Estimation of vehicle speed by motion tracking on image sequences

This paper presents a method for estimating vehicle speed by tracking the motion of a vehicle through a sequence of images. The motion is derived using an equation based on spherical projection which relates the image motion to the object motion. Motion tracking is done via the Kanade-Lucas-Tomasi algorithm. The motion equation is reformulated into a dynamical space state model, for which Kalman and Extended Kalman filter are applied to estimate the object velocity as well as to predict the future location of the features. The proposed algorithm is employed on a real-life traffic video captured using an un-calibrated camera to estimate the speed of individual vehicles in the video frames. The main advantages are that it is a simple yet robust method having lower time complexity and less ambiguity in vehicle speed estimations.