Landmark recognition using motion-derived scene structures

Abstract

One of the most important tasks in computational vision is to determine the structure of a scene in terms of its three dimensional objects and their spatial relationships. Functions such as landmark recognition, scene analysis, context based target recognition, digital map to scene correlation, and motion path planning all depend on accurate determination of the three dimensional description of the scene. In this paper we investigate a technique for determining structure of a scene from motion, based on the analysis of planes that "slice" a spatio-temporal volume. This technique uses a sequence of two dimensional images and generates a set of lines (epipolar-planar lines) corresponding to the planar surfaces in the scene. This technique was then applied to a variety of simulated scenes, sensor position and sensor speeds and the feasibility of the technique was established. For the reacquisition of landmarks we used the outputs of the epipolar-planar lines and performed a matching between the stored landmarks and the observed scene in a joint two dimensional Hough transform domain. This method was shown to be robust and fast in deciding whether a landmark is reobserved by the sensor. The result of this approach have significance in many fields such as scene analysis, autonomous navigation and target tracking.