Detection and tracking of vanishing points in dynamic environments
This thesis is about the detection and tracking of vanishing points in images, its application to rectification of planes, and the additional considerations that the road scenario imposes on these topics. The core element of the dissertation is the obtention of vanishing points, for which a number of contributions are introduced in the eld of robust estimation through optimization procedures. RANSAC (RANdom Sampling And Consensus) and the EM (Expectation-Maximization) algorithm are studied for initialization and refinement purposes, respectively. The results show that these proposals use the information of the images more efficiently and with better performance than other approaches in the literature. Two chapters cover the extraction of the required information from the images in order to feed the estimation methods. On the one hand, images of structured environments (those containing rectilinear elements such as corridors, buildings, roads, etc.), contain line segments that can be detected using the SSWMS (Slice Sampler and Weighted Mean Shift) algorithm, proposed in this thesis. This method has been conceived to work in real-time and to represent a trade-off between accuracy and robustness. On the other hand, focusing on the specific road scenario, the information of the images is extracted through more complex procedures, which are based on recursive Bayesian classi¯cation and model fitting. Finally, the required conditions for the rectification of planes using vanishing points are determined, together with an evaluation of the recovered geometric properties of the plane according to the available information of the scene. Besides, paying special attention to the road scenario, a number of case studies are identified, for which corresponding practical methods for plane rectification are proposed. Along this thesis, results are presented individually for each proposed strategy, comprising feature extraction, vanishing point estimation, road modeling and plane rectification. As a practical application, these strategies are jointly used as a vision-based ADAS (Advanced Driver Assistance Systems) that generates and updates, dynamically, the rectified images and the resulting road model.