What Does a Frame-to-Frame Traffic Detector Compare?

Frame the investigation by showing that a traffic camera only delivers a grid of pixel intensities, and raise the question of what the detector actually compares when it labels something a 'vehicle'.

• A traffic camera outputs a matrix of numbers, not objects
• 'Comparison-based' detectors must turn numbers into a decision
• Question: what is being compared frame to frame?

Ask the learner to commit to a first guess about the comparison operation before any evidence is shown.

• Commit to an initial hypothesis
• Distinguish raw-frame subtraction from smarter references

Let the learner toggle between three real comparison strategies applied to the same short intersection clip: raw frame difference, background subtraction, and adjacent-frame feature comparison, and see the resulting change maps side by side.

• Raw frame difference highlights every moved pixel, including shadows
• Background subtraction ignores static scenery and only flags non-background pixels
• Feature-based comparison tracks corner-like points across frames
• Each strategy produces a visibly different change map

Show the second stage: thresholding, morphological cleanup, connected-component labeling, and bounding-box extraction that turn a change map into vehicle hypotheses.

• Difference values are thresholded into a binary change mask
• Morphological operations remove noise and fill holes
• Connected components are extracted as blob candidates
• Bounding boxes are drawn around each blob

Explain why background models, previous frames, and local features each serve a different purpose: static rejection, fast motion, and occlusion robustness.

• Background models separate moving foreground from static road
• Previous-frame subtraction reacts within a single frame delay
• Feature descriptors remain trackable through partial occlusion
• The detector picks the comparison that matches the road scene

Let the learner apply the same comparison pipeline to a congested daytime highway and a dark, rainy intersection with strong headlights, and observe where each comparison strategy breaks.

• Headlights confuse raw frame differencing with phantom blobs
• Background models adapt slowly when rain changes the road globally
• Feature comparison keeps tracking through glare
• Robust traffic detection uses comparison plus adaptive thresholds

Clarify what these detectors deliberately do not handle: classifying vehicle type, reading license plates, separating two occluded cars, or recovering from long full occlusion.

• They detect change, not semantic class
• Two overlapping vehicles usually merge into one blob
• Stationary vehicles blend into the background and vanish
• Classification and tracking require extra modules beyond comparison

Directly answer the driving question: a comparison-based traffic detector compares a compact per-pixel summary of the current frame against a reference derived from earlier frames, and uses the difference to hypothesize vehicles.

• The reference is usually a background model, a prior frame, or local features
• The comparison produces a per-pixel change signal
• Change signals are thresholded and grouped into vehicle candidates
• That is the full 'frame-to-frame comparison' pipeline