Vision-based Road Recognition and Obstacle Detection for Intelligent Vehicles

`Intelligent vehicles` research has been active for more than two decades and is gaining more attention every year.It aims to develop systems able to sense the driving environment and provide information for vehicle control, driver assistance, or traffic network management. A fundamental requirement for intelligent vehicles is an ability to perceive the environment. This can be achieved thanks to appropriate sensors, together with methods to interpret their outputs. This thesis presents advanced methods to exploit vision-based sensors in order to derive a coherent scene representation of both the road surface (road recognition task), and objects that lie on it (obstacle detection task). The proposed solution is a collection of algorithms, each of which is based on a different concept, exploiting different features from the input data. Depending on the required precision of the application and available computation time, all algorithms, or a subset of them, can be activated simultaneously. The results are then fused into a coherent 3D scene representation.The main contribution of this thesis is about road recognition. Firstly, a novel algorithm for the recovery of the parameters of a 3D clothoid-based road model is introduced. Then a B-snake approach to road border detection is detailed, followed by a stereo-based road surface detection algorithm based on plane fitting. Finally, anovel switching-models method for road region fusion is explained.The obstacle detection task is performed in two steps. Firstly, obstacle hypotheses are generated analyzing the stereo map. Secondly, appearance based algorithms and 3D reasoning rules discard those hypotheses that are unlikely to be obstacles. The remaining obstacle regions can then be classified into different classes of vehicles. The classification is performed either by an algorithm based on 3D wireframe models or by a feature based machine learning schema.The algorithms were developed and tested on a real intelligent vehicle prototype. A set of image sequences covering a wide range of different road and traffic conditions was acquired and manually labeled. Quantitative results on labeled sequences are promising, both in terms of precision and computation time