Visual Landmark Based 3D Road Course Estimation with Black Box Variational Inference

摘要

In this paper we present an approach which estimates the course of a road over long distances based on static and dynamic scene cues detected by a video camera. The approach is based on a clothoid road model, a probabilistic fusion concept as well as a fast variational inference method. Our experimental results show that the approach outperforms a state-of-the-art road marking-based method in challenging real-world driving situations.