calibrated camera settings, consistent exposure, RTK/PPK optional)
3. Control survey (GCPs + checkpoints; stable, visible, distributed)
4. Processing (feature match → bundle adjust → densify)
5. Products (orthophoto, DSM/DTM, contours) + QA/QC (residuals, RMSE/NMAD)

Treat QA/QC as a product: report accuracy on independent checkpoints.
Inspect residuals, look for systematic tilt or doming, and check for artifacts on low-texture surfaces (water, uniform crops, snow).

Common Pitfalls & Fixes

• Low texture / repetitive patterns → add obliques; fly lower; diversify look angles
• Rolling shutter → shorter exposure, slower speed, global shutter if possible
• Doming / bowl surfaces → stronger control (GCPs/RTK), better geometry
• Vegetation canopy bias → model as DSM; don’t compare to DTM
• Shadowing/lighting changes → plan time of day; manual exposure

Most failures are geometric or radiometric: improve image geometry or stabilize exposure.
Always separate DSM vs DTM use-cases, don’t mistake canopy for ground.

What We’ve Learned (Recap)

• Remote sensing basics; passive vs. active sensors
• Why orthorectification is required for measurement
• Inputs needed for reliable mapping (overlap, calibration, control)
• How SfM recovers 3D and when it succeeds or fails

You should now be able to plan a flight for a target GSD, choose control, and defend the accuracy of your products with checkpoints and transparent QA.

Quick Activity (In-Class)

• Define a site, target GSD, and map scale
• Pick overlap, altitude, and speed
• Decide GCP count & placement + independent checkpoints
• List expected products and accuracy targets (e.g., RMSE_XY/Z)

Work in teams for five minutes and sketch a plan.
We’ll compare choices and discuss trade-offs, this mirrors real pre-flight planning.

References & Credits

• Adapted from course slides: UAS Mapping and Analytics – Photogrammetry and Structure from Motion Concepts (NC State CGA)