High-speed range estimation based on intensity gradient analysis

Autonomous vehicles and robots need to understand their surroundings quickly and with reasonable precision before they can move safely. Many sensors can serve this purpose, but cameras are often preferred because they emit no signal of their own and offer strong range and detail. Most camera-based depth methods rely on stereo vision, and recent work has begun pairing stereo with the camera's own movement. Skifstad's dissertation presents a new way to measure depth when the camera's motion is known, and it can estimate distances to objects in a scene with a single camera that moves sideways.

The central idea is to avoid the correspondence problem, which is the hardest and most computationally costly step in stereo and similar methods. Instead of matching features between two arbitrary frames, the approach exploits the known translation of the camera, which makes it simpler to track shifts of a single pixel across a sequence of images. Because these calculations depend only on neighbouring pixels, they can be run in parallel with ease. The method also tends to flag nearby objects before distant ones, a property that is especially useful for navigation.