A Lane Detection, Tracking and Recognition System for Smart Vehicles

As important components of intelligent transportation system, lane detection and tracking (LDT) and lane departure warning (LDW) systems have attracted great interest from the computer vision community over the past few years. Conversely, lane markings recognition (LMR) systems received surprisingly little attention. This thesis proposed a real-time lane assisting framework for intelligent vehicles, which consists of a comprehensive module and simplified module. To the best of our knowledge, this is the first parallel architecture that considers not only lane detection and tracking, but also lane marking recognition and departure warning. A lightweight version of the Hough transform, PPHT is used for both modules to detect lines. After detection stage, for the comprehensive module, a novel refinement scheme consisting of angle threshold and segment linking (ATSL) and trapezoidal refinement method (TRM) takes shape and texture information into account, which significantly improves the LDT performance. Also based on TRM, colour and edge informations are used to recognize lane marking colors (white and yellow) and shapes (solid and dashed). In the simplified module, refined MSER blobs dramatically simplifies the preprocessing and refinement stage, and enables the simplified module performs well on lane detection and tracking. Several experiments are conducted in highway and urban roads in Ottawa. The detection rate of the LDT system in comprehensive module average 95.9% and exceed 89.3% in poor conditions, while the recognition rate depends on the quality of lane paint and achieves an average accuracy of 93.1%. The simplified module has an average detection rate of 92.7% and exceeds 84.9% in poor conditions. Except the conventional experimental methods, a novel point cluster evaluation and pdf analysis method have been proposed to evaluate the performance of LDT systems, in terms of the stability, accuracy and similarity to Gaussian distribution.