Lane detection algorithm for automotive applications: Implementation of a GOLD-based machine vision algorithm on a high-performance system on chip

One of the most important research and commercial topics in the automotive industry today is the design of systems capable to support the driver when traveling by vehicle, with different levels of intervention on the vehicle control. These systems are called Advanced Driving Assistance Systems (ADAS) and they are inserted into a larger research field, which is the autonomous driving task. This topic, also named as Dynamic Driving Task (DDT) following the Society of Automotive Engineering denomination, is one of the most important multidisciplinary engineering challenges of the incoming years. In this scenario, one of the most crucial tasks needed to improve the quality of a system autonomously controlling the vehicle is the environment recognition, namely the determination of areas where the car is allowed to go and the detection of obstacles and hazardous situations. In particular, the lane detection task is the one which is responsible to recognize the drivable area by detecting the lane markers and elaborating the obtained information to describe the geometry of the road ahead of the vehicle. In the first section of this thesis, the autonomous driving and the lane detection topics will be addressed. In the first chapter, a general definition of the autonomous driving task will be given, and the state of the art of the most used system architectures and software/hardware components will be presented. In the second chapter, the lane detection problem will be presented, analyzing the state of the art about this kind of applications, the most common system structures and examples of implementations. In the third chapter of the introduction section, a specific lane detection application will be addressed, the GOLD system. The algorithm used for detecting lane markers in this system will be the base of the implementation of the thesis application. The core topic of this thesis is the implementation of a real-time lane detection system on a high-performance system-on-chip running an Embedded Linux distribution as operating system. The main lane detection algorithm is based on the GOLD system, with some differences in certain low-level operations. The system goal is to recognize the lane markers from a given image of the road ahead the vehicle and to compute a visual and a geometric description of the lane center. The system is composed by: a low-level module that takes as input a stream of frames and provides as output a stream of frames in which the lane markers and the center line of the current lane are highlighted; a graphical application, developed using the Qt framework, responsible to visualize the output stream on a screen. In the second section of the thesis, the software architecture of the system will be firstly presented, describing the methods and the optimizations applied to reach a real-time elaboration of each incoming frame. Then, the description of the hardware components used during the development and the testing phases will be addressed. Finally, in the third section of the thesis, the testing results will be analyzed in terms of computational time performances and in terms of effective correct detections. As appendices, two topics will be addressed: the 3D to 2D image transformation, to explain the process followed during the pre-processing module of the algorithm; the Embedded Linux and Yocto system, to explain how the operating system used in this thesis is composed and how it was adapted to the application needs.