Visual computing techniques for automated LIDAR annotation with application to intelligent transport systems

Abstract

The concept of Intelligent Transport Systems (ITS) refers to the application of communication and information technologies to transport with the aim of making it more efficient, sustainable, and safer. Computer vision is increasingly being used for ITS applications, such as infrastructure management or advanced driver-assistance systems. The latest progress in computer vision, thanks to the Deep Learning techniques, and the race for autonomous vehicle, have created a growing requirement for annotated data in the automotive industry. The data to be annotated is composed by images captured by the cameras of the vehicles and LIDAR data in the form of point clouds. LIDAR sensors are used for tasks such as object detection and localization. The capacity of LIDAR sensors to identify objects at long distances and to provide estimations of their distance make them very appealing sensors for autonomous driving.This thesis presents a method to automate the annotation of lane markings with LIDAR data. The state of the art of lane markings detection based on LIDAR data is reviewed and a novel method is presented. The precision of the method is evaluated against manually annotated data. Its usefulness is also evaluated, measuring the reduction of the required time to annotate new data thanks to the automatically generated pre-annotations. Finally, the conclusions of this thesis and possible future research lines are presented.