Stereo vision gives a robot depth from two cameras the same way two eyes give you depth — by measuring how far each object shifts between the left and right image, then turning that shift into distance.
Stereo vision uses two cameras a fixed distance apart. Nearby objects shift more between the two images than far ones — measure that shift and you get distance, exactly how your two eyes judge depth.
Hold a finger up and blink each eye in turn: the finger jumps side to side. Distant objects barely move. Your brain turns that jump into depth — and a robot with two cameras does exactly the same thing. That's stereo vision.
How it works
Two cameras are mounted a known distance apart (the baseline). The same real-world point lands at slightly different horizontal positions in the two images — this shift is the disparity. Near objects shift a lot; far objects barely shift. Depth follows a simple triangulation formula:
depth = (focal_length × baseline) / disparity
From two images to a depth map
Rectify the two images, match each pixel across them, measure the horizontal shift (disparity), then triangulate that shift into distance for every pixel.
The hard part: matching
The whole method rests on stereo matching — finding, for each pixel in the left image, its partner in the right. This is easy on richly textured surfaces and hard-to-impossible on blank ones: a plain white wall gives no features to match, so depth there is unreliable. Reflective and transparent surfaces fool it too. Modern systems use either classic block-matching / semi-global matching or learned networks, and many add a projected texture pattern (active stereo, as in the Intel RealSense) to give featureless surfaces something to match.
Stereo vs the alternatives
Compared with LiDAR, stereo is cheap, passive (no emitted laser), and gives dense color + depth together — but it's shorter-range, needs texture and good light, and is compute-heavy. Compared with a structured-light or time-of-flight depth camera, stereo works outdoors in sunlight where those often fail. Many robots fuse several via sensor fusion.
Where you'll see it
Stereo rigs sit on drones (obstacle avoidance), the Mars rovers (they navigate almost entirely by stereo), AR headsets, and warehouse robots. The output — a per-pixel depth map — usually becomes a point cloud for downstream planning.
Why it matters
Stereo vision is the cheapest route to dense 3D perception, and understanding disparity is the foundation for everything from depth cameras to visual SLAM.
