🚦 How Computer Vision Learns to Recognise Road Signs

About this simulation

A forward-facing camera captures a road sign, and a compact convolutional neural network — the kind used in real driver-assistance systems — turns that image into a classification in milliseconds. This scene lays the network's stages out in space so you can see the image get smaller and its feature maps get deeper, exactly like a MobileNetV2-style backbone.

How it works

• Input frame — the captured sign, rendered with a canvas texture and optional grime/weather overlay.
• Depthwise blocks (blue) — one small filter per input channel, sliding over space cheaply.
• Pointwise blocks (amber) — 1×1 convolutions that mix those channels into new, richer features — together these form a depthwise-separable convolution, MobileNetV2's key efficiency trick.
• Global pool + classifier — the final feature map is averaged down and turned into class scores for the four sign types.
• Each block's spatial grid shrinks while its channel depth grows deeper into the network — trading resolution for abstraction.

Controls

• Sign on camera — which UK road sign the camera currently sees.
• Weather / light — fog, rain and darkness all reduce image quality and confidence.
• Dirt / occlusion — simulates a muddy, faded or partly blocked sign.
• Network width (α) — MobileNetV2's width multiplier: fewer channels run faster but recognise less reliably; more channels are slower but more accurate.
• Show data flow — toggles animated particles carrying the signal through the network.

Did you know?

UK-specific signage — retroreflective triangles, roundel speed limits, the diagonal-stripe "national speed limit" sign — differs enough from other countries' road signs that ADAS models are usually fine-tuned on regional datasets before deployment, rather than relying on a single global model.