Automatic information and safety systems for driving assistance

Abstract

The main object of this thesis is the study of algorithms for automatic information processing and representation, in particular information provided by onboard sensors (2D and 3D), to be used in the context of driving assistance. The work focuses on some of the problems facing todays Autonomous Driving (AD) systems and Advanced Drivers Assistance Systems (ADAS). The document is composed of two parts. The first part describes the design, construction and development of three robotic prototypes, including remarks about onboard sensors, algorithms and software architectures. These robots were used as test beds for testing and validating the developed techniques; additionally, they have participated in several autonomous driving competitions with very good results. The second part presents several algorithms for generating intermediate representations of the raw sensor data. They can be used to enhance existing pattern recognition, detection or navigation techniques, and may thus benefit future AD or ADAS applications. Since vehicles often contain a large amount of sensors of different natures, intermediate representations are particularly advantageous; they can be used for tackling problems related with the diverse nature of the data (2D, 3D, photometric, etc.), with the asynchrony of the data (multiple sensors streaming data at different frequencies), or with the alignment of the data (calibration issues, different sensors providing different measurements of the same object). Within this scope, novel techniques are proposed for computing a multi-camera multi-modal inverse perspective mapping representation, executing color correction between images for obtaining quality mosaics, or to produce a scene representation based on polygonal primitives that can cope with very large amounts of 3D and 2D data, including the ability of refining the representation as new information is continuously received.