Is the focus of expansion a real scene point?

It is an image location representing the projected translation direction. A visible object need not occupy that location.

What happens when the camera moves backward?

Under the corresponding ideal assumptions, the radial pattern contracts toward a focus of contraction.

Can the focus alone tell us absolute speed?

No. Translation direction and metric speed are different quantities; additional scale information is needed to separate speed from depth.

AI Glossary / Graphs and symbolic AI / Intermediate

Focus of expansion

Definition

The focus of expansion is the image location from which motion vectors appear to spread outward when a camera translates toward a stationary scene without rotating. It represents the projected direction of translation under the relevant camera model. In computer vision, it can help interpret an observer's motion from changes across images.

What it means

Imagine looking forward while moving along a straight corridor. Nearby features appear to move outward across your view, while the point in the direction of travel appears comparatively stationary. The focus of expansion describes the center of that idealized outward pattern.

This term belongs to image-motion geometry. It is not a search algorithm's next node to expand. Also distinguish the geometric motion field, produced by movement of projected scene points, from estimated optical flow, which tracks apparent image changes. Texture, illumination, occlusion, and estimation errors can make measured optical flow depart from the ideal geometric pattern.

How it works

Under a pinhole camera model, translation changes the image positions of stationary scene points. With no rotation and a nonzero forward translation component, their motion directions align with rays through a common image location. Extending those directions backward can help estimate the focus of expansion.

The location depends on translation direction and camera calibration. It need not lie at the image center or even inside the recorded image. Translation exactly parallel to the image plane produces a limiting case with the focus at infinity and parallel motion directions.

Camera rotation adds a different motion component. Independently moving objects can introduce additional patterns. Practical estimation therefore needs appropriate motion assumptions, robust handling of inconsistent tracks, or a method that separates rotational and translational effects. A single intersection drawn from two noisy arrows is rarely sufficient evidence.

Three practical examples

1. A mobile-robot developer examines image tracks from a straight motion segment to compare visual heading with the robot's measured movement direction.

2. A driving-simulation researcher studies how background image motion changes when a virtual camera travels forward or diagonally. The known simulated motion provides a reference for evaluation.

3. A video analyst checks whether a simple outward-flow model plausibly describes a shot. A turning camera or a large independently moving object may require a richer motion model.

Think of it this way

Think of watching scenery through the front of a smoothly moving train. Objects appear to fan outward from the direction ahead, with nearby scenery often sweeping across the view faster.

The analogy illustrates perspective motion, but a real train can turn, vibrate, and pass moving vehicles. Those effects break the pure-translation picture. The focus is a geometric image location, not a physical object placed on the track, and it does not by itself reveal the train's speed.

Worked example

Consider a calibrated pinhole camera moving straight forward at 2 meters per second toward stationary scene points. Assume no rotation. Use normalized horizontal image coordinate x = X/Z, where X is sideways position and Z is forward depth. The focus of expansion is x = 0, y = 0.

For this motion, horizontal image velocity is u = x × 2/Z. A point at x = 0.2 and depth 10 meters therefore has u = 0.04 normalized coordinate units per second. A different point at the same horizontal coordinate but depth 20 meters has u = 0.02. They can have different vertical coordinates, avoiding identical image locations.

Both horizontal components point outward, but the farther point moves more slowly. Doubling both translation speed and all scene depths leaves these image velocities unchanged. Consequently, this observation alone cannot determine absolute speed and depth separately. The calculation is an idealized instantaneous example, not an optical-flow measurement from recorded video.

Illustrative case study

Fictional case: Bracken Robotics tests a camera on a cart moving through an empty demonstration corridor. The team plots tracked feature motions and estimates an outward-flow center during straight segments.

When the cart turns, the estimate becomes inconsistent. Rather than concluding that the corridor moved, the engineer separates the turning segments and considers rotation measurements in the analysis. A person crossing the view is also treated as a possible independently moving region. The team compares estimates with known test motion before using them in a navigation prototype. This scenario illustrates validation choices; it does not claim a measured navigation improvement.

Common misconceptions and mistakes

Do not assume the focus of expansion is always the principal point or the center pixel. Those coincide only under the appropriate motion and calibration conditions.

Do not interpret every expanding image pattern as camera translation. Zooming, object motion, or other effects can also create expansion-like observations. Check the image-formation assumptions.

Finally, the absence of a clear finite focus does not prove the camera is stationary. Sideways translation, rotation, weak texture, and noisy measurements can prevent the simple radial picture from being visible or well estimated.

Business use

For businesses evaluating visual navigation or video-analysis systems, the useful question is whether the tested motion model matches the operating environment. Ask which conditions the system handles: camera turns, moving crowds, vibration, low texture, and changing illumination are distinct challenges.

Record camera calibration, frame timing, tracking quality, and the reference motion used in validation. Report uncertainty and failure cases alongside a heading estimate. An apparently persuasive flow visualization should support an engineering assessment, rather than substitute for measured performance under representative conditions.

Check your understanding

Why does the worked example's farther point have half the horizontal image velocity?

Answer: At the same normalized horizontal position and translation speed, the expression is inversely proportional to depth. Doubling depth from 10 to 20 meters halves that component.

Related terms

Optical flow
Motion field
Perspective projection
Structure from motion
camera calibration
visual odometry
epipole

Learn next

Study perspective projection first, then compare image-motion patterns for pure translation and pure rotation. Practice identifying which conclusions concern direction and which require additional depth or scale measurements.

Perspective projection

Sources

Guido Gerig, Optical Flow II, pure translation and expansion/contraction: https://www.sci.utah.edu/~gerig/CS6320-S2015/Materials/CS6320-CV-S2015-OpticalFlow-II.pdf

Imperial College, Motion and Optical Flow: https://www.doc.ic.ac.uk/~dfg/vision/v16.html

The supplied category is retained; this article uses the computer-vision meaning. Calculations and the fictional case are original teaching examples.

Previous: Focus groups Next: Focus peaking