Route boundary inference with vision and LiDAR

Thesis

The purpose of roads is to carry vehicles. Human drivers can easily distinguish roads and their components (e.g., surfaces, boundaries) using direct and indirect (contextual) clues as they are designed with driving in mind. The colour of road tarmacs, shape of road turns, smoothness of road surfaces, traffic signs, road boundaries, buildings, or even other vehicles provide clues about roads. For autonomous vehicles to safely navigate to a desired location in complex driving scenarios they are required to perceive their surrounding environment even in the presence of occlusions. This requires the use of contextual information in a similar fashion to human perception.

In this thesis, we focus primarily on road boundary detection and present a deep learning based approach to capture contextual information for dealing with occlusions. Many scenes present large-scale occlusion by other road users, preventing direct approaches from fully detecting road boundaries. Conventional neural network architectures fail to infer the exact location of an occluded, narrow, continuous curve running through the image. We tackle this problem with a coupled approach that generates multi-scale parameterised outputs in a discrete-continuous form. We combine the power of deep learning with the data obtained from our novel annotation framework to detect and infer road boundaries irrespective of whether or not the boundaries are visible by taking inspiration from human perception, which uses contextual information to perceive beyond the visible spectrum. Our semi-supervised data annotation framework leverages visual localisation and facilitates the use of deep networks by providing an efficient way to generate thousands of training samples. We present two road boundary detection approaches, camera-based and LiDAR-based, that capture scene context and achieve accurate results. We also demonstrate that the presented approaches have utility in scene understanding and localisation.

Authors: Tarlan Suleymanov

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: curb detection and inference; scene understanding; lateral localisation; deep learning; road segmentation; road boundary detection
• Subjects: Computer vision; Robotics; Neural networks (Computer science); Machine learning