Omni-Directional Structure from Motion using the Inter-temporal Stochastic Map of Vanishing Points and Lines

The paper describes a system for 6 degree of freedom intertemporal stochastic mapping of an unknown environment using only odometery and omni-directional vision. The method uses a feature-based stochastic map where vanishing points and line segments are extracted and tracked in the omni-directional video sequence. To track lines, we use a novel dynamic programming approach to improve ambiguity resolution, and we use delayed states to aid in the initialization of landmarks in the scene. By reobserving vanishing points we get a direct measurement of the robot’s 3-D attitude that is independent from its translation. Furthermore using vanishing points simplifies the representation since parallel lines share the same direction states. We show the performance of the system in various indoor and outdoor environments and compare the generated 3-D map with a 2-D map generated from a gyro-laser based system.