Multi-lane detection using convolutional neural networks and transfer learning

Multi-lane detection is essential in autonomous navigation. Conventional algorithms have difficulties due to diverse image variations caused by illumination variations, occlusions, and shadows. Recently, deep-learning algorithms which use large amounts of training data show dramatic improvements in many areas. In this paper, we address multi-lane detection using convolution neural networks (CNNs) and transfer learning. The CNNs' architectures are used for the detection of multiple lanes, and transfer the learning using pre-learned models from ImageNet to reduce the learning time. A sliding window is used to detect lane candidates on an image and it requires heavy computation time. We present a method to reduce the detection time by changing the input structure when applying trained networks. Images from KITTI are used for training. Experiments are conducted applying trained networks to images obtained in other environments.