Lane Structure Fitting via Row-wise Grid Classification

With people’s attention to driving safety, the development of Advanced Driving Assistance System (ADAS) is becoming more and more mature, and lane detection is a vital part of it. Although most of the lane detection work based on deep learning shows excellent performance, in practical application, we should pursue a better balance between accuracy and speed. Shadow, blur and sparse supervision are still the challenges of lane detection. In this paper, we design a light convolutional neural network (CNN) based on ResNet-18, which achieves the fastest running time under the benchmark of ensuring detection accuracy. Different from the segmentation based method, we use the row-wise grid classification method, which can greatly reduce the amount of calculation. In addition, we propose a loss function with curve fitting function as reference value, which can maintain the lane structure and improve the problem of no visual information. Our model is tested on the TuSimple and CULane datasets. The experimental results show that our method is astonishing in accuracy and speed compared with the existing state-of-the-art methods.