Learning-based method for lane detection using regionlet representation

摘要

Lane detection is an important enabling or enhancing technology for many intelligent applications. A marker line can be segmented into several image blocks, each of which contains lane marking in the centre. This study proposes a learning-based method for lane localisation via detecting and grouping such image blocks. The authors model the marking class using regionlet representation, in which each image block is regarded as a region and is represented by a group of regionlets. A region feature composed of the features extracted from the regionlets contributes a weak classifier. A cascade structure detector is then trained for lane detection. At early stages, it rejects as many negatives as possible. Each layer of the cascade detector is a strong classifier, which consists of several weak classifiers. A real AdaBoost algorithm is adopted to select the most discriminative features and to train the classifiers. Since the use of regionlet features allows desired performance with only a few weak classifiers and the dimensionality of the features is significantly reduced by principal component analysis, the computational burden of their algorithm is much lower than other learning-based methods. Experiment results demonstrate the computational efficiency and robustness of the method.