On feature templates for Particle Filter based lane detection

In this work we propose the application of state-of-the-art feature descriptors into a Particle Filter framework for the lane detection task. The key idea lies on the comparison of image features extracted from the actual measurement with a priori calculated descriptors. First, we demonstrate how a feature expectation can be extracted based on a particle hypothesis. We then propose to define the likelihood function in terms of the distance between the expected feature and the features calculated from the current measurement. We select the Histogram of Oriented Gradients as a descriptor and the Battacharyya distance as a metric. We show that this simple approach is powerful in terms of pattern discrimination and that it opens a new set of possibilities for increasing the robustness of lane detectors.