Robust localization in 3D prior maps for autonomous driving

Summary

Abstract

In order to navigate autonomously, many self-driving vehicles require precise localization within an a priori known map that is annotated with exact lane locations, traffic signs, and additional metadata that govern the rules of the road. This approach transforms the extremely difficult and unpredictable task of online perception into a more structured localization problem---where exact localization in these maps provides the autonomous agent a wealth of knowledge for safe navigation. This thesis presents several novel localization algorithms that leverage a high-fidelity three-dimensional (3D) prior map that together provide a robust and reliable framework for vehicle localization. First, we present a generic probabilistic method for localizing an autonomous vehicle equipped with a 3D light detection and ranging (LIDAR) scanner. This proposed algorithm models the world as a mixture of several Gaussians, characterizing the z-height and reflectivity distribution of the environment---which we rasterize to facilitate fast and exact multiresolution inference. Second, we propose a visual localization strategy that replaces the expensive 3D LIDAR scanners with significantly cheaper, commodity cameras. In doing so, we exploit a graphics processing unit to generate synthetic views of our belief environment, resulting in a localization solution that achieves a similar order of magnitude error rate with a sensor that is several orders of magnitude cheaper. Finally, we propose a visual obstacle detection algorithm that leverages knowledge of our high-fidelity prior maps in its obstacle prediction model. This not only provides obstacle awareness at high rates for vehicle navigation, but also improves our visual localization quality as we are cognizant of static and non-static regions of the environment. All of these proposed algorithms are demonstrated to be real-time solutions for our self-driving car.

Bibtex entry