RoadWay lane detection for autonomous driving vehicles via deep learning

Locomotion is basic to all human needs. Modern-day transport has come a long way but still far away from perfection and all-around safety. Lane Detection is a concept of demarcating lanes on the roads while the vehicle is moving. Lane detection algorithm is crucial aspect in making intelligent driving systems that can be used in autonomous self-driving vehicles, road safety, and accidents prevention systems, testing and analyzing driving skills, etc. Lane detection systems using hand crafted features fails in complex scenarios like adverse weather condition, low illumination, sharp turns and occlusion. Recently, deep learning models have been used remarkable in driving assistance systems and shows a significant improvement in their performance. Although, deep learning based methods has shown significant success in lane detection using hybrid techniques, that includes FCN, CNN and RNN. But, a safe driving assistance system can be used to save lives by avoiding accidents, it is crucial to have a real-time lane detection method. We have proposed a lightweight model that can detect lane with high accuracy and low execution time. The size of model has been kept short to make it hardware deployable and perform in real-time. We have designed and trained a deep Convolutional Neural Network (CNN) model for lane detection since a CNN based model is known to work best for image classification datasets. We have used multiple networks and optimization criteria as hyper-parameters and proposed the one with higher F1 score and execution time in comparison to other methods. The training part is done on Supercomputer NVIDIA DGX V100.