Lane Extraction and Quality Evaluation: A Hough Transform Based Approach

Lane recognition helps to guide vehicles and is important in driving assistance and autonomous driving. While deep learning techniques have advanced the accuracy of lane recognition, they only aim to minimize the pixel-wise difference between prediction and ground truth, which does not always reflect the true quality of lane recognition. The predicted road marking (white/yellow lines) pixels also do not directly reflect the lane parameters which are essentially straight lines. In this paper, a novel Hough transform based approach is proposed to simultaneously extract lane parameters and give an intuitive quality score on how the lane parameters match the ground truth. This approach can be used with any pixel-wise lane recognition algorithms.