Semantic Road Detection for Non-Ideal Rural Roadways with Limited Apriori Maps

Abstract

In this work we provide a system that has the ability to detect the entire road area ahead and provide semantic information about where to move within that area for an autonomous vehicle on a wide array of road types in rural environments. This approach combines current state of the art image road segmentation, image lane detection and lidar ground segmentation techniques by fusing the output of algorithms based on these techniques to produce a costmap to be used by a local planner for motion planning and vehicle control.

One of the main goals of this thesis was to prove whether this is a sensible task given the current state of the art in the relevant algorithms and computational resources, and to propose improvements that address the limitations of this work. These limitations were determined by applying the proposed system to the real-world driving scenarios in simulation and by evaluation on a test vehicle driven on real public rural roadways.

We have presented a functional system that works within the assumptions and limitations set, and shows higher accuracy results compared to the traditional approach.