Fitting parametric road models to spatio-temporal derivatives

We present an algorithm for extracting parametric descriptions of roads from motion cues inherent in static video of traffic. First, statistics of spatio-temporal derivatives are accumulated from a static video. We derive new energy terms for fitting B-spline snakes to roads by aligning their direction and speed to be maximally consistent with the spatio-temporal derivatives. Due to the constraint on the derivative of the snake, we are able to produce stable, open-ended snakes, without specifying the location of endpoints. We explore other energy terms, including terms to expand the width of the snake allowing it to cover the entire road.