Real-Time Lane Detection Based on Deep Learning

As the research and development of autonomous vehicles has become more active, lane detection technologies for providing road information have become key elements. There are limits to detecting lanes in dynamic driving environments in conventional machine vision research, as the approaches are generally dependent on expert scenarios and fine-tuned heuristics. Deep learning has shown good performance in classifying target information with this distribution of nonlinear data; thus, many studies have actively applied deep learning to lane detection. However, most of these studies have focused on improving the accuracy, rather than on the operating speed. For the work reported herein, a benchmarking deep-learning framework for lane detection was applied with lightened feature extraction modules and decoder modules. These were used to compare performances and to present an indicator for selecting a model for optimizing real-time performance and accuracy. The VGG-16, MobileNet, and ShuffleNet networks were used for the encoder module, whereas frontend dilation and UNet were used for the decoder module. The limitations of the benchmarking framework were analyzed, and perspective loss concepts were applied to the processing of the network using front-view images to ensure improvements in the accuracy and operating speed. All of the candidate networks obtained objective performance indicators based on a large-scale benchmark dataset (TuSimple) and network training with a dataset collected and verified via performance on public roads in Singapore.