Road Mapping and Localization Using Sparse Semantic Visual Features

We present a novel method for visual mapping and localization autonomous vehicles, by extracting, modeling, optimizing semantic road elements. Specifically, our integrates cascaded deep models to detect standardized elements instead of traditional point features, seek improved pose accuracy map representation compactness. To utilize the structural we model lights signs their representative keypoints skeleton boundary, parameterize lanes via piecewise cubic splines. Based on build complete pipeline localization, which includes a) image processing front-end, b) sensor fusion strategies, c) optimization back-end. Experiments public datasets testing platform have demonstrated effectiveness advantages outperforming approaches.