Automotive

Camera Sensors

Of the three main ADAS sensors, the camera is the only one that "reads" a scene the way a human driver does — recognizing a lane marking as a lane marking, a stop sign as a stop sign, a pedestrian as a pedestrian. Radar and LiDAR confidently say something is there; the camera is what usually says what.

Quick reference

What it is: An image sensor and lens feeding a computer-vision pipeline that detects and classifies lane markings, signs, pedestrians, and vehicles.
Mental model: The one ADAS sensor built to recognize what something is, not just that something is there.
Key numbers: Typically 1–8 MP at 30–60 fps; dynamic range (handling sun glare, headlights, tunnel entrances) matters more day-to-day than raw resolution.
Reach for it when: Classification matters — telling a pedestrian from a shopping cart, or reading a lane marking or a speed-limit sign — not just detecting that an object exists.

From photons to pixels

An automotive camera's image sensor is a grid of photodiodes, each one counting how many photons hit it during the exposure — the same basic photodiode principle behind semiconductor photodetection in general. A color filter over each photodiode (almost always a Bayer pattern of red/green/blue filters) lets the sensor infer color, and an ISP (Image Signal Processor) reconstructs a full-color image from that raw, per-photodiode data before any computer-vision algorithm ever sees it.

Mono vs. stereo: getting depth from a 2D image

A single camera captures a flat, 2D projection — depth isn't directly measured, only inferred (from known object sizes, the ground plane, or a trained model's learned priors). A stereo pair — two cameras a known distance apart — recovers real depth by triangulation, the same baseline-and-angle idea behind radar's angle of arrival, just applied to light instead of radio:

The object lands at a slightly different horizontal pixel position in each camera's image — that pixel difference (disparity) is inversely proportional to depth: nearby objects shift a lot between the two images, distant ones barely shift at all.

That inverse relationship is also stereo's core limitation: disparity shrinks fast with distance, so stereo depth accuracy is naturally best up close and degrades the farther out an object is — often the exact opposite of where radar's range accuracy shines.

The camera processing pipeline

Radar and LiDAR chains are mostly signal processing; the camera chain is mostly machine learning — a genuinely different compute profile even though all three feed the same ADAS ECU.

What cameras actually detect

Lane markings: The basis for Lane Departure Warning and Lane Keep Assist — detecting painted lane lines and the vehicle's position relative to them.
Traffic signs and signals: Traffic Sign Recognition (TSR) reads speed limits, stop signs, and similar signage, often displayed back to the driver.
Pedestrians and cyclists: Classification that radar and LiDAR alone struggle to make confidently — critical input to Automatic Emergency Braking.
Other vehicles: Detecting, classifying, and often estimating the pose/orientation of surrounding vehicles for adaptive cruise control and collision-avoidance logic.

Camera vs. radar vs. LiDAR

See the full comparison table on the Radar page rather than repeating it here — the short version is that camera is the classification specialist, at the cost of being the sensor most degraded by darkness, glare, fog, and rain.

Real-world applications

Lane Departure Warning / Lane Keep Assist — a forward camera watching lane markings is the primary sensor behind both features.
Traffic Sign Recognition — reading speed-limit and regulatory signs and surfacing them on the dashboard.
Automatic Emergency Braking — camera-based pedestrian/cyclist classification is frequently what triggers or confirms an AEB event alongside radar's range data.
Driver monitoring systems — an inward-facing camera watching the driver's eyes/head pose for drowsiness or distraction, increasingly required by regulation.
Surround-view / parking cameras — multiple wide-angle cameras stitched into a top-down view for low-speed maneuvering.

Gotchas

Field notes

Camera performance drops sharply in low light, direct sun glare, fog, heavy rain, or simply a dirty or iced-over lens — the sensor most exposed to real-world weather and cleanliness problems of the three.
The CNN inference stage is genuinely compute-heavy compared to radar or LiDAR's signal processing — automotive-grade camera systems typically need a dedicated AI accelerator, not a general-purpose CPU alone.
Monocular depth estimation is inherently an inference, not a direct measurement — it can be fooled by unusual object sizes or scenes the training data didn't cover, in a way stereo triangulation or radar/LiDAR range measurement can't be.
A stereo camera's calibration (the precise baseline and alignment between the two lenses) can drift from vibration or a minor impact over a vehicle's lifetime, silently degrading depth accuracy without an obvious failure symptom — periodic recalibration is a real maintenance concern.