An algorithm to estimate mean traffic speed using uncalibrated cameras

Abstract

We present a novel approach to estimate traffic speed using a sequence of images from an uncalibrated camera. We assert that exact calibration is not necessary to estimate speed. Instead, we use: 1) geometric relationships inherently available in the image, 2) some common-sense assumptions that reduce the problem to a one-dimensional geometry, 3) frame differencing to isolate moving edges and track vehicles between frames, and 4) parameters from the distribution of vehicle lengths to estimate speed.