Lane recognition in urban environment using optimal control theory

In order to recognize lanes in an environment where lane boundaries are not necessarily present, a world representation is generated that holds the sensory information of a binocular vision system. Results from both the early and the advanced levels of image understanding are input to the planar, 2D map which, in a loose sense, is interpreted to encode the collision risk. In that domain, a trajectory is determined that avoids collisions and at the same time satisfies the optimality criterion of drive comfort represented by minimum lateral acceleration. Deviations from these distributed, desired properties are measured by a composite penalty functional. By doing so, lane recognition, is treated as an optimal control problem. The involved variational calculus is transformed into multivariate function optimization by describing the trajectory via piecewise constant curvature rates. The key point is that by augmenting the system with a planning capability, lane recognition is made possible when the lane shape is not precisely defined